This PDF contains the front matter associated with SPIE Proceedings Volume 11155, including the title page, copyright information, table of contents, and author and conference committee lists.

Modern atmospheric gas monitoring applications demand progressively better performances with regards to spatial, spectral and temporal resolutions. In this context, great potential is shown by a newly developed family of cutting-edge snapshot imaging spectrometers based on Fabry-Perot interferometry, whose conceptual design was patented under the name ImSPOC. Three different sensor prototypes based on the ImSPOC concept are under development: 1) in the near infrared wavelength range for CH₄ or H₂S detection, 2) in ultra-violet and visible range for NO₂, O₄, O₃, and O₂ characterisation and 3) specifically for CO₂ monitoring. After the realisation of these prototypes there is the need arose to provide intelligible and well-calibrated acquisitions for the final users. This study presents the ImSPOC concept from the signal processing point of view, framing the optical transformations performed in the instruments under an appropriate mathematical model formulation. Additionally, preliminary developments are presented to address the first step of the signal processing pipeline for this instrument: the estimation of the thickness of each interferometer. This is a fundamental step for obtaining calibrated acquisitions that could then be used for gas monitoring.

The rectification quality of EUMETSAT´s Meteosat Second Generation (MSG) level 1.5 imagery is routinely monitored for geometrical performance, summarily referred to as geometric quality assessment (GQA). The monitoring includes the absolute navigation error, image-to-image relative geometric accuracy and the inter-channel relative geometric accuracy. Recently, an inquiry from the user community indicated a structured residual error for certain channel pairs, when measuring the inter-channel relative accuracy using direct correlation of image channels. The residual relative interchannel geometrical accuracy shows a clear spatial banding structure in North -South direction. Within each band (i.e. in East-West direction), the measured value of the relative inter-channel geometric accuracy is not changing much, but when going from one band to the next the residual errors change more significantly. The effect is only measurable when comparing a channel located on the one focal plane with a channel on the other focal plane of MSG SEVIRI instrument. Channel - pairs located on the same focal plane do not show the effect. It has therefore been hypothesized that this effect could be associated with a problem in the ground-characterised pointing table for the SEVIRI instrument that associates a scan step with a look angle. Any error in the look angle characterisation would have the same effect on channels of one focal plane, but have a different effect on channels of different located focal planes. While the overall quality of the image is still within specification, an improvement to the processing could be achieved by updating the pointing table with in-flight measurements.

The Multi-viewing, Multi-channel, Multi-polarization Imager (3MI) of the EUMETSAT Polar System - Second Generation (EPS-SG) is a radiometer dedicated to aerosol characterization for climate monitoring, air quality forecasting, and Numerical Weather Prediction (NWP). The 3MI aims to provide multi-spectral, multipolarization, and multi-angular images of the Earth upward radiances at the top of the atmosphere (TOA). This particular design, inspired from the POLarisation and Directionality of the Earth's Reflectances (POLDER) instrument, will lead to a better understanding of the microphysical properties of aerosols and clouds. In terms of design, the instrument's calibration will be done using natural targets, exclusively. This particular constraint requires to further study the already existing calibration techniques that were used on the predecessors of 3MI. The main reason is that the instrument will provide spectral bands (mainly in the short-wave infrared) that will require a reliable characterization. In this paper, we present the results based on measurements obtained with the Observing System Including PolaRisation in the Solar Infrared Spectrum (OSIRIS) instrument, a 3MI airborne simulator developed at the Laboratoire d'Optique Atmosphérique.

In the context of Long Term Data Preservation, ESA, with the support of the Instrument Data Evaluation and Quality Analysis Service (IDEAS+) team, reprocessed over 600’000 Landsat 1-5 Multi-Spectral Scanner (MSS) products acquired in the period 1975 - 2001 from the European Ground Stations complementing the existing 1 million Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) Level 1 products, with a geographical coverage from Greenland to Continental Europe and North Africa. The reprocessing software and quality control tool developed for the reprocessing present new features aimed at improving the radiometric and geometric accuracy, image data recovery and including additional quality assurance information. An accurate geometric and radiometric accuracy evaluation across the over 40-years long operative life of Landsat is vital for the exploitation of long time series analysis and for the future development of multi-mission Analysis Ready Data (ARD). This paper proposes to evaluate the geometric accuracy and the radiometric calibration accuracy of the Landsat Level 1 products delivered by ESA. It is obvious that across Landsat historical mission the accuracy is not the same; data are acquired with the MSS, the TM and the ETM+ sensors with improved characteristics for the most recent ones. All these differences between remote sensing systems, did not preclude to engineer processing algorithms with one main objective to harmonize physical measurement and ensure interoperability with the ongoing missions such as Landsat 8 - OLI / Sentinel 2 – MSI. This approach is a key aspect for the future development of multi mission ARD. The geometric quality assurance parameters are exposed in the Level 1 products. These scene based parameters allows to filter out and select, for a given region, if needed the most accurate products. Also, to demonstrate that these parameters are consistent is fundamental. Because of the ageing of missions, specifically for MSS missions, the processing cannot solely rely on information coming from telemetry. It is mandatory to apply ground model and to estimate both external and internal orientation parameters for what concerned the geometric model. Furthermore, for some parameters, the use of single scene for calibration of geo referencing model is not sufficient and the estimate of parameters becomes more robust when considering, instead of one scene, all data recorded in the acquisition period and downlinked to the receiving station. Also, the proposed methodology herein validates the stability of geometric accuracy for a long period of time within the orbit. The product quality assurance parameters are compared with the same ones but inferred from an image matching methods comparing disparity between two geometric grids; the input ESA image and geometric reference image (Global Land Survey data). A comparison with relative USGS products is also performed.

Present scientific space instruments generate a high amount of raw data while deep-space missions only have a very limited telemetry rate. Because the computation of the scientific relevant parameters is usually accompanied with the reduction of the data, the processing is desired to be carried out already on-board. To accomplish this, the following paper presents a flexible image processing framework which makes use of a heterogeneous data processing module consisting of a space-grade General Purpose Processor (GPP) as well as two dynamically reconfigurable Field-Programmable Gate Arrays used for hardware acceleration. The flexibility and capabilities of the presented framework are proven by means of three exemplary processing tasks of the Polarimetric and Helioseismic Imager (PHI) on-board Solar Orbiter.

In this paper, we present a modified version of a popular component-substitution (CS) pansharpening method, namely the hyperspherical color space (HCS) fusion technique. Unlike other improvements of HCS, the proposed method is insensitive to the format of the data, either calibrated spectral radiance values or uncalibrated digital numbers (DNs), thanks to the use of a multivariate linear regression between the squares of the interpolated MS bands and the squared lowpass filtered Pan, in order to find out the intensity component peculiar of CS methods. The regression of squared MS, instead of the Euclidean radius used by HCS, makes the color space hyper-ellipsoidal instead of hyper-spherical and the intensity component more similar to the lowpass-filtered Pan, such that the extracted detail, namely Pan minus intensity, is more accurate. Furthermore, before the regression is calculated, the interpolated MS bands are diminished by their minima, in order to build a multiplicative injection model with approximately de-hazed components, thereby benefiting from the haze correction, as for all methods exploiting the multiplicative model. Experiments on true GeoEye-1 images show consistent advantages over the baseline HCS and its improvements achieved over time, and a performance comparable with some of the most advanced methods.

We developed a novel SRR system, called Multi-Angle Gotcha image restoration with Generative Adversarial Network (MAGiGAN), to produce resolution enhancement of 3-5 times from multi-pass EO images. The MAGiGAN SRR system uses a combination of photogrammetric and machine vision approaches including image segmentation and shadow labelling, feature matching and densification, estimation of an image degradation model, and deep learning approaches, to retrieve image information from distorted features and training networks. We have tested the MAGiGAN SRR using the NVIDIA® Jetson TX-2 GPU card for onboard processing within a smart-satellite capturing high definition satellite videos, which will enable many innovative remote-sensing applications to be implemented in the future. In this paper, we show SRR processing results from a Planet® SkySat HD 70cm spaceborne video using a GPU version of the MAGiGAN system. Image quality and effective resolution enhancement are measured and discussed.

Super-resolving a satellite imagery from its low-resolution one has a significant impact on the remote sensing industry. There are many potential applications that can directly benefit from this technique. A convolutional neural network (CNN) has recently achieved great success for image super-resolution (SR). However, most deep CNN architectures do not properly handle the inherent trade-off between localization accuracy and the use of global context. In this paper, we propose a stacked lossless deconvolutional network (SLDN) for remote sensing SR. We fully exploit global context information while guaranteeing the recovery of fine details. Specifically, we design a lossless pooling by reformulating the pixel shuffle operator, and incorporate it with a shallow deconvolutional network. The resulting lossless deconvolution blocks (LDBs) are stacked one by one to enlarge the receptive fields without any information loss. We further design an attentive skip connection to improve gradient flows throughout the LDB. The SLDN can reconstruct high-quality satellite images without noticeable artifacts. We also provide an extensive ablation study showing that all the components proposed in this paper are useful for the remote sensing SR. Experimental comparisons demonstrate the superiority of the proposed method over state-of-the-art methods both qualitatively and quantitatively.

High-resolution (HR) remote sensing images are characterized by rich and detailed ground object information with more complex structures of the ground object which make the interference information is more difficult to process. It has always been the focus of domestic and foreign researchers that how to obtain more accurate and higher quality ground object information from these images. The GF-4, the world's first geostationary orbit with high spatial resolution remote sensing satellite, can provide high temporal resolution, large width and 50m pixel resolution of remote sensing data by using area array imaging technology. However, the GF-4 image is a medium resolution and low resolution (LR) image data with relatively vague details of ground objects and not obvious relationships between objects which limit the acquisition of the ground object information to some extent. Therefore, in this paper, we analyze the influence of various factors in the imaging process and construct an image degradation model according to the characteristics of GF-4 satellite images. We adopted the super resolved (SR) method based on Mixed sparse representations (MSR) to increase the spatial resolution of the GF-4 image by twice as much, which not only enriched the detailed information of the image, but also improved the image quality. For the results of SR of GF-4 imagery, we adopted the Maximum Likelihood Classification (MLC) method to perform image classification test and result verification. The experimental area selected in this paper is Yantai City, Shandong Province, China, the LANDSAT 8 OLI data is used as a training sample to calculate the overall accuracy and Kappa coefficient after classification. The results show that the overall accuracy of the superreconstructed result data is 40% higher than that of the source image data from GF-4, especially when the spectral characteristics of the ground objects are obviously different, the accuracy is more obvious. The Kappa coefficient increased 0.4, the extracted outline is more complete and the classification details are more refined.

This paper aims at a novel super-resolution imaging design scheme for CX6-02 micro-satellite. In the design of super resolution reconstruction algorithm, under the framework of the hierarchical bayesian, an improved L1 norm prior model based on gradient adaptation is proposed, to avoid the amplification of noise while deconvolution. In imaging mode and system design, the satellite attitude control deviation and ground velocity compensation are used to obtain multi-frame images with sub-pixel information, and in order to exploit the potential of the reconstruction algorithm, the parameters of the optical system are designed by means of high availability redundancy. These make the CX6-02 satellite’s imaging resolution increased from 2.8 to 1.4 meters in the 700km orbit altitude, the whole satellite is 66kg.

Affinity propagation (AP) is one of the most recent unsupervised classification methods used. Its property is very interesting, but its use reveals the existence of three drawbacks which seriously hamper its usage. These concern its sensitivity to: (i) the presence of identical objects in the population to be partitioned; (ii) the value of the preference parameter p and (iii) the value of the regularization parameter λ introduced for the updating of responsibilities and availabilities criteria. In this paper, an optimization of AP clustering method is proposed which provides an appropriate solution to each drawback underlined above. For the first two, we adapt the value of the preference parameter p for a given object, to the values of the similarities between this object and the others only. In the similarity matrix, any null distance for a given row is reassigned a preference value p, which corresponds to the minimum or average value of this row. Finally, to eliminate the regularization parameter while ensuring the convergence of the algorithm and the optimality of the solution, a smoothing operation on the elements of the responsibilities and the availabilities is introduced, involving the current estimated value and the three previous ones. This approach is tested on synthetic data and hyperspectral images and shows its efficiency. We show both the relevance of the results and their stability without the identified three drawbacks.

Detection and segmentation of motorways, railroads and other roads with similar features are significant for comprehension of both low and high resolution synthetic aperture radar (SAR) imagery. Separation of transportation network from other fields or features is important to understand area contained in SAR image (i.e. the road density can inform about characteristic of that area). Standard image processing methods are inadequate to detect multiple linear targets correctly where computer vision, especially deep learning, provides more insight about features for different type of roads which help better discrimination of multiple linear features like roads and railroads. State-of-art deep learning algorithms are proposed as solutions for understanding road characteristics and extraction of multiple roads. In this paper, a method which uses deep convolutional neural network (DeepLabv3+) backbone architecture is proposed to detect road and railways concurrently. Semantic segmentation of roads using SAR imagery is challenging since these images differ as ground sample distance changes with sensor types which creates a setback for establishing dataset for all sensors. Training set contains 3 classes (road, railway, other) with collected signatures from TerraSAR-X Spotlight images for classification. Proposed method shows robust performance when applied to other sensor and results are presented.

Machine learning has been very actively used in analysis of remote sensing data and for a wide range of applications such as agriculture, land and ice surface change detection, etc. In this paper, we explore use of machine learning methods for forest inventory, since it has great impact on economic and ecological sustainability. Our approach is based on UAV data from hyperspectral and LiDAR sensors. A particular case described in the present paper showcases how aerial survey data achieves high classification accuracy of forest tree species. Moreover, applicability of different data types such as LiDAR and hyperspectral (NIR and VIS) for the task at hand is investigated. In the paper we present the solution in its full scope, starting from remote and in-situ data acquisition and pre-processing, to automatic tree species classification and recognition using machine learning algorithms, as well, as further analysis of the results. In the initial stage, during an expedition to the northern region of Arkhangelsk (Russia), airborne data (LiDAR and hyperspectral) and manually recording of trees locations and species were collected. Afterwards, an object-based classification has been performed by detecting individual tree crowns beforehand with the help of LiDAR data. Various hyperspectral features have been calculated for use in classification algorithms complemented by on-ground spectroscopic benchmark data. In this paper, we prove applicability of the proposed method and workflow in real life application. Results validation was done using data from the observation parcels, where trees were manually labeled. The next step is design of a system with finely tuned filters, which will make possible robust species classification at a cost much lower than hyperspectral imaging.

Tracking a spacecraft with the satellite-based imaging system is an essential task to on-orbit servicing, status monitoring, and visual navigation in the field of remote sensing. Tracking-by-segmentation uses the silhouette to describe the target spacecraft instead of the bounding box representation, which is robust to multiple views of different poses and more suitable for spaceborne applications. But space imagery has the characteristics of low carrier-to-noise ratio, unbalanced illumination and low contrast, which hinders reliable on-board tracking. We propose a tracking-by-segmentation algorithm GCT that adaptively models target appearance with an on-line GMM learning, and applies an external segmentation technique GrabCut to realize the silhouette tracking. The on-line learning GMM provides pixel-level multimodal appearance, which can model the target with sequential viewpoint changing. On the basis of the appearance model, we use GrabCut to construct an objective function based on graph model, which can be minimized with min-cut/maxflow algorithm to track the target’s silhouette. Iterative tracking and appearance learning make our tracker suitable for the target’s multiple views and the space’s extreme imaging condition. Based on the benchmark dataset SegTrack, contrast experiments demonstrate that the proposed algorithm GCT performs favorably on performance and processing speed against the state-of-the-art algorithms such as FLS, SPT and HFT. We also evaluate GCT on the spacecraft datasets including simulation dataset and the ESA’s Speed dataset, and the results show that the proposed technique is scalable to spacecraft of different views as well as robust to the extreme conditions of space.

Remote sensing hyperspectral sensors, with high spectral resolution, allow precise classification of endmembers present in imaged areas. These sensors have a limited spatial resolution, which results in mixed pixels. The mixture is usually assumed to be linear and blind linear spectral unmixing (LSU) methods are used to unmix all observed pixel spectra. Most blind LSU approaches assume that each endmember is represented by a unique spectrum in all image pixels. But, in many practical applications, this assumption is not valid and more complex models are needed to describe other phenomena, e.g. when each endmember needs to be represented by slightly different spectra in all image pixels. This spectral variability must be handled by replacing the concept of endmembers by classes of endmembers, to avoid errors when processing the considered data. In this paper, a new linear mixing model is firstly introduced in order to handle the spectral variability. In the proposed model, the endmember spectra are additively tuned. Then, an algorithm, based on pixel-by-pixel nonnegative matrix factorization, is proposed for unmixing the considered data. That algorithm, which derives, for each class of endmembers, slightly different estimated spectra in all pixels, optimizes a cost function and uses additional constraints that are related to the introduced linear mixing model. Experiments, based on realistic synthetic data, are conducted to evaluate the performance of the proposed algorithm. The obtained results are compared to those of three approaches from the literature. These test results show that the proposed approach outperforms all other tested methods.

In this communication, we address the problem of semi-supervised classification under conditions where (i) learning samples are available only for specific classes and potentially mislabeled, and (ii) the actual number of classes is unknown. For this, we propose a semi-supervised extension of a Nearest-Neighbor - Density Based clustering method, namely the Graph WatershEd using Nearest Neighbor (GWENN) method. We show how an incomplete, erroneous learning sample (LS) set can be incorporated in the algorithm in order to produce efficient labeling decisions partly guided by a priori information, and to discover new classes and correct mislabeled objects. The efficiency of the proposed method, named GWENN-SS, is demonstrated experimentally. We first evaluate its robustness with simulated data for which an erroneous and incomplete LS set is given. We then assess the reliability of GWENN-SS on real hyperspectral images and we show that it can outperform a recent similar semi-supervised approach.

The detection and classification of maritime objects in a harbour environment or coastal areas using a terrestrial hyperspectral system in combination with a high-resolution RGB sensor is a challenging task since the large number of spectral channels requires a robust analytical method. Recently, deep learning methods have shown a good performance in many computer vision applications. In this paper, we present a general analysis workflow for terrestrial ship detection and classification based on fused RGB and hyperspectral images, which employs a deep learning network for the localization of ships in the high-resolution images and a following convolutional neural network based multi-input model for the classification of each detected object. During a measurement campaign, images of various ship types were collected under distinct weather conditions for the training and evaluation of the neural network model. In the first part of the workflow, ship candidates were located using the Mask R-CNN framework based on the RGB images. For the following classification process, which was trained to separate different ship type classes, we developed a multi-input convolutional neural network using the RGB and the hyperspectral images as data samples. For the pre-processing procedure of the hyperspectral data a principal component analysis was applied to reduce the number of input channels for the network while still maintaining a large fraction of the initial information. For the architecture of the RGB classification branch, the structure and the weights of a pre-trained model was integrated and fine-tuned. Since only limited training data was available, regularization methods and data augmentation were employed. The detection and multi-input classification network was finally evaluated and showed that the classification performance can be increased when integrating additional information from a hyperspectral sensor.

We present an algorithm for sub-pixel target detection in hyperspectral images, considering both the common additive target model, and a replacement target model where the target’s spectrum partially replaces that of the background. We implement an LRT based decision rule, estimating the underlying distributions using cluster detection in feature subsets of a decorrelated image. We select these subsets in subspaces corresponding to sets of consecutive eigenvalues of the data’s empiric covariance. The densities are approximated using products of lower dimensional Gaussian mixture models. We utilize the estimated density functions to compute maximum likelihood estimates of the target’s relative portion of the observed pixel spectrum, and obtain a GLRT based test statistic. Performance analysis of this proposed detector shows promising results when compared to the detection capabilities of the matched filter, which is used as a benchmark.

Convolutional Neural Network (CNN) has established as an effective deep learning model for hyperspectral image classification by considering both spectral and spatial information. In this study, the performance of two-dimensional (2D) CNN architecture is evaluated at hyperspectral and multispectral resolution. Two types of multispectral data are analyzed viz., original and transformed multispectral data. Hyperspectral bands are transformed to spectral resolution of multispectral bands by averaging the reflectances of specific hyperspectral narrow bands which are falling within the spectral ranges of multispectral bands. The well-known Pavia University dataset and a new dataset of Pear orchard are investigated in this study. In case of Pear orchard dataset, classification is performed with both types of multispectral data. All the experiments are carried out with the same 2D CNN architecture. In case of Pavia University dataset, hyperspectral and transformed multispectral data achieve OA(%) of 94.29±1.28 and 94.27±2.01 respectively considering 20% samples as training. In case of Pear orchard dataset, hyperspectral, multispectral and transformed multispectral data achieve OA(%) of 91.59±0.89, 88.65±1.35, and 93.24±0.16 respectively considering 20% samples as training. It is evident that transformed multispectral data, which comprises of inherent hyperspectral information, provides similar or better performance compared to hyperspectral data. Further, with the use of 3D CNN architecture, classification performance improves in case of Pavia University dataset, whereas it remains statistically similar in case of Pear orchard dataset. The present promising results illustrates the performance of CNN even in small dataset which is comparable to several published state-of-the art results on the same dataset.

Automatic building segmentation from remote sensing images is critical in the remote sensing image semantic segmentation. The success of deep neural networks has led to advances in using fully convolutional neural networks (FCN) to extract buildings from the high-resolution image. However, the downsampling processing inevitably leads to loss of details of the segmentation results. To solve this problem, some methods try to refine the results of FCN by using probability graph models such as fully connected CRF (Conditional Random Fields). Nevertheless, many fully connected CRF based methods are too time-consuming and not suitable for building segmentation tasks in some situations. In this paper, we propose a novel time-efficient end-to-end CRF model with the domain transform algorithm called DT-CRF. In the proposed model, in order to accelerate the message passing in the mean-field approximate inference algorithm, we take the edge maps as the joint image for DT-CRF and use the domain transformation algorithm to calculate the pair-wise potential instead of the Gaussian kernel function. Meanwhile, we design a multi-task network which can generate masks and edges simultaneously, and the network can make the DT-CRF to easily optimize the segmentation results using model information. The evaluation of remote sensing image datasets verifies the time and space efficiency of the proposed DTCRF and demonstrates a distinct improvement.

Automated 3D reconstruction of urban models from satellite images remains a challenging research topic, with many interesting outcoming applications, such as telecommunications and urban simulation. To reconstruct 3D city from stereo pairs of satellite images, semi-automatic strategies are typically applied, which are based either on procedural modeling, or on the use of both image processing and machine learning methods to infer scene geometries together with semantics. In both cases, human interaction still plays a key role, in particular for the rooftop buildings extraction. In the last decade, the use of deep learning algorithms, notably convolutional neural networks (CNNs), has shown a remarkable success for automatic image interpretation. We propose an approach using CNN architecture to automatize the procedure of building contour extraction with the final purpose to automatize 3D urban reconstruction chain and improve the quality of the generated city models. The developed algorithm consists of three steps: 1) We apply a mask-based normalization technique to the input image. 2) CNN network is applied to obtain a raster map of buildings. 3) A polygonization algorithm is designed, which processes a raster map of building to output an ensemble of building contours. We have adopted a U-Net neural network for building segmentation task. We compare the use of several U-Net architectures with the purpose to retain the best suited model. To train models, we have built a dataset of high-resolution satellite images over 15 different cities, and the corresponding building masks. The experimental results show that the proposed approach succeeds in predicting building polygons in a short time, and exhibits good generalization properties to be applied on diverse Earth areas. The developed algorithm combined with the existing LuxCarta reconstruction chain improves 3D urban scene modeling results, and thus supplies an important step towards the automatic reconstruction of 3D city scenes.

Convolutional neural networks (CNNs) can effectively detect objects in satellite images. They use a large number of labeled samples for training appropriate feature extractor. However, creating labels requires significant concentration and increases the workload of users, because satellite images cover quite large areas relative to the scale of the objects. We propose a human cooperative semi-self-training (SST) framework to reduce users' burden for training a CNN. The SST cooperatively collects training samples from unlabeled samples by repeating the following two phases: self-training and user intervention. The self-training automatically constructs the training dataset and trains the CNN, while the user intervention expands the number of accurately labeled samples. Notably, the SST requests users to label samples only if the automatic training stagnates. We improve the self-training phase to reduce the frequency of user intervention through the following two modules: intelligent dataset construction and pre-training. The intelligent dataset construction module automatically collects only effective training samples based on automatic labeling and evaluation of the collected samples utilizing tentatively trained CNN, while the pre-training module facilitates training of feature extractor in the CNN by allowing the CNN to learn handcrafted image features. Introducing the pre-training module into the dataset construction can effectively improve the self-training because the quality of the automatic labeling based on the pre-trained CNN can be enhanced. The experimental results demonstrated that the CNN trained by the proposed method yielded 92% of performance with only 2.6% of the labeled samples relative to the model trained with the complete dataset.

In the remote sensing area, how to automatically and accurately extract buildings from images is a hot and challenging topic in these years. With the rapid development of sensor and computer hardware technologies, it gets easier to gain remote sensing images with very high-resolution and extract buildings from them by the popular deep learning models such as Fully Convolutional Networks (FCN). However, current FCN based models always lead to blurred building boundaries and have poor abilities on extracting small buildings. Therefore, in this paper, we propose the Gaussian Dilate Convolution, which is a cascade of a trainable Gaussian Filter and an dilate convolution with proper hyperparameter initializations. Also, we carefully design a hierarchical dense feature fusion structure following the dense connection manners. Finally, we embed the Gaussian Dilate Convolution into the hierarchical dense fusion structure and name it as Dense Hierarchical Spatial Gaussian Pool (Dense-HSGP). More specifically, the Gaussian Dilate Convolution has the advantages of the original dilate convolution but preserves much more context information, while the hierarchical dense connection structure of Dense-HSGP provides more abundant receptive fields and higher feature reused abilities within the model. We execute the experiments on the widely used Inrial Labelling Dataset to verify the efficiency of the proposed model. The experimental results show that the proposed model achieves 96.45 % average accuracy and 77.17% IoU respectively, which are distinct improvements rather than several recent state-of-the-art building extraction models.

Semantic segmentation on satellite images is used to automatically detect and classify objects of interest over very large areas. Training a neural network for this task generally requires a lot of human-made ground truth classification masks for each object class of interest. We aim to reduce the time spent by humans in the whole process of image segmentation by learning generic features in an unsupervised manner. Those features are then used to leverage sparse human annotations to compute a dense segmentation of the image. This is achieved by essentially labeling groups of semantically similar pixels at once, instead of labeling each pixel almost individually using strokes. While we apply this method to satellite images, our approach is generic and can be applied to any image and to any class of objects on that image.

This paper presents a novel approach based on the direct use of deep neural networks to approximate wavelet sub-bands for remote sensing (RS) image scene classification in the JPEG 2000 compressed domain. The proposed approach consists of two main steps. The first step aims to approximate the finer level wavelet sub-bands. To this end, we introduce a novel Deep Neural Network approach that utilizes the coarser level binary decoded wavelet sub-bands to approximate the finer level wavelet sub-bands (the image itself) through a series of deconvolutional layers. The second step aims to describe the high-level semantic content of the approximated wavelet sub- bands and to perform scene classification based on the learnt descriptors. This is achieved by: i) a series of convolutional layers for the extraction of descriptors which models the approximated sub-bands; and ii) fully connected layers for the RS image scene classification. Then, we introduce a loss function that allows to learn the parameters of both steps in an end-to-end trainable and unified neural network. The proposed approach requires only the coarser level wavelet sub-bands as input and thus minimizes the amount of decompression applied to the compressed RS images. Experimental results show the effectiveness of the proposed approach in terms of classification accuracy and reduced computational time when compared to the conventional use of Convolutional Neural Networks within the JPEG 2000 compressed domain.

Water region estimation is considered as one of the fundamental classification tasks in remote sensing. Several previous research works focused on traditional practices such as spectral analysis, and statistical approaches for water region estimation. However, producing a consistent global scale water estimation results are still considered as relatively challenging task. On the other hand, in computer vision applications Convolutional Neural Network (CNN) emerged as greater tool for classification tasks. Recently, Recurrent Convolutional Neural Network(R-CNN) proposed for improved classification results. Therefore, inspired from R-CNN, this research proposes a Recurrent feedback Encoder-Decoder without max-pooling for global scale water region estimation using temporal Landsat-8 images. The proposed R-CNN uses three Landsat-8 images which consist of current observation (t0) to predict water region and two previous observation of the same location (t 􀀀 1, t 􀀀 2), and these three temporal observation of the same location were employed for training with the ground truth labelled data (water/non-water) from the current observation. Proposed R-CNN model uses temporal input data and results in multi-temporal output for water region estimation. Experiments show promising results especially while using concatenated recurrent feedback features. The model significantly outperforms baseline model and UNet (without recurrent and feedback structure). Detailed comparison study on temporal Landsat-8 images that highly affected by sunglint, cloud and other atmospheric conditions shows that the proposed model has a potential to produce reliable water region estimation where UNet, baseline model R-CNN single model fail.

Deep learning-based hyperspectral image classification has become very popular recently. It has been widely used in the domain of classification of remote sensing hyperspectral images and has shown encouraging results. Although presence of many spectral bands in a hyperspectral image provides valuable features for the classification purposes, lack of adequate training samples makes training a deep learning model a challenging task. In this paper, we employed stacked auto-encoder (SAE) as the deep learning architecture to extract deep spectral features of the input data and to address the problem of absence of enough training samples, we proposed an effective data augmentation approach to boost the number of training data. Also, in order to alleviate the noise in the output image, we used the majority voting strategy to smooth the final classification map. We applied our method on the Indian Pines hyperspectral dataset including very few training examples. Experimental results show the superiority of our method comparing to some of the state of the arts algorithms.

Automatic dense labeling of multispectral satellite images facilitates faster map update process. Water objects are essential elements of a geographic map. While modern dense labeling methods perform robust segmentation of such objects like roads, buildings, and vegetation, dense labeling of hydrographic regions remains a challenging problem. Water objects change their surface albedo, color, and reflection in different weather and different seasons. Moreover, rivers and lakes can change their boundaries after floods or droughts. Robust documentation of such seasonal changes is an essential task in the field of analysis of satellite imagery. Due to the high variance in water object appearance, their segmentation is usually performed manually by a human operator. Recent advances in machine learning have made possible robust segmentation of static objects such as buildings and roads. To the best of our knowledge, there is little research in the modern literature regarding dense labeling of water regions. This paper is focused on the development of a deep-learning-based method for dense labeling of hydrographic in aerial and satellite imagery. We use the GeoGAN framework and MobileNetV2 as the starting point for our research. The GeoGAN framework uses an aerial image as an input to generate pixel-level annotations of five object classes: building, low vegetation, high vegetation, road, and car. The GeoGAN framework leverages two deep learning approaches to ensure robust labeling: a generator with skip connections and Generative Adversarial Networks. A generator with skip connections performs image→label translation using feed-forward connections between convolutional and deconvolutional layers of the same depth. A GAN framework consists of two competing networks: a generator and a discriminator. The adversarial loss improves the quality of the resulting dense labeling. We made the following contributions to the GeoGAN framework: (1) new MobileNetV2-based generator, (2) adversarial loss function. We term the resulting framework as HydroGAN. We evaluate our HydroGAN model using a new HydroViews dataset focused on dense labeling of areas that are subject to severe flooding during the spring season. The evaluation results are encouraging and demonstrate that our HydroGAN model competes with the state-of-the-art models for dense labeling of aerial and satellite imagery. The evaluation demonstrates that our model can generalize from the training data to previously unseen samples. The developed HydroGAN model is capable of performing dense labeling of water objects in different seasons. We made our model publicly available.

Agricultural fields are monitored for the purpose of EU subsidy eligibility checks. A precondition to make automatic monitoring of fields for this purpose possible is that the object geometric boundary is correct. This precondition can be addressed to some extent by performing image time series analysis to identify changes. Accurate object change detection in agricultural fields on satellite images requires separating object class changes such as new ditches, buildings, or roads from other changes, such as crop development, crop management practices, seasonal variation, or shadows from adjacent objects. In this paper we present an approach to identify unchanged agricultural fields using Deep Neural Networks. We propose a combination of CNNs for semantic segmentation and ConvLSTMs for change detection, applied to multitemporal satellite image time-series of arbitrary length. The neural networks were trained on images acquired over the Netherlands in 2017 by the TripleSat and PlanetScope constellations (0.8 and 3.5m resolution respectively) with RGB and NIR bands. We introduce techniques to create artificial change training data, reducing the need for real training data. The results demonstrate that (1) a neural network is not required to be deep to achieve usable semantic segmentation performance for satellite images for this application, that (2) ConvLSTMs can to some extent compensate imperfect image alignment and pixel misclassification, that (3) longer time series significantly increase the performance of the change detection and that (4) expanding and densifying the time series with lower resolution imagery does not improve accuracy in this particular configuration.

Information regarding both the spatial distribution and the quantity of vegetation components is of great relevance in different fields. Of particular interest is the detection of Non-Photosynthetic Vegetation (NPV) against Photosynthetic Vegetation (PV) and Bare Soil (BS). In-situ approaches exist that identify NPV, but are time and cost expensive. In this context, remote sensing is useful thanks to its ability to provide information at different temporal, spatial and spectral scales. While commonly used vegetation indexes, such as the Normalized Difference Vegetation Index (NDVI), provide robust information to highlight PV, the distinction of NPV from BS is less trivial. Some indices designed for Landsat and Sentinel-2 (S2) bands identify only a part of NPV, the one related to tillage, but they do not provide the proper differentiation of NPV and BS. The Cellulose Absorption Index (CAI) exists that may highlight the presence of NPV. Nevertheless, available broadband multispectral sensors like MODIS, Landsat or S2 do not spectrally resolve these narrow wavelength features, thus CAI cannot be directly extracted. This paper presents a surrogated index for the identification and differentiation of NPV, PV and BS in high spatial resolution S2 Satellite Image Time Series (SITS). To do so, inspiration is taken from the paper presented by Guerschman et al., where a surrogated CAI (CAI*) for MODIS sensor is presented, and moves one step forward in order extend it to high spatial resolution sensors. The S2 CAI* was qualitatively analysed on four different climate zones covering grassland and cropland areas.

A processing chain jointly employing Sentinel-1 and Sentinel-2 data is presented. It aims at monitoring changes in the status of the vegetation cover by integrating the three red-edge (RE) bands of Sentinel-2, at 20 m pixel spacing, with textural features calculated from the Interferometric Wide (IW) Ground Range Detected (GRD) product of Sentinel-1, available at 10 m pixel spacing. Hyper-sharpening on Sentinel-2 multispectral bands and modulation-based integration of Sentinel-1 SAR acquisitions are applied on a multitemporal dataset acquired before and after a recent fire event. New color-composite results of the proposed fusion procedure may be useful for visual inspection and analysis by experts in vegetation status monitoring.

Rapid identification of areas affected by changes is a challenging task in many remote sensing applications. Sentinel-1 (S1) images provided by the European Space Agency (ESA) can be used to monitor such situations due to its high temporal and spatial resolution and indifference to weather. Though a number of deep learning based methods have been proposed in the literature for change detection (CD) in multi-temporal SAR images, most of them require labeled training data. Collecting sufficient labeled multi-temporal data is not trivial, however S1 provides abundant unlabeled data. To this end, we propose a solution for CD in multi-temporal S1 images based on unsupervised training of deep neural networks (DNNs). Unlabeled single-time image patches are used to train a multilayer convolutional-autoencoder (CAE) in unsupervised fashion by minimizing the reconstruction error between the reconstructed output and the input. The trained multilayer CAE is used to extract multi-scale features from both the pre and post change images that are analyzed for CD. The multi-scale features are fused according to a detail-preserving scale-driven approach that allows us to generate change maps by preserving details. The experiments conducted on a S1 dataset from Brumadinho, Brazil confirms the effectiveness of the proposed method.

The coherent nature of the SAR imaging process opens up the opportunity to create interferometric image pairs, which carry a large amount of information about the scene. In this paper, the interferometric coherence is investigated in detail. Coherence is a measure for the temporal stability of the scene with respect to the phase information. Classically, coherence is used for the task of co-registration of the image pair, with the goal of coherence maximization, since such a co-registration will yield the most reliable interferometric phase information. The second important field of application is coherent change detection, i.e. the detection of changes in the scene that most often do not change the backscattering properties of the images and thus are not detectable in the amplitude images. For such an application it is of importance to maximize the contrast between the incoherent changed parts of the scene and the coherent surroundings. With these two applications in mind, in this paper, several published coherence estimation schemes are investigated. The different coherence estimators are applied to an airborne data set, and results regarding coherence maximization and coherence contrast maximization are shown.

Modern remote sensing (RS) systems produce a huge amount of data that should be passed to potential users from sensors or saved. Then, compression is an operation that is extremely useful where lossy compression has found many applications. A requirement to it is not to loose useful information contained in RS data and to provide a rather high compression ratio (CR). This has to be done in automatic manner and quickly enough. One possible approach to ensure minimal or appropriate loss of useful information is to provide a desired visual quality of compressed images where introduced distortions are invisible. In this paper, we show how this can be done for coders based on discrete cosine transform (DCT) that employ either uniform or non-uniform quantization of DCT coefficients. For multichannel images that contain sub-band images with different dynamic range, it is also proposed to carry out preliminary normalization. Additionally, compression performance can be improved if sub-band images are compressed in groups. Then, either introduced distortions are smaller for a given CR or a larger CR is provided for a given level of compressed data quality. Examples for real-life data are presented.

The accurate monitoring and understanding of glacier dynamics are of high relevance for climate science and water-resources management. The glacier parameters are typically estimated by data assimilation methods which inject field measurements into the numerical simulations with the aim of improving the physical model estimates. However, these methods often are not able to capture and model the complexity of the estimation problem. To solve this problem, this paper proposes a method that integrates remote sensing (RS) data, in-situ observations and a physical-based model to accurately estimate the Glacier Mass Balance (GMB). The RS data are used to represent the physical properties of the glaciers by characterizing their topography and spectral properties. Instead of assimilating the observations into the model, the in-situ measurements are used to perform a data-driven correction of the GMB estimates derived from the physically-based simulations in the informative RS feature space. The method is applied to the Alpine MUltiscale Numerical Distributed Simulation ENgine (AMUNDSEN) hydro-climatological model. In the experimental analysis, the multispectral images used to define the feature space are high-resolution Sentinel-2 images. The method is validated on three glaciers in Tyrol (Hintereis, Kasselwand and Varnagt glaciers), in 2015 and 2016. The obtained results show the effectiveness of the method in improving the GMB estimates.

This paper investigates the problem of blind noise parameters estimation (BNPE) of multispectral/hyperspectral images (M/HSI) using deep convolutional neural networks. In contrast to single-band images, M/HSI possess property of interband correlation that can effectively improve quality and accuracy of BNPE. Therefore, in this work, we extend a previously proposed CNN for BNPE of single-band images, called NoiseNet, to vector case. The proposed vNoiseNet Convolutional Neural Network (CNN) can be applied to three-band images including RGB images from Digital Single- Lens Reflex (DSLR) cameras, and subsets of M/HSI bands. Training data for the proposed CNN were obtained from three sources: calibrated images captured by DSLR Nikon D80 camera, AVIRIS data with accurately estimated noise parameters, and Sentinel-2 data with synthetic noise. The vNoiseNet estimates both sub-band (component) image noise variance and uncertainty of this estimate from 32×32×3 image patches. On the basis of a set of such estimates, both signal-independent (SI) and signal-dependent (SD) noise component parameters can be robustly estimated. Experiments on NED2012 database, AVIRIS data and Sentinel-2 data demonstrate high accuracy of the proposed CNN.

Traditional analysis of smoke extent from satellite imagery relies largely on spectral analysis using multispectral data thereby requiring large data volumes or subjective and time-consuming evaluation. These methods are not scalable to observing capabilities of the new generation of remote sensing platforms. We propose an automated, deep learning based detection model capable of identifying smoke plumes from shortwave reflectance for the Geostationary Operational Environmental Satellite R series of satellites. Hand-labelled, past instances of smoke plumes from the NOAA Hazard Mapping System, quality controlled for spatiotemporal accuracy by a subject matter expert, comprises the reference truth dataset. The detection pipeline comprises of pre-process, detection, and post-process stages. A Convolutional Neural Network (CNN), trained on smoke events with varying optical thicknesses and sun-satellite viewing geometry is used to predict the probability score for a given pixel containing smoke. The model is able to detect smoke over both low and high reflectance surfaces and discriminate smoke from clouds though challenges remain in identifying optically thin smoke. Finally, we discuss a web-based interface to visualize daily smoke prediction and analyze the predictions over time.

Within line of sight pointing and stabilisation of EO (Electro-optic) systems operating under motion disturbances it is desirable to measure the inertial orientation of different parts of the system, not just the line of sight - this would allow additional information to be added to the control loop. To implement this a framework to fuse the multiple inertial sensors of the EO system is considered, with an example implemented. The fusion of higher performance sensors located at the line of sight is implemented within the proposed framework, to improve the performance of the estimate at the location of the lower performance sensor. The fusion framework makes use of cascaded Multiplicative Extended Kalman Filter that estimate the multiplicative error of the quaternion orientation estimate.

The upcoming Earth Observation satellite Synthetic Aperture Radar (SAR) missions offer sensing capabilities that imply an increasing complexity of the image formation process. Among the major innovations, the ultra-high resolution Spotlight acquisition, performed exploiting long integration time with steering schemas with a wide range of angles obtained through innovative hardware design; more satellite agility is also essential to acquire wider areas even in nearcontiguous frames within a reduced time period, addressed by Spotlight squinted geometry obtained through both fast and precise attitude maneuvers and electronic antenna steering. It is worth stating that the acquisition of high resolution squinted spotlight scenes, especially in the case of terrains with strong topography variation, could result in a significant degradation of the SAR data focusing performance, mostly for the algorithms working in the Fourier-domain. Considering the challenging scenario described above, this article aims at presenting a new accurate and efficient Spotlight image formation algorithm that uses a focusing core implemented in the time-domain, where the precise accommodation of high topography variations can be easily handled working on pixel basis. The proposed solution has actually a hybrid design: the focusing of the signal in across-track direction uses a frequency-domain approach, while in the along-track direction the focusing is performed in the time-domain. The focusing core of the signal in the azimuth direction uses a Back-Projection approach, which allows to obtain excellent performances where the traditional algorithms in the frequency-domain can have severe limitations, but it will dramatically worsen the temporal performance making the processing algorithm generally not suitable for operational purposes. For this reason, a subaperture processing optimization has been proposed that guarantees significant efficiency improvements primarily for not squinted images and takes into account further considerations in order to obtain a comparable reduction of the computational complexity when applied to data representative of highly squinted geometry. The combination of the described techniques together with the increasing hardware computational power, will allow for a new generation of time-domain focusing algorithms to be implemented in the real world operational scenarios. The performance of the algorithm have been verified and validated mainly by simulation. The accurate Impulse Response Function (IRF) has been measured on a simulated dataset of very high resolution SAR X-Band Spotlight data with squinted and not squinted geometry, with point targets placed at different altitudes using different schemes of topography. The solution has finally been verified on real data from the COSMO-SkyMed (CSK) mission, acquired in the Spotlight Enhanced mode. These high quality X-Band SAR data allow verifying the behavior of the algorithm in terms of performance and stability of the focusing core in a real world scenario. The computational burden of the proposed solution vs its processing execution time has also been assessed.

Performance assessment of image processing systems may be carried out using large volumes of data with known ground truth. Unfortunately such data, collected in sensor trials, can be challenging to source for many problems of interest. In particular, trials collection may require the acquisition of imagery in a range of scenario settings, imaging geometries and environmental conditions. An alternative to trials data collection uses synthetically generated imagery of objects and environments configured into realistic scenarios. For performance assessment of image processing chains, large volumes of synthetic imagery may be required in order to characterise individual algorithmic steps or for complete system assessment. In order to generate sufficiently large volumes for such characterisations the simulation approach must also be fast to execute. This paper presents a process for the generation of simulated Synthetic Aperture Radar (SAR) imagery which is fit-for-purpose for the task of algorithm and systems performance assessment of image processing for Automatic Target Detection, Recognition and Identification (ATDRI) tasks. The approach taken is based on the exploitation of computational geometry primitives. It uses a simplified imaging model and correctly treats both layover effects and shadowed regions on both the target object and within the background region. For speed and simplicity the simulation process synthesises single bounce reflections only. This means that the simulation is effective only up to the intermediate resolutions which are typically used for ATDRI applications. The input models are comprised of three-dimensional triangulations representing the geometric structure of the scene content, with each triangle having a parameterised scattering response based on distributional models often used for SAR imagery. The synthesis process generates a collection of two-dimensional arrays of distributional parameters of the same size as the image to be produced. It is straightforward to use these to generate representations of, for example, mean scattering response, or realistic-looking simulated SAR images with speckle ‘noise’. Results are presented for different scene content and sensor configurations, including target aspect and sensor depression angles.

ABSTRACT Today, ice charts in Greenland waters are produced manually by the Danish Meteorological Institute (DMI) for selected regions depending on season and shipping routes. The project “Automated Downstream Sea Ice Products for Greenland Waters” or shorter “Automated Sea Ice Products” (ASIP) attempts to automate this process by means of fusion of data from instruments with different resolutions and modalities. As a part of this process data from the Advanced Microwave Scanning Radiometer (AMSR2) will be interpolated to the geometry of the SAR data acquired by Sentinel-1. In a preparatory leave-one-out cross-validation (LOOCV) study, different interpolation methods including ordinary kriging (OK) are compared. Using bias and root-mean-squared error (RMSE) as measures of precision, OK using 20-30 nearest neighbours outperforms other often used methods such as inverse distance (ID) weighting. This comes at a cost: more work needs to be done by both the operator and the computer. INTRODUCTION The project “Automated Downstream Sea Ice Products for Greenland Waters” or shorter “Automated Sea Ice Products” (ASIP) is a cooperation between the Danish Meteorological Institute (DMI), two departments at the Technical University of Denmark (DTU), the National Space Institute (DTU Space) and the Department of Applied Mathematics and Computer Science (DTU Compute), and Harnvig Arctic & Maritime. The project is funded by Innovation Fund Denmark. The objective of ASIP is to develop an automatic sea ice product service for Greenland waters which can meet the increasing demands for sea ice information coming from the growing group of users operating in Greenland waters. In the span from traditional, manually produced ice charts and daily downstream sea ice products at coarse resolution, there is a lack of high resolution products delivered in near-real time. ASIP intends to meet this demand by taking advantage of the vast amount of new data from the Copernicus Sentinel satellites and by using a new and innovative data fusion approach and state-of-the-art mathematical/statistical data processing methods: utilization of data from satellite sensors with different modalities/capabilities will facilitate the making of ice products that are reproducible, independent of operator, daylight, weather and season and will result in a significant increase in product temporal frequency and geographical coverage compared to existing ice products. The statistical algorithms work directly in the Sentinel-1 scene coordinate system. In order to make use of the information in the AMSR2 data along with the radar data an alignment of the AMSR2 data to the radar coordinate system is therefore necessary. In this process of interpolating AMSR2 data to the Sentinel SAR data, in a preparatory study six methods are compared by means of leave-one-out cross-validation (LOOCV) 1. nearest neighbour (NN, one neighbour only), 2. triangulated irregular network (TIN, three neighbours only), 3. local mean (LM), 4. inverse distance (ID), 5. inverse squared distance (ID2), and 6. ordinary kriging (OK). Bias and and root-mean-squared error (RMSE) are used as measures of precision. OK with 20-30 nearest neighbours obtains a LOOCV bias of around 0.001 K and RMSE of around 1.1 K. The second best of the six methods is ID2 which with 5-10 nearest neighbours gives a LOOCV bias of around 0.01 K and RMSE of around 3 K. When we use kriging we must estimate semivariograms and model them, this takes operator as well as computer time. The project homepage https://asip.dk will be launched soon.

Radar sounders mounted on airborne platforms have acquired data of the subsurface of the Earth's icy areas over the last decades. These data, called radargrams, contain information on the dielectric discontinuities in the ice-sheets, and thus on the buried geological structures and the related processes. Conventionally, these structures have been characterized and mapped by visually inspecting the radargrams. However, visual inspection is subjective and time-consuming and can lead to misinterpretations. Recently, state-of-the-art automatic techniques are proposed to map the position of the bedrock, the ice layering, and the noise in the radargram. However, there are no automatic techniques for mapping the basal refreezing, which is an important ice target that controls the rate of sea-ward ow of the ice-sheets. This paper proposes an automatic method to map the refreezing ice in radargrams. We model the refreezing ice considering its geophysical and radiometric properties. Then, we design a set of features considering this model to perform a classification of the radargrams into four classes, i.e., ice layering, echo-free zone (EFZ) and thermal noise, bedrock, and the refreezing ice. We applied the proposed method to radargrams acquired in the north Greenland by Multichannel Coherent Radar Depth Sounder (MCoRDS3), a radar sounder designed by the Center for Remote Sensing of Ice Sheets (CReSIS). The results indicate a good overall accuracy. The accuracy of refreezing ice is high, while that of the other classes is comparable with the state-of-the-art techniques. The results indicate the effectiveness of the proposed features in mapping the refreezing ice.

Inverse Synthetic Aperture Radar (ISAR) relies on the motion of the observed target to synthetically generate an image. While the motion of the target might be known for controlled turn-table experiments, this is not true for the general case of non- cooperative objects. For The purpose of obtaining a focused high resolution image of the object, the relative motion between the radar and the object must be estimated. The process of estimating the relative motion components based on the received signals only and then compensating them to produce the ISAR image is called ISAR Autofocus. Compressed Sensing (CS) tackles the problem of recovering an unknown signal from fewer measurements than that required by the Shannon's theory. CS assumes that the signal to be recovered is sparse either directly or under some other representation basis. In general, the object's reflectivity distribution is not sparse. However, in some cases, the ISAR measurements can be approximated by the superposition of the echoes of a group of scattering centers. This interpretation of ISAR images allows for the use of CS algorithms for the reconstruction of ISAR images. In this paper, we propose a Compressed Sensing (CS) based algorithm for estimating the relative motion between the radar and the object as a first step for focusing ISAR images. We verify, with simulated data, the ability of the proposed algorithm to estimate both the relative translational and rotational motion of the observed object with respect to the radar. Future work will test the performance of the algorithm with real data as well as modify the algorithm to work with three dimensional motion components.

The analytic solution of sparse signal reconstruction algorithm based on L₁ regularization is a biased estimation, which leads to the underestimation of target intensity when applied to sparse SAR imaging, resulting in the bias effect and affecting the reconstruction accuracy. In this paper, we quantitatively analyze the bias effect in SAR imaging applications, and analyse the influence of target intensity, signal-to-noise ratio, intensity ratio of adjacent targets in the observation scene on the reconstruction bias. In order to suppress the bias effect and improve the reconstruction accuracy, we adopt a class of algorithms based on nonconvex penalty, and verify the performance of these algorithms using simulations and real data.

The robust small and dim target detection is a key technique for infrared search and track system, and it is still a challenging task due to the complex scenarios and noise. An approach based on total variation regularization and principal component pursuit (TV-PCP) method is proposed to achieve robust target detection performance in non-smooth and non-uniform scene, because total variation (TV) regularization could describe the inner smooth and crisp edges of background. Nevertheless, there are still two drawbacks: 1) TV-PCP model only considers the spatial information of single frame and 2) the vanilla nuclear norm adopted in TV-PCP model is only suitable for the scenarios of sufficient edge samples, which would cause some sparse background residuals and increase the false alarm rate. Inspired by this, a novel small target detection approach based on a new reweighted infrared patch image (IPI) model and TV regularization is proposed by utilizing both spatial and temporal information. Then, the TV regularization and weighted nuclear norm minimization are adopted to separate the target and background. For the low-rank background part, we adopt the weighted nuclear norm and TV regularization to describe the smoothness. For sparse target part and noise part, we use the reweighted l1 norm and Frobenius norm term to characterize, respectively. Finally, the proposed model can be solved efficiently by the Alternating Direction Method of Multipliers (ADMM) method. Extensive experiments demonstrate that the background suppression ability and target detection probability of the proposed method is better than the other competitive methods.

In this paper, we propose a constrained sparse representation (CSR) based algorithm for target detection in hyperspectral imagery. This algorithm is based on the concept that each pixel lies in a low-dimensional sub- space spanned by target and background training samples. Therefore, it can be linearly represented by these samples weighted by a sparse vector. According to the spectral linear mixture model (LMM), the non-negativity constraint and sum-to-one constraint are imposed to the sparse vector. According to the Karush Kuhn Tucker (KKT) conditions, the upper bound constraint on sparsity level is removed. Besides, to alleviate the effects of target contamination in the background dictionary, an upper bound constraint is given to the weights corresponding to the atoms in the background dictionary. Finally, this constrained sparsity model is solved by a fast sequential minimal optimization (SMO) method. Different from other sparsity-based models, both the residuals and weights are used to detect targets in our algorithm, resulting in a better detection performance. The major advantage of the proposed method is the capability to suppress target signals in the background dictionary. The proposed method was compared to several traditional detectors including spectral matched filter (SMF), adaptive subspace detector (ASD), matched subspace detector (MSD), and sparse representation (SR) based detector. The commonly used receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) are adopted for performance evaluation. Extensive experiments are conducted on two real hyperspectral data sets. It is demonstrated that our CSR method is robust to different target contamination levels in the background dictionary. From these experiments, it can be seen that our CSR method achieves a much higher target detection probability than other traditional methods at all false alarm rates. Meanwhile, our CSR method achieves the highest AUC value, which is significantly larger than most traditional methods. Moreover, the proposed method also have a relatively low computational cost.

Semantic segmentation of remote sensing data such as multispectral imagery has been boosted recently using deep convolutional neural networks (CNN). However, segmentation of multispectral images using supervised machine learning algorithms such as CNN requires a significant number of pixel-level annotated data, often unavailable, making the task extremely challenging. To address this, this paper puts forward a semi-supervised framework, based on generative adversarial networks (GAN). The proposed solution consists of a generator network to provide photo-realistic images as extra training data to a multi-class classifier acting as a discriminator and trained on a small annotated dataset. Performance of the proposed semi-supervised GAN is evaluated on two benchmarks multispectral semantic segmentation datasets collected from urban scenes of Vaihingen and Potsdam. Results indicate that the proposed framework achieves competitive performance compared to state-of-the-art semantic segmentation methods and show the potential of GAN-based methods for the challenging task of multispectral image segmentation.

Nowadays, Satellite images are used for various analysis, including building detection and road extraction, which are directly beneficial to governmental applications, such as urbanization and monitoring the environment. Spatial resolution is an element of crucial impact on the usage of remote sensing imagery. High spatial resolution means satellite images provide more detailed information. To improve the spatial resolution at the sensor level, many factors are ought to be taken into consideration, such as the manufacturing process. Moreover, once the satellite is launched, no further action can be taken from the perspective of hardware. Therefore, a more practical solution to improve the resolution of a satellite image is to use Single Image Super Resolution (SISR) techniques. This research proposal deals with the re-design, implementation, and evaluation of SISR technique using Deep Convolutional Neural Network with Skip Connections and Network in Network (DCSCN) for enlarging multispectral remote sensing images captured by DubaiSat-2 (DS-2) and estimating the missing high frequency details. The goal is to achieve high performance in terms of quality, and to test whether training the network using luminance channel only, which is extracted from YCbCr domain, can achieve high quality results. For this purpose, DCSCN is trained, evaluated, and tested using a dataset collected from DS-2. A single low resolution DS-2 image is used to construct its high resolution version by training the model from scratch and fine-tuning its hyper-parameters to produce optimal results. The performance is evaluated using various quality indices, such as Structural Similarity Index Measurement (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Wavelet domain Signal-to-Noise Ratio (WSNR). The performance is compared to other state-of-the-art methods, such as Bil-inear, Bi-cubic, and Lanczos interpolation.

Subsurface investigation of the Jovian icy moons is expected to disclose interesting information on the Jovian system. The Radar for Icy Moon Exploration (RIME) is the instrument in charge of characterizing the subsurface of the three icy moons Ganymede, Europa and Callisto. To provide a key for interpretation for the real acquired data, simulations of different possible scenarios on Ganymede are presented in this work. In this paper, we present an approach to performance analysis of RIME based on the 3D modelling and electromagnetic simulations of selected icy moon targets. These simulations are carried out using the Finite-Difference Time-Domain (FDTD) technique, which has been used in recent years to support radar sounder applications. In this work, we analyze in detail some interesting targets: 1) the dark terrain regolith, 2) the bright terrain dielectric profile, and 3) the grooved bright terrain. Our analysis is performed in two levels. First, the contribution of individual features is analyzed, varying their geometry and composition to understand how the measured fields vary accordingly. Second, a more realistic geological arrangement of a combination of subsurface features is considered. The results are very promising and indicate that the subsurface response is detectable in most of the cases.

Precise quantification of land surface dynamics finds its way in various applications ranging from hydrological modeling to disaster management studies. The elevation accuracy of global DEMs generated though stereo images and interferometric techniques are prone to bias due to vegetation and built-up structure. Thus, the global elevation models depict the surface elevations rather than bare earth elevation. Recent advancement in point cloud processing from LiDAR has paved way for isolating ground points i.e., bare earth points from nonground points, thus improving bare earth model generation. Point clouds generated through photogrammetric techniques from very high resolution stereo images has been a promising alternative to LiDAR data. The objective of this study is to understand the capability of point cloud data from spaceborne LiDAR and high resolution stereo images for creating accurate bare earth models for Indian region. This study employs point cloud obtained from ICESat mission and point cloud generated from high-resolution satellite stereo images from Cartosat-1 mission. Various parameters extracted from ICESat waveform were used to extract the bare earth footprints. Filtering techniques were used to extract ground points from photogrammetrically derived point cloud. The bare earth points were interpolated to generate DEM. Resultant DEM were compared to global DEM for elevation changes with respect to different land cover classes.

Ship classification in remote sensing images has been rarely studied because of relative scarcity of publicly available datasets. It is well known that datasets have played an important role in object classification research, especially for CNN-based algorithms which have been proved to perform well. In this paper, we introduce a public Dataset for Ship Classification in Remote sensing images (DSCR). We collect 1,951 remote sensing images from DOTA, HRSC2016, NWPU VHR-10 and Google Earth, containing warships and civilian ships of various scales. For object classification, we cut out ships of different categories from the collected images. The whole dataset contains about 20,675 instances which are divided into seven categories, i.e. aircraft carrier, destroyer, assault ship, combat ship, cruiser, other military ship and civilian ship. Each image contains ships of the same category, which is labeled by the category name. Since our dataset contains most models of major warships, it is relatively comprehensive for ship classification. To build a benchmark for ship classification, we evaluated six popular CNN-based object classification algorithms on our dataset, including ResNet, ResNext, VGG, GoogLeNet, DenseNet, and AlexNet. Experiments demonstrates that our dataset can be used for verifying ship classification algorithms and may advance the development of ship classification in remote sensing images.

Land cover classification and change detection analysis based on remote sensing images using machine learning algorithm has become one of the important factors for environmental management and urban planning. We select Yangon as the study area because the government faces many problems in urban planning sectors due to the population growth and urban sprawl. Therefore, the proposed method aims to perform the land cover classification in Yangon using Random Forest (RF) classifier in Google Earth Engine (GEE) and post-classification change detection between 1987 and 2017 with 5 years interval periods are evaluated. Despite land cover classifications using satellite imagery have been executed in the past decades, the classification of remotely sensed data integrating with multiple spectral, temporal and textural features and processing time for classification using time series data still have limitations. To overcome these limitations, features extracted from Sentinel-2, Landsat-8, Landsat-7, Landsat-5 and Open Street Map (OSM) are executed for classification and cloud-based GEE platform is used to reduce the processing time. Some spectral indexes such as NDVI, NDBI and slope from SRTM are calculated to achieve better classification. Land cover classification is performed by using the RF classifier with the different bands’ combination. Land cover classification map with 7 classes (Shrub Land, Bare Land, Forest, Vegetation, Urban Area, Lake and River) is obtained with the overall accuracy of 96.73% and kappa statistic of 0.95 for 2017. Finally, change detection analysis over 30 years is performed and the significant changes in build-up, bare land, and agriculture have been resulted.

Object detection in aerial images plays an important role for a wide range of applications. Although many efforts have been done in the last decade, it is still an active and challenging problem because of the highly complex backgrounds and the large variations in the visual appearance of objects caused by viewpoint variation, occlusion, illumination, etc. Recently, many object detectors based on deep learning demonstrate the great advantages for significantly improving the detection performance in aerial images. However, the most accuracy neural networks usually have hundreds of layers and thousands of channels, thus requiring huge computation and memory consumption. Besides, the state-of-the-art object detectors are usually fined-tuned from the models pretrained on classification dataset ImageNet, which limits the modification of network architecture and also leads to learning bias because of the different domains. In this paper we trained a lightweight convolutional neural network from scratch to perform object detection in aerial images. When designing the lightweight network, Concatenated Rectified Linear Units (CReLU) and depthwise separable convolution operation were employed to reduce the computation cost and model size. When training the lightweight network from scratch, we employ Group Normalization (GN) in each convolution layer, which makes smoother optimization landscape and has more stable gradients. A serial of ablation experiments is conducted on the recently published large-scale Dataset for Object detection in Aerial images (DOTA), and the results show that the proposed object detection methods with lightweight network trained from scratch achieves competitive performance but has smaller model size and lower computation cost.

In order to accomplish the remote sensing detection demands of drones, which is aimed at potential hazards such as explosives and pollutants on the disaster site, the uav-based hyperspectral imaging remote sensing detection methods should be researched. Taking into account the spectral information on disaster site is frequently weak, the detection mechanism of hyperspectral imaging technology leads to its low sensitivity to weak information. The integrated multiplexing method of complex optical path, high-sensitivity hyperspectral spectroscopic imaging function, integrated design method, and the optical aberration optimization of wide band in large field of view are studied, the lightminiaturization high-sensitivity hyperspectral imager on drones has been developed, which breaks through the key technologies of high-sensitivity spectral imaging. Meanwhile, in order to ensure the timeliness of explosive information detection, the rapid and high-precision extracting processing methods of the large amount of hyperspectral data for targets such as explosives on the disaster site must be deliberated, while developing the rapid and high-precision imaging hyperspectral data processing software, to provide rapidly detailed information supporting for accident disposal and rescue deployment on the disaster site, and implement the practical application of hyperspectral imaging technology in disaster and environmental monitoring.

In recent 10 years, forest damage caused by forest fires in Korea has increased significantly compared to previous years. Therefore, interest and concern about damage caused by forest fires are very important in terms of environmental and ecosystem. According to various domestic and international research results, forests perform functions such as reporting of life resources, prevention of desertification, and adjustment of micro climate. There are many studies to extract the damage areas based on hyper spectral aerial image, high resolution satellite image, vegetation index and factors affecting the forest environment. However, there are limitations that the indexes have different threshold values depending on the region and season, and the threshold value must be continuously adjusted in order to detect the concentration of the damage areas. In this study, we detected forest disaster damaged areas through satellite image data and deep learning. We collected image data on Landsat satellite and applied to the detection of damaged area using U-net [1] and SegNet [2] models. We tried to verify the applicability of semantic segmentation for remote sensing, compare and evaluate each model, and build an optimal forest disaster detection model.

Payload simulation technology is a method to simulate the various working modes and working states of real payloads through a mathematical model, so that people can know better the drawbacks of the payload under certain working conditions before the payload is manufactured. It is significant for payload manufacture, because it can discover the potential design flaws and guides the design of the optical remote sensor, provides a reference for the manufacture and performance evaluation of the remote sensor further. In the simulation procession, the optical imaging system aberrations are fully considered, and the dynamic characterization model of the aberration and optical imaging system performance degradation is established. The comparison between the simulation data and the experimental sample acquisition data proves the correctness of the simulation procession.

The paper addresses the further advance in our complex research in the field of multisensory image fusion based on generative adversarial models [1-2] and their application to such practical tasks as visual representation of fused images, acquired in different spectral ranges (e.g. TV and IR), and changes detection on images, acquired in different conditions (e.g. season-varying images). A developed architecture of a neural network based on pix2pix model is presented, which can solve the both tasks mentioned above. A technique for generating training and test datasets including data augmentation process is described. The results are demonstrated on real-world images.

Differential Interferometric Synthetic Aperture Radar (DInSAR) is considered to be one of the most accurate geodetic techniques for monitoring Earth's surface change phenomena. Corner reflectors (CR) are widely used to validate the DInSAR generated deformation results because of their better phase stability and higher radar cross-section. A controlled experiment has been performed for carrying out a sensitivity analysis of DInSAR technique with the controlled vertical movement of CR in between two COSMO-SkyMed tandem passes over Jharia Coal Field. From the results, it has been observed that CR-based DInSAR experiment followed the similar displacement trend with a maximum of 10% error limit.

Precise high-resolution Digital Elevation Models (DEMs) are essential for creation of terrain relief and associated terrain hazard area maps, urban land development, smart cities and in other applications. The 3D modelling system entitled the UCL Co-registration Ames Stereo Pipeline (ASP) Gotcha Optimised (CASP-GO) was demonstrated on stereo data of Mars to generate 3D models for around 20% of Martian surface using cloud computers which was reported in 2018. CASP-GO is an automated DEM/DTM processing chain for NASA Mars, lunar and Earth Observation data including Mars 6m Context Camera (CTX) and High Resolution Imaging Science Experiment (HiRISE) 25cm stereo-data as well as ASTER 18m stereo data acquired on the NASA EOS Terra platform. CASP-GO uses tie-point based multi- resolution image co-registration, combined with sub-pixel refinement and densification. It is based on a combination of the NASA ASP and an adaptive least squares cor- relation and region growing matcher called Gotcha (Gruen-Otto-Chau). CASP-GO was successfully applied to produce more than 5300 DTMs of Mars (http://www.i- Mars.eu/web-GIS). This work employs CASP-GO to obtain DEMs from high resolution Earth Observation (EO) satellite video system SSTL Carbonite-2. CASP- GO was modified to work with multi-view point-and-stare video data including subpixel fusion of point clouds. Multi-view stereo video data are distinguished from still image data by a richer amount of information and noisier water areas.

Object detection and recognition is one of the important problems in many remote sensing applications such as monitoring, security, rescue mission and other data analysis tasks. So a large number of approaches and techniques have been developed to improve the quality of object detection. Recently, deep learning methods have made significant progress in detecting objects. But when dealing with objects having small size in the image (which are the often case in monitoring or rescue mission), the quality of detection noticeably decreases. To study the performance and the limits of deep learning abilities for object detection in remote sensing imagery with degrading object resolution in the image, a special dataset of aerial images containing objects of interest (human, car) at different resolution has been collected. The dataset consists of images acquired at different distances to the objects of interest, providing representative subsets of object images of various scale. Two state-of-the-art object detection convolutional neural networks (Faster-RCNN and SSD) was evaluated on the collected dataset. The aim of the study was to find out how the object size in the image influences on the detection performance and to estimate the value of object image size at which the performance drops significantly. Also the approaches for improving the small size object recognition were developed and evaluated. First approach uses multimodal image fusion, the second one applies deep learning to increase the resolution of small objects in the image. The performed tests have proved that the developed approaches allow to improve the quality of object recognition when dealing with low resolution object images.

Hyperspectral image can acquire hundreds of bands with wavelengths ranging from visible spectrum to infrared, and the rich discriminant information has led to the widespread applications in the military and civil fields. However, due to the influence of imaging devices, some bands are polluted by noise, which bring great inconvenience to subsequent processing. Therefore, in order to quickly and accurately find low quality and noisy bands in hyperspectral image, we propose a hyperspectral image quality analysis method based on band dictionary representation. Firstly, a representative band dictionary is constructed to accurately represent the dominant information in hyperspectral image. Then, the band subset dictionary is used to reconstruct all the remaining bands in the hyperspectral image and obtain the reconstruction coefficients of each band. Finally, representation error is calculated to analyze the quality of each band. By comparing the representation errors of the bands, the quality of each band can be estimated. Experimental results on three real-world hyperspectral images demonstrate the proposed method can effectively and quickly select low quality bands and noisy bands without any priors.

Mapping of built-up areas were always a main concern to researchers in the field of remote sensing. Thus, several techniques have been proposed to saving technicians from digitizing hundreds of areas by hand. Multiclass classifiers exhibit a very promising performance in terms of classification accuracy. However, they require that all classes in the study area to be labeled. In many applications, users may only be interested in a specific land class. This referred to as one-class classification (OC) problem. In this paper, we compare a Binary Support Vector Machine (BSVM) classifier, with two OC classifiers, OC SVM (OCSVM), and Presence and Background Learning (PBL) framework for the extracting built-up areas from Gaofen-2 and Aster satellites imagery. The obtained classification accuracies show that PBL provides competitive extraction results due to the fact that PBL is a positive-unlabeled method based on neural network in which large amounts of available unlabeled samples is incorporated into the training phase, allowing the classifier to model the built-up class more effectively.

In this study, deep neural network is utilized as another approach for improving accuracy of the precipitation based on microwave-sensor. And The ensemble Bayesian model averaging(EBMA), which employs a weighting scheme for each member using posterior probability, in order to produce a more improved blending of precipitation from multi-satellite and to evaluate the effect of accuracy improvement. Experiments to improve rain rate were carried out based on data obtained from Global Precipitation Measurement (GPM) Microwave Imager (GMI). Input data for the DNN model include 7 brightness temperatures (Tb), ice water path (IWP), convective rain rate, scattering index (SI) and land sea mask is used. The experiment for blending of precipitation product was performed using rain rate product of three satellites and sensors, namely GMI of GPM core observatory, special sensor microwave imager/sounder (SSMI/S) of the Defense Meteorological Satellite Program (DMSP) F16 and microwave humidity sounder (MHS) of NOAA-18. In both experiments, precipitation product of the Dual-frequency Precipitation Radar (DPR) of CO was used as reference data. The probability density function(PDF) of gamma distribution combined with logistic regression is used to estimate the probability and quantity of precipitation for considering the characteristics of precipitation. And then, the exponent for these two functions and the percentile threshold of the cumulative density function were set by optimizing simulations. After that, the validation statistics of the blending precipitation through comparison with precipitation obtained from DPR is carried out.

A growing number of offshore wind farms have been deployed in recent years, comprising up to hundreds of wind turbines. The high number of wind turbines and the effort to reach an offshore wind turbine motivate the use of unmanned aerial vehicles (UAV) or drones for remote visual inspection. The images taken by the drones need to be inspected to detect and classify various kinds of damages. To leverage the effort for the detection and classification of damages, it is a good practice to segment the wind turbines from the background in a first step. In this paper, we compare the speed and segmentation accuracy of a k-Nearest-Neighbor classifier when applied to pixels or superpixels. Pixel-by-pixel classification is compared to superpixel classification using two different kinds of superpixel representations: Superpixels gained by Single Linear Iterative Clustering (SLIC) and Boundary Aware Superpixel Segmentation (BASS). Our results clearly show that superpixel based segmentation increases the segmentation speed by a relevant factor while the segmentation accuracy is at least on the same level or even improves in one dataset.

When characterizing the imaging performance of tactical airborne imagers, one must account for the inherent subjectivity of human operators interpreting objects within a scene. One common approach to measuring imaging performance is the National Imagery Interpretability Rating Scale (NIIRS). However, the ground resolvable distance (GRD) is often a preferred metric because it can more easily be predicted through analysis, measured with field data, and traced to the laboratory measurements of system components. Although GRD is a calculable metric, it ultimately contains some ambiguity when linking it to an operator’s experience. Because of this, the ISR community sometimes utilizes different approaches to modeling and even measuring GRD. This can lead to confusion and disagreements when trying to compare the performance of different sensors. This paper explores different methods for modeling and measuring GRD. Emphasis is placed on airborne hyperspectral imaging (HSI) sensors. The additional spectral information provided by HSI sensors allows one to explore the relative importance of spatial resolution within the framework of HSI target detection capabilities. Modeled HSI performance will be validated against real world measurements performed with different airborne hyperspectral sensors. Different modeling and measurement approaches will be evaluated based on flight imagery, and the relative strengths and weaknesses of these different approaches will be discussed. In particular, a new model is presented, which is based on methods from laser physics of measuring the width of the point spread function.

In recent decades, global sea level is rising year by year and some low and close to the coastal regions will face to be swallowed up by the sea. It is necessary and important to monitor and even forecast the sea level change to study the evolution of land-ocean system and its impact on human activities. In this study, satellite altimeter retrieved monthly average sea level height products from 1993 to 2017 are used to extract the components of their trends, seasonal and residual. Its trends which are obtained by the linear regression method in South China Sea increases about 4.03 mm annually within 25 years. Using the Fourier spectrum analysis and Harmonic analysis, its period mainly has some cycles of 12, 30 and 150 months. The residual component is analyzed in temporal and spatial by Empirical Orthogonal Function and the changes of their temporal coefficient are forecasted by Auto-Regressive Moving Average model. The result shows that the contribution of residual components is mainly to display the details of sea level height, and they affect the accuracy of the sea level prediction. By some post-prediction experiments, the estimated sea level in South China Sea, which are the linear combination of their trends, seasonal and residual components forecasting, are consistent with satellite observations. In addition, there are some deviations in details on some regions which are decided by the prediction of residual components. In the future, those conclusions are helpful in assessing the hazards of sea level rise from long time-series by satellite altimeter.

Scanners are widely used in modern industry and daily life, and they have a wide range of applications. They are used not only for medical measuring and imaging equipment but also for cutting and processing equipment in the industrial field, as well as for information processing, display and data collection in various technologies, such as projectors and 3D image sensors. With the development of technology, the use of scanners has been increasing as automation has progressed. In particular, with the development of unmanned mobile technology, sensors using scanners are being developed. It is a common problem that it is difficult to precisely control a mechanically moving scanner, although the manner of controlling the scanner varies according to the application technology. When a scanner is operated at a high speed, a scan error occurs because of the difference of scan speed between the acceleration region and the constant region. Scan error also leads to serious problems in a LIDAR sensor. In a LIDAR sensor, scan error prevents the user from achieving the objective of detecting the exact position of the object by collecting or transmitting information that is different from what the user desires. In this paper, we present a technique to compensate for the scanning error that occurs in a LIDAR sensor. In most cases where a commercial scanner is used, the user has restricted methods to control the scanner. It is difficult to control the scanner precisely because the scanner needs to be controlled using only one or two pieces of information. Also, conditions that can be changed depending on application technology may be limited. We propose a method to reduce the scan error and to improve the 3D image of the LIDAR sensor through the limited scanner control.

Crop classification is a representative problem in multispectral remote sensing image (RSI) classification, and has significance in country food security, ecological security, production estimate, crop growth supervision, and so on. It has attracted increasing attention of many researchers around the world especially after the development of convolutional neural networks (CNN). General CNN-based multispectral RSI classification methods may be not suitable for labeled samples with limited numbers and areas. Other pixel-based classification methods are always affected by noise and ignore spatial information. Focusing on these problems, this paper presents an approach based on lightened CNN for crop classification with a small number of tiny size labeled samples in multispectral images. The contribution of this work is to construct a lightened CNN model for crop classification with small samples in multispectral image. It avoids overfitting of deep CNN and reduces the requirement for the size of training samples. We adopt two-layer fully convolutional network (FCN) to extract features. The first layer uses a convolutional kernel of size 1 and outputs 16-band feature map to obtain spectral band information. Spatial information is extracted in the sequential layer using convolutional kernel of size 3, step 1 and padding 1. Thus the feature map after FCN and the labeled area have the same size. Finally, we use a fully connected layer and a softmax classifier for classification. Our experiment was conducted on 8-band multispectral image of size 50362-by-17810 pixels. There are 5 classes in the multispectral image, namely rice, soy, corn, non-crop, and uncertainty. The experimental result which achieves 86.28% accuracy indicates the good performance of our network for crop classification in multispectral RSIs.

The frequent hazy weather in North China has drawn people's attention. The anthropogenic emission by fossil fuel power plants is one of the main pollution resource, so the environmental protection administration need to monitor power plants. Thus, a power plant detection system is needed to locate power plants and judge their working status. In this paper, we propose a power plant monitoring framework based on Feature Pyramid Network (FPN) to automatically detect the chimneys and condensing towers of the power plants and judge their working status in high resolution remote sensing images (RSIs). We improve the original FPN by changing the number of layers and scales of feature pyramid to get better performance. Experimental results show that our improved FPN framework can effectively detect the chimneys and condensing towers of fossil-fuel power plants and judge their working status with mean average precision up-to 0.8591, showing good potential for power plant monitoring.

Accurate onboard-camera pose estimation is one of the challenges of satellite systems. Improving remote sensing camera pose accuracy never ceases for various applications, including autonomous navigation, 3D reconstruction and continuous city modeling. 3D products of very high spatial accuracy can be created with 3m@SE90 (3 meters error with SE90, which is the abbreviation for Spherical Error 90%) with leading companies, for example, Vricon company in USA. Aiming at the problem of the accuracy of pose estimation, a new method from captured images with the reference 3D products is presented in this paper. Distinguished from the existing methods, our method employs the 3D model to calibrate the pose of the remote sensing camera. Firstly, the high-precision 3D digital surface model is projected onto image space using a virtual calibrated camera. Then, the camera motion parameters of the neighboring moment are estimated by the information of the adjacent frames. This process consists of three steps: i) feature extraction; ii) similarity measurement, and feature matching; iii) camera pose estimation and verification. Finally, the camera pose of the captured image can be determined. Experiment results were compared with the initial exterior orientation parameters used to achieve perspective transformation of the captured images. Furthermore, the method proposed in this study is tested by hardware experiment which simulates remote sensors and platform. Results showed that acceptable accuracy of camera pose can be achievable by using the proposed approach.

In the paper is described clouds segmentation algorithm based on convolutional neural network. It has been made an analysis of existed convolutional neural networks topologies and it was made a decision of using the modifying U-Net topology. The preliminary data processing has been made taking into account a source data specific. Learning dataset has been made using real high spatial resolution remote sensing data and manual segmented clouds mask. Methodology of using learning dataset in network learning process has been proposed. Results of learned network implementation on real data are shown in the paper.