JARS thanks the reviewers who served the journal in 2023.

Guest editors Kaixu Bai, Simone Lolli, and Yuanjian Yang introduce the Special Section on Integrating Remote Sensing, Machine Learning, and Data Science for Air Quality Management.

Atmospheric transmittance, turbulence, and wind play a crucial role in the field of laser atmospheric transmission. In response to the demand for comprehensive detection of atmospheric optical parameters, a LiDAR system for comprehensive measurement of atmospheric transmittance, turbulence, and wind (ACW-LiDAR) has been developed through integrated optical and mechanical design. The remote sensing measurement of atmospheric transmittance, turbulence, and wind simultaneously and along a common path has been realized by the ACW-LiDAR. By proposing a limb scanning algorithm, the problem of preference for atmospheric transmittance estimation has been overcome, and the problem of inconvenient turbulence measurement at near-surface has been effectively solved. It is also possible to obtain wind information on the laser transmission route. The experimental results based on the ACW-LiDAR system indicate that the detection distance of atmospheric transmittance is greater than 10 km @ 1064 nm. The detection distance of turbulence is greater than 10 km @ 532 nm. The detection distance of wind is greater than 4 km @ 1550 nm. The comparison between the ACW-LiDAR system and the ground meteorological automatic observation system shows that the variation trend in extinction coefficient, turbulence, and radial wind velocity is consistent, with a correlation better than 0.69, verifying the accuracy of the developed ACW-LiDAR system. The analysis of comprehensive scanning detection indicates that the three key atmospheric parameters mentioned above are interrelated and mutually influencing. So the measurement of the same time and common path of atmospheric transmittance, turbulence, and wind is of great significance. These can provide a theoretical and experimental basis for long-term observation and accumulation of atmospheric parameters, as well as the correction of atmospheric parameters.

Most health studies have used residential addresses to assess personal exposure to air pollution. These exposure assessments may suffer from bias due to not considering individual movement. Here, we collected 45,600 hourly movement trajectory data points for 185 individuals in Nanjing from COVID-19 epidemiological surveys. We developed a fusion algorithm to produce hourly 1-km PM_2.5 concentrations, with a good performance for out-of-station cross validation (correlation coefficient of 0.89, root-mean-square error of 5.60 μg / m³, and mean absolute error (MAE) of 4.04 μg / m³). Based on these PM_2.5 concentrations and location data, PM_2.5 exposures considering individual movement were calculated and further compared with residence-based exposures. Our results showed that daily residence-based exposures had an MAE of 0.19 μg / m³ and were underestimated by <1 % overall. For hourly residence-based exposures, the MAE exhibited a diurnal variation: it decreased from 0.58 μg / m³ at 09:00 to 0.44 μg / m³ at 12:00 and then continuously increased to 0.74 μg / m³ at 17:00. The biases also depended on activity types and distances from home to activity locations. Specifically, the largest MAE (3.86 μg / m³) occurred in visits that were among the top four types of activity other than being at home. As distances changed from <10 to >30 km, the degree of underestimation for hourly residence-based exposures increased from 1% to 6%. This trend was more obvious for work activities, suggesting that personal exposure assessments should consider individual movement for work cases with long commuting distances.

It is of great significance to explore the spatial-temporal variations and estimate the relative importance of the influencing factors of PM_2.5 and O₃ pollution. The study established nationwide surface O₃, NO₂, and SO₂ estimation models using the extreme gradient boosting model and the data fusion method. The cross-validation results indicated that the forecasted models performed well (R-values from 0.86 to 0.95). The results revealed that the pollution levels of O₃, PM_2.5, NO₂, and SO₂ in the North China Plain (NCP) were the highest in China. Subsequently, a multi-task learning model was utilized to estimate the relative importance of influential factors on the PM_2.5 and O₃ pollution in the NCP. The sensitivity analysis results indicated that the O₃ pollution from 2010–2020 in the NCP was susceptible to meteorological factors such as ultraviolet radiation and temperature, as well as anthropogenic precursors such as NO_X, and PM_2.5 pollution in the NCP was constrained by both meteorological factors (44.62%) and anthropogenic emissions (16.86%). The impact of NO₂ on PM_2.5 pollution was similar to its impact on O₃ pollution; therefore, the importance of NO₂ emission reduction to PM_2.5 pollution is as important as that of O₃ pollution, whereas the impact of SO₂ on PM_2.5 was much greater than its impact on O₃ pollution, so SO₂ emission reduction is more important for PM_2.5.

Air quality monitoring plays a vital role in the sustainable development of any country. Continuous monitoring of the major air pollutants and forecasting their variations would be helpful in saving the environment and improving the quality of public health. However, this task becomes challenging with the available observations of air pollutants from the on-ground instruments with their limited spatial coverage. We propose a multimodal deep learning network (M²-APNet) to predict major air pollutants at a global scale from multimodal temporal satellite images. The inputs to the proposed M²-APNet include satellite image, digital elevation model (DEM), and other key attributes. The proposed M²-APNet employs a convolutional neural network to extract local features and a bidirectional long short-term memory to obtain longitudinal features from multimodal temporal data. These features are fused to uncover common patterns helpful for regression in predicting the major air pollutants and categorization of air quality index (AQI). We have conducted exhaustive experiments to predict air pollutants and AQI across important regions in India by employing multiple temporal modalities. Further, the experimental results demonstrate the effectiveness of DEM modality over others in learning to predict major air pollutants and determining the AQI.

Fine particles with an aerodynamic diameter ≤2.5 μm are called PM_2.5, and accurate prediction of PM_2.5 concentration can help prevent the harmful effects of heavy pollution on humans. At present, the distribution of ground-based PM_2.5 monitoring stations in China’s cities is relatively sparse. Hence, the aerosol optical depth (AOD) obtained from satellite remote sensing provides an effective means for large-scale routine PM_2.5 monitoring. In this study, the multi-angle implementation of atmospheric correction AOD (1 km resolution) product from the moderate resolution imaging spectroradiometer (MODIS), meteorological factors, and ground measurements of daily PM_2.5 concentrations from 2016 to 2020 for Dalian were input into a backpropagation neural network (BPNN) to predict PM_2.5 concentrations. To improve the prediction accuracy and stability, a genetic algorithm (GA)-optimized BPNN was further established based on the BPNN to achieve a comparative PM_2.5 concentration prediction. Results showed that the BPNN and GA-BPNN achieved the PM_2.5 concentration by integrating AOD, meteorological factors, and air pollutants with the model test set R² of 0.77 and 0.83 and root mean square error (RMSE) of 11.83 and 9.80 μg m ^{− 3}, respectively. GA-BPNN decreased the SD_RMSE from 0.44 to 0.23 μg m ^{− 3} compared with BPNN and improved the model stability. The spatial distribution of annual averaged estimated PM_2.5 was predicted using GA-BPNN in Dalian from 2016 to 2020. The spatial distribution of PM_2.5 concentrations was generally consistent over 5 years, and the PM_2.5 concentrations exhibited an overall decreasing trend. The BPNN before and after optimization achieved longer-term interannual PM_2.5 daily concentration prediction, and the GA-BPNN had a better prediction effect for extreme values, handled complex fuzzy mapping relationships, and had lower computational complexity. Hence, GA-BPNN was found to be more suitable for practical applications with more advantages for PM_2.5 concentration prediction than BPNN.

Aerosol–cloud–precipitation interaction is currently a research hotspot that is challenging but also one of the most prominent sources of uncertainty affecting climate change. We have identified 1082 mesoscale convective systems (MCSs) over eastern China from April to September in 2016 and 2017. Overall, the occurrence frequency and MCS area increased when altitude increased, as demonstrated by the t-test at 95% confidence. More MCSs appeared and matured fully, although they moved slowly, in a selected urban agglomeration area compared to a selected rural area, owing to the urbanization impact. With an increase in the concentration of particulate matter with particle size below 10 μm (PM10) averaged by the first 3 h of MCS initiations, the cloud top brightness temperature and MCS area decreased, resulting in weakened precipitation intensity and a smaller MCS area. The t-test was passed with 90% confidence, confirming this finding. In addition, high-humidity circumstances can produce enough water vapor to support the creation of many higher and deeper MCSs.

Dielectric measurements of plastic explosives using a loaded waveguide technique via vector network analyzer and banded millimeter wave extender modules operating at V-band (50 to 75 GHz) are performed. A portion of an explosive sample is inserted into a waveguide shim 2 mm in length and trimmed flush with the faces of the shim. Two-port S-parameter measurements are conducted on the explosive; the empty shim is similarly characterized. Using standard waveguide equations and the measured length of the shim, the complex S-parameter data obtained with the filled shim is optimized to four free parameters—complex permittivity and distance offsets for the two sample faces relative to the calibration planes. Permittivity data obtained from measurements of the plastic explosives C-4, Primasheet 1000, Primasheet 2000 and Semtex 10 are presented. Results obtained for C-4 and Primasheet 1000 are comparable to other data in the literature, and the data on Primasheet 2000 and Semtex 10 are the first known published permittivity values in this range. Excellent agreement between the experiment and the fit is obtained using a constant permittivity across the waveguide band, indicating that dispersion is not significant for these materials.

Although deep learning-based methods have made remarkable achievements in polarimetric synthetic aperture radar (PolSAR) image classification, these methods require a large number of labeled samples. However, for PolSAR image classification, it is difficult to obtain a large number of labeled samples, which requires extensive human labor and material resources. Therefore, a new PolSAR image classification method based on multi-scale contrastive learning is proposed, which can achieve good classification results with only a small number of labeled samples. During the pre-training process, we propose a multi-scale contrastive learning network model that uses the characteristics of the data itself to train the network by contrastive training. In addition, to capture richer feature information, a multi-scale network structure is introduced. In the training process, considering the diversity and complexity of PolSAR images, we design a hybrid loss function combining the supervised and unsupervised information to achieve better classification performance with limited labeled samples. The experimental results on three real PolSAR datasets have demonstrated that the proposed method outperforms other comparison methods, even with limited labeled samples.

Forestry pests pose a significant threat to forest health, making precise extraction of infested trees a vital aspect of forest protection. In recent years, deep learning has achieved substantial success in detecting infestations. However, when applying existing deep learning methods to infested tree detection, challenges arise, such as limited training samples and confusion between forest areas and artificial structures. To address these issues, this work proposes a two-stage hierarchical semi-supervised deep learning approach based on unmanned aerial vehicle visible images to achieve the individual extraction of each pine wilt disease (PWD). The approach can automatically detect the positions and crown extents of each infested tree. The comprehensive framework includes the following key steps: (a) considering the disparities in global image representation between forest areas and artificial structures, a scene classification network named MobileNetV3 is trained to effectively differentiate between forested regions and other artificial structures. (b) Considering the high cost of manually annotating and incomplete labeling of infested tree samples, a semi-supervised infested tree samples mining method is introduced, significantly reducing the workload of sample annotation. Ultimately, this method is integrated into the YOLOv7 object detection network, enabling rapid and reliable detection of infested trees. Experimental results demonstrate that, with a confidence threshold of 0.15 and using the semi-supervised sample mining framework, the number of samples increases from 53,046 to 93,544. Precision evaluation metrics indicate a 5.8% improvement in recall and a 2.6% increase in mean average precision@.5. The final test area prediction achieves an overall accuracy of over 80% and the recall rate of over 90%, indicating the effectiveness of the proposed method in PWD detection.

In recent years, graph convolutional networks (GCNs) have attracted increased attention in hyperspectral image (HSI) classification through the utilization of data and their connection graph. However, most existing GCN-based methods have two main drawbacks. First, the constructed graph with pixel-level nodes loses many useful spatial information while high computational cost is required due to large graph size. Second, the joint spatial-spectral structure hidden in HSI are not fully explored for better neighbor correlation preservation, which limits the GCN to achieve promising performance on discriminative feature extraction. To address these problems, we propose a multiscale graph convolutional residual network (MSGCRN) for HSI classification. First, to explore the local spatial–spectral structure, superpixel segmentation is performed on the spectral principal component of HSI at different scales. Thus, the obtained multiscale superpixel areas can capture rich spatial texture division. Second, multiple superpixel-level subgraphs are constructed with adaptive weighted node aggregation, which not only effectively reduces the graph size, but also preserves local neighbor correlation in varying subgraph scales. Finally, a graph convolution residual network is designed for multiscale hierarchical features extraction, which are further integrated into the final discriminative features for HSI classification via a diffusion operation. Moreover, a mini-batch branch is adopted to the large-scale superpixel branch of MSGCRN to further reduce computational cost. Extensive experiments on three public HSI datasets demonstrate the advantages of our MSGCRN model compared to several cutting-edge approaches.

Turbulence estimation theory is presented and demonstrated by imaging a series of spatially encoded quick response (QR) codes in ambient radiation through atmospheric scintillation. This remote sensing concept was verified though preliminary feasibility experiments and detailed MATLAB simulations using QR codes displayed on a low-power digital e-ink screen. Of note, knowledge of propagation range and QR code dimensions are not required ahead of time, as each code contains information detailing its block size and overall physical size, enabling automated calculations of spatial resolution and target range. Estimation algorithms leverage the extracted resolution and range information to determine path-integrated optical turbulence, as quantified by the Fried parameter, r0. The estimation criterion is obtained by cycling a series of QR code sizes on an e-ink screen and determining the transition point at which the QR code can no longer be read, resulting in a system capable of automatically estimating path-integrated optical turbulence.

Owing to accelerated urbanization, land subsidence has damaged urban infrastructure and impeded sustainable economic and social development in Qingdao City, China. Combining interferometric synthetic aperture radar (InSAR) and generic atmospheric correction online service (GACOS), atmospheric correction has not yet been investigated for land subsidence in Qingdao. A small baseline subset of InSAR (SBAS InSAR), GACOS, and 28 Sentinel-1A images were combined to produce a land subsidence time series from January 2019 to December 2020 for the urban areas of Qingdao, and the spatiotemporal evolution of land subsidence before and after GACOS atmospheric correction was compared, analyzed, and verified using leveling data. Our work demonstrates that the overall surface condition of the Qingdao urban area is stable, and subsidence areas are mainly concentrated in the coastal area of Jiaozhou Bay, northwestern Jimo District, and northern Chengyang District. The GACOS atmospheric correction could reduce the root-mean-square error of the differential interferometric phase. The land subsidence time series after correction was in better agreement with the leveling-monitored results. It is effective to perform GACOS atmospheric correction to improve the accuracy of SBAS InSAR-monitored land subsidence over a large scale and long time series in coastal cities.

Timely and accurate prediction of winter wheat yield contributes to ensuring national food security. We propose a CNN- bidirectional gated recurrent unit method with triple attention for winter wheat yield prediction, named CBTA. This deep learning model uses convolutional neural networks to mine the spatial spectral information in hyperspectral remote sensing images. Furthermore, the bidirectional gated recurrent unit is used to adaptively learn the time dependence between the various stages of winter wheat growth. Data from Henan Province, China, is used in this study to train the model and also verify its prediction performance and stability. The results from our experiment show that our proposed model has an excellent effect on yield prediction in the county, with root-mean-square-error, mean absolute error, and R2 of 0.469 t/ha, 0.336 t/ha, and 0.827, respectively. Moreover, our findings suggested that the precision of our model using the data from sowing to heading-flowering stage was very close to that from sowing to ripening stage, which proves that the CBTA model can accurately predict the yield of winter wheat 1 to 2 months in advance.

Vehicle detection using aerial thermal infrared images has received significant attention because of its strong capability for day and night observations to supply information for vehicle tracking, traffic monitoring, and road network planning. Compared with aerial visible images, aerial thermal infrared images are not sensitive to lighting conditions. However, they have low contrast and blurred edges. Therefore, a combinational and sparse you-only-look-once (ComS-YOLO) neural network is put forward to accurately and quickly detect vehicles in aerial thermal infrared images. Therein, we adjust the structure of the deep neural network to balance the detection accuracy and running time. In addition, we propose an objective function that utilizes the diagonal distance of the corresponding minimum external rectangle, which prevents non-convergence when there is an inclusion relationship between the prediction and true boxes or in the case of width and height alignment. Furthermore, to avoid over-fitting in the training stage, we eliminate some redundant parameters via constraints and on-line pruning. Finally, experimental results on the NWPU VHR-10 and DARPA VIVID datasets show that the proposed ComS-YOLO network effectively and efficiently identifies the vehicles with a low missed rate and false detection rate. Compared with the Faster R-CNN and a series of YOLO neural networks, the proposed neural network presents satisfactory and competitive results in terms of the detection accuracy and running time. Furthermore, vehicle detection experiments under different environments are also carried out, which shows that our method can achieve an excellent and desired performance on detection accuracy and robustness of vehicle detection.

Mangroves maintain coastal balance and have the greatest potential for carbon sequestration. Most mapping studies on mangroves have focused on their extent and distribution and rarely featured mangrove species. Therefore, the objective of our study is to investigate mangrove species mapping from integrated Sentinel-2 imagery and field spectral data using a random forest (RF) algorithm. Study areas are located in East and South Lampung, Indonesia. The field samples used represented 144 points of mangrove species. The classification method used an RF algorithm and four models with varying parameters: model 1 with Sentinel-2; model 2 with both Sentinel-2 and field spectral data; model 3 with Sentinel-2, field spectral data, and spectrally transformed data; and model 4 only with spectrally transformed data. The results showed that Rhizophora mucronata, Sonneratia alba, Avicennia lanata, and Avicennia marina were the most common mangrove species in these areas, with reflectance values in the range of 0.002 to 0.493, 0.006 to 0.833, 0.014 to 0.768, and 0.002 to 0.758. Permutation importance (PI) that affects the classification model is the red band, near-infrared, and green normalized difference vegetation index, where the most PI in model 3 is 0.283. The highest level of agreement for mangrove species is found in model 3. Model 3 is the best parameter for RF classification that showed the best mapping accuracy, with the overall accuracy, user accuracy, producer accuracy, and kappa value being 81.25%, 81.68%, 81.25%, and 0.80, respectively.

The generation of synthetic multispectral satellite images has not yet reached the quality level achievable in other domains, such as the generation and manipulation of face images. Part of the difficulty stems from the need to generate consistent data across the entire electromagnetic spectrum covered by such images at radiometric resolutions higher than those typically used in multimedia applications. The different spatial resolution of image bands corresponding to different wavelengths poses additional problems, whose main effect is a lack of spatial details in the synthetic images with respect to the original ones. We propose two generative adversarial networks-based architectures explicitly thought to generate synthetic satellite imagery by applying style transfer to 13-band Sentinel-2 level1-C images. To avoid losing the finer spatial details and improve the sharpness of the generated images, we introduce a pansharpening-like approach, whereby the spatial structures of the input image are transferred to the style-transferred images without introducing visible artifacts. The results we got by applying the proposed architectures to transform barren images into vegetation images and vice versa and to transform summer (res. winter) images into winter (res. summer) images, which confirm the validity of the proposed solution.

A total solar eclipse occurred on April 20, 2023, with the umbral shadow touching the Australian continent over the Ningaloo coastal region, near the town of Exmouth, Western Australia. Eclipse totality lasted ∼1 min, reaching totality at ∼03:29 UTC and happened under cloudless skies. Here, we show that the speed of the Moon’s shadow over the land surface can be estimated from 10 min sampling in both the infrared and visible bands of the Himawari-9 geostationary satellite sensor. The cooling of the land surface due to the passage of the Moon’s shadow over the land is investigated, and temperature drops of 7 K to 15 K are found with cooling rates of 2±1.5 mK s−1. By tracking the time of maximum cooling, the speed of the Moon’s shadow was estimated from thermal data to be 2788±21 km h−1 and from the time of minimum reflectance in the visible data to be 2598±181 km h−1, with a notable time dependence. The methodology and analyses are new and the results compare favorably with NASA’s eclipse data computed using Besselian elements.

Citrus black spot (CBS) is a fungal disease caused by Phyllosticta citricarpa that poses a quarantine threat and can restrict market access to fruits. It manifests as lesions on the fruit surface and can result in premature fruit drops, leading to reduced yield. Another significant disease affecting citrus is canker, which is caused by the bacterium Xanthomonas citri subsp. citri (syn. X. axonopodis pv. citri); it causes economic losses for growers due to fruit drops and blemishes. Early detection and management of groves infected with CBS or canker through fruit and leaf inspection can greatly benefit the Florida citrus industry. However, manual inspection and classification of disease symptoms on fruits or leaves are labor-intensive and time-consuming processes. Therefore, there is a need to develop a computer vision system capable of autonomously classifying fruits and leaves, expediting disease management in the groves. This paper aims to demonstrate the effectiveness of convolutional neural network (CNN) generated features and machine learning (ML) classifiers for detecting CBS infected fruits and leaves with canker symptoms. A custom shallow CNN with radial basis function support vector machine (RBF SVM) achieved an overall accuracy of 92.1% for classifying fruits with CBS and four other conditions (greasy spot, melanose, wind scar, and marketable), and a custom Visual Geometry Group 16 (VGG16) with the RBF SVM classified leaves with canker and four other conditions (control, greasy spot, melanoses, and scab) at an overall accuracy of 93%. These preliminary findings demonstrate the potential of utilizing hyperspectral imaging (HSI) systems for automated classification of citrus fruit and leaf diseases using shallow and deep CNN-generated features, along with ML classifiers.

Climate change is a critical concern that has been greatly affected by human activities, resulting in a rise in greenhouse gas emissions. Its effects have far-reaching impacts on both living and non-living components of ecosystems, leading to alarming outcomes such as a surge in the frequency and severity of fires. This paper presents a data-driven framework that unifies time series of remote sensing images, statistical modeling, and unsupervised classification for mapping fire-damaged areas. To validate the proposed methodology, multiple remote sensing images acquired by the Sentinel-1 satellite between August and October 2021 were collected and analyzed in two case studies comprising Brazilian biomes affected by burns. Our results demonstrate that the proposed approach outperforms another method evaluated in terms of precision metrics and visual adherence. Our methodology achieves the highest overall accuracy of 58.15% and the highest F1 score of 0.72, both of which are higher than the other method. These findings suggest that our approach is more effective in detecting burned areas and may have practical applications in other environmental issues such as landslides, flooding, and deforestation.

Accurate orchard spatial distribution information can help government departments to formulate scientific and reasonable agricultural economic policies. However, it is prominent to apply remote sensing images to obtain orchard planting structure information. The traditional multidimensional remote sensing data processing, dimension reduction and classification, which are two separate steps, cannot guarantee that final classification results can be benefited from dimension reduction process. Consequently, to make connection between dimension reduction and classification, this work proposes two neural networks that fuse stack autoencoder and convolutional neural network (CNN) at one-dimension and three-dimension, namely one-dimension and three-dimension fusion stacked autoencoder (FSA) and CNN networks (1D-FSA-CNN and 3D-FSA-CNN). In both networks, the front-end uses a stacked autoencoder (SAE) for dimension reduction, and the back-end uses a CNN with a Softmax classifier for classification. In the experiments, based on Google Earth Engine platform, two groups of orchard datasets are constructed using multi-source remote sensing data (i.e., GaoFen-1, Sentinel-2 and GaoFen-1, and GaoFen-3). Meanwhile, DenseNet201, 3D-CNN, 1D-CNN, and SAE are used for conduct two comparative experiments. The experimental results show that the proposed fusion neural networks achieve the state-of-the-art performance, both accuracies of 3D-FSA-CNN and 1D-FSA-CNN are higher than 95%.

Accurate estimation of forest individual tree characteristics and biomass is very important for monitoring global carbon storage and carbon cycle. In order to solve the problem of calculating individual biomass of various tree species in complex stands, we take terrestrial laser scanning data, unmanned aerial vehicle-laser scanning data, and multispectral data as data sources and extract spectral characteristics, vegetation index characteristics, texture characteristics, and tree height characteristics of diverse forest areas through multispectral classification of tree species. Based on the random forest (RF) algorithm, the extracted features were superimposed and optimized, and the tree species were classified according to the multispectral data combined with field investigation. Then multispectral classification data combined with light detection and ranging (LIDAR) point cloud data were used to classify point cloud species, and then individual tree parameters were extracted for the divided point cloud species, and stand biomass was obtained using the tree biomass calculation model. The results showed that all kinds of tree species could be identified based on RF algorithm by combining multispectral data and LIDAR data. The overall classification accuracy was 66% and the kappa coefficient was 0.59. The recall rate of poplar, cypress, and lacebark-pine was about 75%, except for willow and clove trees, which were blocked by large crown width and caused multiple detection and missed detection. The R2 of diameter at breast height was 0.85, and the root-mean-square error (RMSE) was 5.90 cm. The R2 of the tree height was 0.90, and the RMSE was 1.78 m. Finally, the biomass of each tree species was calculated, and the stand biomass was 66.76 t/hm2, which realized the classification of the whole stand and the measurement of the biomass of each tree. Our study proves that the application of combined multisource remote sensing data to forest biomass estimation has good feasibility.

This study was based on analysis of Sentinel-1 (SAR) data to estimate the normalized difference vegetation index (NDVI) and the normalized difference water index (NDWI) during the period 2019 to 2020 in a region with a range of different land uses. The methodology adopted involved the construction of four regression models: linear regression (LR), support vector machine (SVM), random forest (RF), and artificial neural network (ANN). These models aimed to determine vegetation indices based on Sentinel-1 backscattering data, which were used as independent variables. As dependent variables, the NDVI and NDWI obtained via Sentinel-2 data were used. The implementation of the models included the application of cross-validation with an analysis of performance metrics to identify the most effective model. The results revealed that, based on the post-hoc test, the SVM model presented the best performance in the estimation of NDVI and NDWI, with mean R2 values of 0.74 and 0.70, respectively. It is relevant to note that the backscattering coefficient of the vertical-vertical (VV) and vertical-horizontal (VH) polarizations emerged as the variable with the greatest contribution to the models. This finding reinforces the importance of these parameters in the accuracy of estimates. Ultimately, this approach is promising for the creation of time series of NDVI and NDWI in regions that are frequently affected by cloud cover, thus representing a valuable complement to optical sensor data. This integration is particularly valuable for monitoring agricultural crops.

The temperature of hot mix asphalt (HMA), base, and subgrade layers plays a significant role in pavement performance, because temperature influences the strength of the materials. Therefore, a model to predict temperature throughout the entire pavement structure is desirable. However, most existing models only focus on predicting the temperature of the road surface or the HMA layer, and these models usually need some information related to boundary conditions or material properties that is difficult to obtain. This research aims to demonstrate that machine learning (ML) model can be a powerful generalized approach to predict the temperature within a pavement structure at multiple depths. Data collected from sensors (thermistors and time domain reflectometers) installed in the Integrated Road Research Facility test road in Edmonton, Alberta, Canada, were used to train ML models. Sensitivity analysis was performed to analyze the influence of several input parameters on asphalt and soil temperature. ML models with three input parameters—average daily air temperature, day of the year, and depth—resulted in better performance compared to ML models based on other combinations of parameters. Three ML models were established to predict the average daily temperature, minimum daily temperature, and maximum daily temperature of the pavement structure. To validate model performance, the three ML models were compared with four existing models, and of these the ML models showed the highest accuracy with the coefficient of determination values above than 0.97 and root mean square error values below than 2.21. These results demonstrate that ML models can be used to give accurate predictions of road temperature at multiple depths with only one environmental predictive parameter, average daily air temperature.

Long-time coherent integration is known as a powerful method to detect the weak target. However, its effectiveness is limited by the target motion across range and Doppler bins. For the high-speed target, it is highly possible that the range bin crossing (RBC) problem occurs, and for maneuvering target, the Doppler bin crossing (DBC) problem cannot be neglected. In this paper, we propose a Radon dynamic path optimization and fixed point iteration method to deal with the RBC and DBC problem, and thus make the radar able to detect the high-speed maneuvering weak target effectively. Radon transform is essentially a parameter searching method to find the target range moving path. We derive a cost function based on the property of the slow time time-frequency and frequency-time matrix, and solve it with the dynamic path optimization and fixed point iteration algorithm. The proposed method does not demand any a priori information,and is free of the ambiguity of the velocity or the acceleration caused by the potential undersampling of the slow time. Both the simulated and real Radar echo signals validate the effectiveness of the proposed method.

In recent years, methods based on convolutional neural networks (CNNs) have achieved significant results in the problem of target classification of synthetic aperture radar (SAR) images. However, the challenges of SAR image data labeling and the characteristics of CNNs relying on a large amount of labeled data for training have seriously limited the further development of this field. In this work, we propose an approach based on attention mechanism and feature complementary fusion (AFCF-CNN) to address these challenges. First, we design and construct a feature complementary module for extracting and fusing multi-layer features, making full use of limited data and utilizing contextual information between different layers to capture more robust feature representations. Then, the attention mechanism reduces the interference of redundant background information, while it highlights the weight information of key targets in the image to further enhance the key local feature representations. Finally, experiments conducted on the moving and stationary target acquisition and recognition dataset show that our model significantly outperforms other state-of-the-art methods despite severe shortages of training data.

Maize water stress from rainfall variability is a key challenge in producing rain-fed maize farming, especially in water-scarce regions, such as southern Africa. Hence, quantifying maize foliar water content variations throughout the phenological stages is valuable in detecting smallholder maize moisture stress and supporting agricultural decision-making. The emergence of unmanned aerial vehicles (UAVs) equipped with multispectral sensors offers a unique opportunity for robust and rapid monitoring of maize foliar water content and stress. The combination of near-real-time spatially explicit information acquired using UAV imagery with physiological indicators, such as equivalent water thickness (EWT) and fuel moisture content (FMC), provides viable options for detecting and quantifying maize foliar water content and moisture stress in smallholder farming systems. Therefore, we evaluated the utility of UAV-based multispectral datasets and random forest regression in quantifying maize EWT and FMC throughout the maize phenological growth cycle. Results showed that EWT and FMC could be determined using the near-infrared and red-edge wavelengths to a relative root mean square error of 2.27% and 1%, respectively. Specifically, the spectra acquired during the early reproductive growth stages between silking and milk stages demonstrated a high sensitivity to the variation in maize moisture content. These findings serve as a fundamental step toward creating an early maize moisture stress detection and warning system and contribute to climate change adaptation and resilience of smallholder maize farming.

The super-resolution algorithms of single remote sensing image (SRSI) based on generative adversarial network have made a breakthrough recently, which can effectively learn local details to generate more realistic RSIs with high resolution. However, most of them ignore the characteristics of large size and small targets, resulting in the loss of details at the edges of the generated images, resulting in a strong sense of blur. To solve these problems, this paper proposes an improved architecture named DAE²GAN based on attention mechanism and transformer. First, to process large-size RSIs, the vision transformer is selected as a discriminator to compensate the lack of global information attention of convolutional generator. At the same time, to make the generator consider small objects of RSIs better, channel attention is introduced to realize the focus on the local contour with high frequency. Then, an edge loss is designed to constrain the training process, so that the edge details of the generated image can be kept more complete. Experiments show that the proposed method can more effectively improve the reconstruction quality in visual effect of SRSI, presenting clearer and richer detailed information, and the PSNR and Structural Similarity of reconstructed images generated by the DAE2GAN are improved well, where the highest increase is 1.68/0.078 compared with the existing mainstream methods. Therefore, the proposed DAE2GAN can efficiently assisting the implementation of various remote sensing tasks, such as urban road identification, agricultural monitoring, and geological exploration.

Bathymetry, the measurement of sea depth, has conventionally been conducted using echo-sounders on vessels. However, various factors limit conventional shipborne surveys in coastal regions, including data continuity, geographic obstacles, diplomatic concerns, and marine infrastructures. Remote sensing technology can address these limitations, particularly with the advancement of satellite imaging technology. Indeed, many studies are underway to develop machine learning-based water depth estimation technologies. However, previous studies have focused on clear waters with low turbidity or uniform seabed sediment. Therefore, in this study, we developed a satellite-derived bathymetry (SDB) model using the random forest machine learning algorithm, which was applied to three coastal areas around the Korean Peninsula with distinct characteristics: clear waters (Samcheok), high turbidity (Cheonsuman), and varied seabed sediments (Hallim). We then compared the accuracy of the bathymetric mapping data derived in these three areas. The estimated depth values exhibited the highest accuracy in Samcheok, followed by Hallim and Cheonsuman. Based on Worldview-3 images and on-site surveys, we confirmed the presence of basalt on the seabed. However, the remote reflectance was attenuated due to the effect of the black rock, leading to an overestimation of the depth. In the future, additional satellite images will be applied as training data for the machine learning model to advance the SDB technology using turbidity and seabed sediment distribution data for each area. Ultimately, the SDB results will be applied as depth monitoring data to facilitate safe ship passage in coastal areas, including ports that require periodic and consistent coastal bathymetry. In addition, they can be applied as input data for numerical ocean models, contributing to various fields.

Natural disasters are calamitous events with causes and effects that need to be examined. Among them, floods are one of the most devastating natural disasters that demand constant monitoring. The flood in Pakistan in 2022 caused massive destruction of lives and property due to heavy rainfall. The utilization of remote sensing technology is crucial for managing these disastrous events. This research was carried out to map class-level damages caused by the 2022 Pakistan flood using the fuzzy-based modified possibilistic c-means (MPCM) classifier with the individual sample as mean (ISM) training approach. This study targeted seven different flood-affected classes: fields with crop submerged by August 8 and 29, 2022; fields with crop and later fallow submerged by August 8 and 29 2022; open areas with vegetation submerged by August 29, 2022; roads and parks submerged by August 29, 2022; and ponds submerged by August 29, 2022. The temporal class-based sensor independent normalized difference vegetation index (CBSI-NDVI) database prepared using pre-flood, flood, and post-flood images of PlanetScope was used for this research work. The classified outputs were evaluated using Mean Membership Difference (MMD) and F-score. Based on the MMD values of 0.01 to 0.02, and an F-score of nearly 1, the classifier indicated a good classification result. This study claimed that the MPCM algorithm with the ISM training approach handled heterogeneity within the class and produced effective flood-based damage extraction results. Furthermore, it was found that the utilization of temporal images reduced the spectral overlap between classes. This research offers insights to study and better understand the applicability of the MPCM algorithm with the ISM training approach in disaster studies. Moreover, this technique of assessing damaged classes is an ingenious single-step class-specific approach performed with optimized temporal images.

The fungus Pestalotiopsis sp. causes aberrant leaf fall, resulting in an annual reduction in rubber production. The loss of latex yield has a significant impact on the income of smallholder farmers. The objective of this study is to estimate NDVI using reflectance from the upper canopy based on Sentinel-2 NDVI time series data and daily climate data, analyze phenological changes of rubber trees using a harmonic model and time-lag cross-correlation, and develop a prediction model of vegetation indices using the random forest regressor (RFR) algorithm. The study was conducted on the rubber plantation of the Indonesian Rubber Research Institute in Sembawa, Indonesia. Three rubber clones aged 10 to 11 years were used to study the wintering trend. The study found that the overall trend of the NDVI decreased significantly from 2019 to 2022. The cycle of defoliation and refoliation changed due to disease and climate change. The time lag effects of vegetation index and climate variables are crucial for predicting vegetation dynamics. Increasing temperature and changing precipitation play significant roles in influencing pathogen incidence. The mean absolute percentage error (MAPE) was used to evaluate the trend of the vegetation indices. The RFR algorithm predicts the vegetation indices with a MAPE of 6.07%, 5.96%, and 8.18% of BPM24, GT1, and RRIC 100, respectively. Our finding indicated that the analysis and prediction model can be used to understand the phenological pattern and predict the wintering pattern to prevent disease spread and minimize latex yield loss due to infestation outbreaks.

Recently, high spatial resolution remote sensing image scene classification has had a wide range of applications and has become one of the hotspots in the field of remote sensing research. Due to the complexity of the scenes in remote sensing images, it is impossible to annotate all ground object classes at once. To adapt to different application scenarios, high spatial resolution remote sensing image scene classification models need to have zero-shot generalization ability for unseen classes. To improve the zero-shot generalization ability of classification models, the existing methods often start from the perspective of image features, thus ignoring the high-order semantic information in the scene. In fact, the association between higher-order semantic information in the scene is very important for the generalization ability of the classification model. People often use image information and its corresponding higher-order semantic information to complete remote sensing image scene understanding. Therefore, this work proposes a text guided remote sensing image pre-training model for zero-shot classification of high spatial resolution remote sensing image scenes. First, the transformer model is used to extract the embedded features of text and remote sensing images. Then, based on the aligned text and remote sensing image data, a contrast learning method is used to train the model to learn the correspondence between text and image features. After the model training is completed, the nearest neighbor method is used to complete zero-shot classification on the target data. The effectiveness of the proposed method was verified on three remote sensing image scene classification benchmark datasets.

The changes in ground roads, buildings, and occurrences of natural disasters lead to mismatches between the actual ground conditions and existing maps. Through style transfer between real-time remote sensing images and maps, map content can be rapidly generated and updated. However, in existing methods for generating maps from remote sensing images based on SmapGAN, we first found that using ResBlock as the style conversion module fails to establish long-distance relationships between features. In addition, the small receptive field of convolution layers in ResBlock leads to poor global information capture, resulting in inferior image restoration during upsampling. Second, using transpose convolution as the upsampling method can result in the issue of blurred content in the generated maps. To address these problems, we propose corresponding improvements: on one hand, a style conversion module combining multi-headed self-attention (MHSA) with residual modules, named MHSA-ResBlock, is introduced to address the difficulty in capturing long-distance relationships between features when dealing with a large number of pixel features, and to better capture global information in images. On the other hand, an upsampling method combining transpose convolution with the CARAFE upsampling operator, named TC-Carafe, is proposed to tackle the issues of content loss and blurring associated with traditional transpose convolution upsampling. Furthermore, experimental results show that the MHSA-ResBlock establishes inter-pixel feature relationships and leverages the advantages of fine-grained upsampling operations with TC-Carafe, thereby utilizing inter-pixel feature relationships and neighborhood information to further improve the quality of map generation. Compared to SmapGAN, our research method has shown improvements of 0.6133 and 0.0042 in PSNR and SSIM, respectively. In addition, it has reduced RMSE by 0.72, outperforming SmapGAN in all metrics.

Image deblurring has the goal of restoring a sharp image from a degraded one. Currently, most deblurring methods are designed for natural images; these methods may not perform well when applied to remote sensing images. There are many differences between remote sensing and natural images, e.g., shooting distance, content complexity, and clarity. Therefore, a blind motion deblurring method specifically designed for remote sensing images called dual scale parallel spatial fusion network (DSPF-Net) is proposed. It has three innovative aspects: a dual-scale connection module is added between the two scales of the bottleneck layer and the decoder to realize the fusion of spatial detail and semantic features. Second, an adaptive spatial selection module is designed, which adds the function of selecting global and local spatial features. Finally, the cross-scale fusion (CSF) module is designed to restore the edge details and main structures by fusing the multi-scale features between the encoder and decoder. Extensive experiments are established on the synthetic dataset Blur-RS, the averaged peak signal-to-noise ratio and structural similarity are improved by 0.7916% and 0.0265%, respectively, compared to the best-performing comparison method. It shows that DSPF-Net has advantages in the task of blind motion deblurring of remote sensing images.

For the evaluation of traffic infrastructure, asphalt pavement aging conditions are crucial. Due to the complexity of identifying and monitoring asphalt pavement aging conditions, many current studies tend to use satellite remote sensing methods. We conducted an extraction experiment on the aging status of asphalt pavement using on-site measured Pavement Surface Condition Index data and Gaofen-2 satellite (GF-2) high-resolution remote sensing images based on comprehensive references to previous research results. Based on our experimental results, the difference health index, ratio health index, and normalized difference health index can reflect asphalt pavement aging to varying degrees, but the correlation is relatively weak. The purpose of this paper is to propose a new asphalt pavement aging index (PAI), namely the PAI, based on sufficient experimental analysis. In addition to identifying asphalt pavement aging perfectly, PAI has a good ability to discriminate between road interference information, such as shadows and vehicles, after it has been verified. There is a significant linear relationship between its correlation coefficient R and pavement surface condition index, which is 0.894. The evaluation results of three sets of ground verification points obtained by applying PAI also demonstrate its practicality. Therefore, the combination of PAI and GF-2 high-resolution remote sensing images can be used to evaluate the aging status of asphalt pavements.

Estimating the accurate position of the shoreline over time can be challenging. We propose a new approach to delimit the shoreline from remotely sensed images simply and accurately. This approach treats pixels as simultaneous radiometric measurements and, by examining the distance between iso-radiometry lines, assumes the shoreline in the position where the radiometric behavior changes more suddenly. The analyses conducted applying this coastline extraction approach underscore its significant flexibility. Specifically, the approach yields promising outcomes even if applied to a variety of different and dissimilar types of remotely sensed products. Indeed, the approach’s intrinsic straightforwardness and low computational load qualify it as a promising tool for time-series production.

Forest stands are the basic units of forest management, whose delineation is a fundamental and significant step for creating stand-level field inventories. Traditional and recent methods for delineation often rely on expensive data and suffer from limited spatial coverage. The Sentinel-2 mission has special bands aimed at vegetation mapping and provides the highest spatial resolution publicly available imagery of the globe. We aim to take the first step for using solely the Sentinel-2 imagery for forest stand delineation. A superpixel segmentation scheme based on a modified version of the simple linear iterative clustering (SLIC) superpixel algorithm, Gaussian processes regressor SLIC (GPR-SLIC) is employed with a feature set composed of Sentinel-2 time-series spectral bands and vegetation indices. GPR-SLIC replaces the spectral distance metric in the original SLIC algorithm with a Gaussian processes regressor, which is trained using the Sentinel-2 feature set and a ground truth forest stand map. GPR-SLIC, a modified SLIC algorithm, leverages multi-temporal Sentinel-2 spectral bands and vegetation indices as features. It replaces the original SLIC’s spectral distance metric with a trained GPR. Original SLIC, simple non-iterative clustering, linear spectral clustering, and entropy rate superpixels algorithms are employed for producing reference segmentations based on the LiDAR-derived canopy height model. The findings demonstrate that by incorporating Sentinel-2 time-series features, the GPR-SLIC algorithm yields more precise results for boundary adherence performance metrics and undersegmentation error. The results underscore the potential of employing a machine learning-based superpixel segmentation algorithm with multi-temporal Sentinel-2 images for effectively delineating forest stands.

The extraction of burned areas and the monitoring of forest type distribution are often affected by image classification methods. We aim to compare two image classification methods, convolutional neural network (CNN) and support vector machine (SVM), for identification of forest types and burned areas. A single post-fire Landsat 8 OLI image, forest management inventory data, and forest fire data were used to determine the optimal sample dataset. The CNN utilized PSPNet for training, while the ResNet34 served as the skeleton network to identify burned areas and forest types simultaneously. To compare and evaluate the effectiveness of the CNN model, the SVM was also used to classify the Landsat 8 OLI image with the same amount of sample data. The results indicate that the CNN model for per-pixel classification of seven classes (burned area, coniferous forest, broadleaved forest, mixed forest, residential area, water, and the other class) achieved an overall accuracy of 92.25% with a kappa coefficient of 0.8823. In contrast, the overall accuracy of the SVM algorithm was 86.72%, with a kappa coefficient of 0.8219. The results suggest that the CNN can achieve a higher classification accuracy than the SVM, and that the CNN is more reliable to support forest resources monitoring and management after a fire.

Sediment plumes are generated from both natural and human activities in benthic environments, increasing the turbidity of the water and reducing the amount of sunlight reaching the benthic vegetation. Seagrasses, which are photosynthetic bioindicators of their environment, are threatened by chronic reductions in sunlight, impacting entire aquatic food chains. Our research uses unmanned aerial vehicle (UAV) aerial video and imagery to investigate the characteristics of sediment plumes generated by a model of anthropogenic disturbance. The extent, speed, and motion of the plumes were assessed as these parameters may pertain to the potential impacts of plume turbidity on seagrass communities. In a case study using UAV video, the turbidity plume was observed to spread more than 200 ft over 20 min of the UAV campaign. The directional speed of the plume was estimated to be between 10.4 and 10.6 ft/min. This was corroborated by observation of the greatest plume turbidity and sediment load near the location of the disturbance and diminishing with distance. Further temporal studies are necessary to determine any long-term impacts of human activity-generated sediment plumes on seagrass beds.

Unmanned aerial vehicles (UAVs) exhibit the ability to flexibly conduct aerial remote-sensing imaging. By employing deep learning object-detection algorithms, they efficiently perceive objects, finding widespread application in various practical engineering tasks. Consequently, UAV-based remote sensing object detection technology holds considerable research value. However, the background of UAV remote sensing images is often complex, with varying shooting angles and heights leading to difficulties in unifying target scales and features. Moreover, there is the challenge of numerous densely distributed small targets. In addition, UAVs face significant limitations in terms of hardware resources. Against this background, we propose a lightweight remote sensing object detection network (LRSNet) model based on YOLOv5s. In the backbone of LRSNet, the lightweight network MobileNetV3 is used to substantially reduce the model’s computational complexity and parameter count. In the model’s neck, a multiscale feature pyramid network named CM-FPN is introduced to enhance the detection capability of small objects. CM-FPN comprises two key components: C3EGhost, based on GhostNet and efficient channel attention modules, and the multiscale feature fusion channel attention mechanism (MFFC). C3EGhost, serving as CM-FPN’s primary feature extraction module, possesses lower computational complexity and fewer parameters, as well as effectively reducing background interference. MFFC, as the feature fusion node of CM-FPN, can adaptively weight the fusion of shallow and deep features, acquiring more effective details and semantic information for object detection. LRSNet, evaluated on the NWPU VHR-10, DOTA V1.0, and VisDrone-2019 datasets, achieved mean average precision of 94.0%, 71.9%, and 35.6%, with Giga floating-point operations per second and Param (M) measuring only 5.8 and 4.1, respectively. This outcome affirms the efficiency of LRSNet in UAV-based remote-sensing object detection tasks.

The change detection in built-up areas within very high resolution synthetic aperture radar images is a very challenging task due to speckle noise and geometric distortions caused by the unique imaging mechanism. To tackle this issue, we propose an object-based coarse-to-fine change detection method that integrates segmentation and uncertainty analysis techniques. First, we propose a multi-temporal joint multi-scale segmentation method for generating multi-scale segmentation masks with hierarchical nested relationships. Second, we use the neighborhood ratio detector and Jensen–Shannon distance to produce both pixel-level and object-level change maps, respectively. These maps are fused using the Demeter–Shafer evidence theory, resulting in an initial change map. We then apply a threshold to classify parcels within the initial change map into three categories: changed, unchanged, and uncertain. Third, we perform uncertainty analysis and implement progressive classification by support vector machine for uncertain parcels, moving from coarse to fine segmentation levels. Finally, we integrate change maps across all scales to obtain the final change map. The proposed method is evaluated on three datasets from the GF-3 and ICEYE-X6 satellites. The results show that our approach outperforms alternative methods in extracting more comprehensive changed regions.

Domain adaptation (DA) aims to reduce the effects of the distribution gap between the source domain where a model is trained and the target domain where the model is deployed. When a deep learning model is deployed on an aerial platform, it may face gradually degrading weather conditions during its operation, leading to gradually widening gaps between the source training data and the encountered target data. Because there are no existing datasets with gradually degrading weather, we generate four datasets by introducing progressively worsening clouds and snowflakes on aerial images. During deployment, unlabeled target domain samples are acquired in small batches, and adaptation is performed continually with each batch of incoming data, instead of assuming that the entire target dataset is available. We evaluate two continual DA models against a baseline standard DA model under gradually degrading conditions. All of these models are source-free, i.e., they operate without access to the source training data during adaptation. We utilize both convolutional and transformer architectures in the models for comparison. In our experiments, we find that continual DA methods perform better but sometimes encounter stability issues during adaptation. We propose gradient normalization as a simple but effective solution for managing instability during adaptation.

Synthetic aperture radar (SAR) image change detection has been widely applied in a variety of fields as one of the research hotspots in remote sensing image processing. To increase the accuracy of SAR image change detection, an algorithm based on saliency detection and an attention capsule network is proposed. First, the difference image (DI) is processed using the saliency detection method. The DI’s most significant regions are extracted. Considering the saliency detection characteristics, we select training samples only from the DI’s most salient regions. The regions in the background are omitted. This results in a significant reduction in the number of training samples. Second, a capsule network based on an attention mechanism is constructed. The spatial attention model is capable of extracting the salient characteristics. Capsule networks enable precise classification. Finally, a final change map is obtained using capsule network to classify images. To compare the proposed method with the related methods, experiments are carried out on four real SAR datasets. The results show that the proposed method is effective in improving the exactitude of change detection.

In recent years, addressing spectral variability in hyperspectral data has improved blind hyperspectral unmixing performance and gained attention in endmember detection applications. Current approaches to address the problem of spectral variability associate the variabilities with the valid endmember and attempt to mitigate the ill-effects caused by them. However, intrinsic variabilities induced by material-specific compositional changes are crucial for identifying within-class materials like diverse soil types, forest species, and urban areas. Despite this significance, no studies have attempted a direct implementation to explicitly identify the intrinsic variants of an endmember. In this paper, we propose a framework to solve two important problems: first, to separate the intrinsic variants from illumination-based variants, and second, to simultaneously estimate the number of intrinsic variants and extract their spectral signatures without any knowledge of the number of such sources. The proposed method utilizes a spectral analysis technique with local minima/maxima to remove illumination-based variabilities, followed by a simplex-volume maximization-based reordering of potential endmembers and an iterative reconstruction error-based technique to simultaneously count the number of intrinsic variants and capture their signatures. The approach is validated on synthetic and real datasets, showcasing comparable results with state-of-the-art methods.

Remote sensing images are characterized by complex feature backgrounds and large target scale differences, so object detection for remote sensing images is a challenging problem. This work proposes a one-stage structure remote sensing image object detection model called GODANet. First, the GODANet incorporates a Global Context Network (GCNet) in the feature extraction structure. The GCNet focuses the model on the image region of interest from a global perspective. Second, the output layer utilizes an omni-dimensional dynamic convolution technique, allowing for more flexible adaptation to targets or edges in specific regions. Finally, an adaptive spatial feature fusion structure, IR-ASFF, which fuses improved-RFB (IRFB) modules is proposed to fuse the critical information of multiple levels of features to realize the adaptability to object detection at different scales. The GODANet efficiently aggregates network performance and possesses two main advantages: adaptability to multi-scale targets and focus on features of interest. The mean average precision (mAP) on the DIOR dataset and the NWPU VHR-10 dataset reached 93.7% and 92.9%, respectively, and compared with YOLOv7, the mAP was improved by 3.1% and 0.3%, respectively. Therefore, we believe the GODANet suits remote sensing image detection tasks.

In high performance optical instruments, post-processing methods are employed to reduce the stray light (SL) level below what can be achieved by design. Often, algorithms are limited to SL coming from sources located inside the field of view (FOV) of the instrument; however, SL could also come from sources located outside the FOV. We describe a method for correcting out-of-field SL, in particular for the case of the Earth observation instrument Metop-3MI. The proposed approach is a variant of the in-field (IF) method developed previously. We estimate the out-of-field SL by a linear combination of the SL kernels modulated by the input scene radiance. The correction is computed for fields on a regularly spaced polar grid, providing reasonable variations of the solid angle sustained around individual fields. Compared with the case of IF SL correction, a difficulty is that the out-of-field input radiance is unknown. We implement a mirroring technique, which is shown to be effective in most situations. This method will be used to correct the in-flight data of Metop-3MI, hence providing an SL level sufficiently low to fulfill the mission requirement.

Hyperspectral image (HSI) classification aims to assign labels to pixels to be classified. The high-dimensional form of HSI and the introduction of spatial information can introduce challenges such as redundant spectral bands and interference pixels. Recently, many methods based on convolutional neural networks and attention mechanisms have been employed to address these issues. However, existing methods often fail to adequately utilize spatial information and exhibit interference pixel diffusion, which may result in these methods being unable to extract discriminative features. In addition, convolutions are not capable of extracting robust features, which results in the poor robustness of recent methods when HSI is rotated. To address these challenges, a center-similarity spectral-spatial attention network model (CS3AN) is proposed. First, a center-similarity spectral attention (CSSpeA) module is proposed to exploit the input spatial information rationally to better reduce the impact of redundant bands. Next, a center-similarity rectified spatial attention (CSRSpaA) module that selectively weighs neighboring pixels based on interference pixels to prevent the diffusion phenomenon is provided. Second, a similarity spatial aggregation module is employed to extract robust spectral-spatial features. Finally, the robust features are input into the softmax to get the label. The validity of the CS3AN is validated on three different databases.

The classification of synthetic aperture radar (SAR) image scenes is widely used in military and civilian fields. However, some SAR images pose challenges in feature extraction due to complex scenes and low image resolution, making accurate SAR image classification difficult. To address this, an SAR image scene classification method based on the multi-layer fuzzy convolutional gray level co-occurrence matrix network (ML-FGLCMNet) approach is proposed. The multi-layer convolutional network module in this method can extract low-resolution features from SAR images, and the fuzzy gray-level co-occurrence matrix module combined four sets of second-order statistics can collect texture features from multiple angles. These features, along with the multi-class support vector machine classifier, are used to achieve SAR image scene classification. The Terra SAR-X and GS-SAR6 datasets were used for experimental analysis and comparison with the full-band channel attention network, multi-feature fusion global-local convolutional network and our method of extracting features with only the multi-layer convolutional network. The experimental results indicate that our method, ML-FGLCMNet, has good accuracy and Kappa index, which can reach more than 90% and 80%, respectively. In the confusion matrix, the classification accuracy of various SAR remote sensing scenes is higher than 0.86. It achieves smaller initial loss, higher training accuracy, and faster convergence compared to the comparative approaches.

Crop classification is necessary to extract information about the crop, such as its type, the area in crop that may be used to estimate health, and the yield of the crop. Remote sensing is a technique used to extract earth surface information from satellite and aerial imagery. However, to extract crop parameters accurately requires high spatial and spectral information of the data, which can be achieved through image fusion. We focus on crop classification by selecting the suitable band combination of satellite imagery and unmanned aerial vehicle (UAV) data through image fusion technique. Various fusion techniques are available to produce high spatial and spectral resolution data. Principal component analysis PAN sharpening method is used to fuse Sentinel-2A and UAV data for crop classification, and the best combinations of different bands are assessed based on the image quality metrics and classification accuracy. Random forest classification algorithm is performed to classify the fused images and the classification accuracy is assessed by using the F-score, Kappa coefficient, and overall accuracy. The best accuracy of 93.14% and a kappa value of 0.87 is achieved using random forest classification technique using fused images of red band of UAV imagery with BGRNIR (blue, green, red, and NIR) bands of Sentinel-2A imagery is high compared to other band combinations.

Distributed acoustic sensing (DAS) using fiber optic cables over an extensive length of railroads is a well-suited technique for condition monitoring (CM) of railroads. Regardless of the type of indication in railroad CM, the original large and noisy dataset from the DAS system is a major challenge in DAS data analysis. Different data analysis strategies, such as conventional peak finding or neural networks, can be considered for DAS data analysis depending on the purpose of the study and characteristics of the railroad. We aim to investigate the robustness of deep learning (DL) models based on long-shot-term memory (LSTM) and gated recurrent unit (GRU) approaches. The average trend of the recorded technical data management streaming signals was used to extract the train presence or absence conditions along the railroad. This investigation showed that DL approaches could be efficient for DAS signal processing and CM in railroad infrastructures and can be expanded in the future for other CM purposes such as flaw detection. Meanwhile, for train position monitoring, the proposed model based on the GRU architecture indicated a 94% detection rate compared with 93% by the LSTM model. In all, the proposed models show promising potential for efficiently detecting railroad conditions, such as anomalies and flaws that require further investigation.

Remote sensing change detection refers to the process of identifying and extracting changes in objects within the same geographical region over multiple periods. With the increasing spatial resolution of remote sensing images, the detection of minor changes has become a challenging task. We introduce a multilevel feature aggregation and enhancement network to tackle this issue. Specifically, we propose a multilevel feature aggregation module to aggregate the distinct features extracted from each image, which strengthens the feature representation capability. Subsequently, a difference parallel mapping module is designed to perceive information at different scales by refining the fused features. In addition, our guided change enhancement module captures local and long-range dependencies in multilevel features, improving the network’s accuracy in identifying changing regions. Based on a basic shared weight Siamese backbone without complex structures, our model outperforms other state-of-the-art methods on three datasets in terms of both efficiency and effectiveness.

This work presents a method for remote sensing object detection (RSOD) based on target feature enhancement and bounding box (Bbox) auxiliary regression. Due to the characteristics of dense distribution, easy feature loss, and difficult Bbox regression (ground truth boxes of medium and small objects in remote sensing images usually only contain a few pixel sizes, making they difficult to regress from the global image), the problem of low accuracy of RSOD arises. Especially, it is easy to lose the features of medium and small objects in RSOD. This work proposes a target feature enhancement module, which enhances the feature transmission of medium and small objects by mapping deep and shallow features. Furthermore, considering that the existence of low-quality labeled data in remote sensing datasets will affect the training process of detection methods, this work proposes a wise intersection over union (Wise-IoU) loss. The Wise-IoU loss focuses on important ordinary-quality labels and improves the overall performance of RSOD. To solve the problem of difficult Bbox regression caused by the small size in remote sensing objects, this paper also proposes a coarse-to-fine Bbox regression model. The new model improves the regression speed and accuracy of medium and small object Bbox by using the auxiliary IoU loss. In addition, the validity and versatility of the method were verified on the Min-DOTA dataset, UCAS-AOD dataset, RSOD dataset, and Min-AI-TOD dataset. The results show that compared with other methods (see, e.g., FCOS, YOLOx, YOLOv7, and YOLOv8), our method has better detection performance and meets real-time detection requirements.

Fusing LiDAR point cloud and camera image for 3D object detection in autonomous driving has emerged as a captivating research avenue. The core challenge of multimodal fusion is how to seamlessly fuse 3D LiDAR point cloud with 2D camera image. Although current approaches exhibit promising results, they often rely solely on fusion at either the data level, feature level, or object level, and there is still a room for improvement in the utilization of multimodal information. We present an advanced and effective multimodal fusion framework called EPAWFusion for fusing 3D point cloud and 2D camera image at both data level and feature level. EPAWFusion model consists of three key modules: a point enhanced module based on semantic segmentation for data-level fusion, an adaptive weight allocation module for feature-level fusion, and a detector based on 3D sparse convolution. The semantic information of the 2D image is extracted using semantic segmentation, and the calibration matrix is used to establish the point-pixel correspondence. The semantic information and distance information are then attached to the point cloud to achieve data-level fusion. The geometry features of enhanced point cloud are extracted by voxel encoding, and the texture features of image are obtained using a pretrained 2D CNN. Feature-level fusion is achieved via the adaptive weight allocation module. The fused features are fed into a 3D sparse convolution-based detector to obtain the accurate 3D objects. Experiment results demonstrate that EPAWFusion outperforms the baseline network MVXNet on the KITTI dataset for 3D detection of cars, pedestrians, and cyclists by 5.81%, 6.97%, and 3.88%. Additionally, EPAWFusion performs well for single-vehicle-side 3D object detection based on the experimental findings on DAIR-V2X dataset and the inference frame rate of our proposed model reaches 11.1 FPS. The two-layer level fusion of EPAWFusion significantly enhances the performance of multimodal 3D object detection.

We present an approach for grassland management using uncrewed aerial vehicles (UAV) LIDAR data and statistical modeling techniques integrated within a software-based multi-level information system (SMI). The primary objective is to utilize UAV LiDAR data and statistical modeling techniques within an SMI to accurately estimate compressed sward height (CSH) and above-ground biomass for precision farming applications. As a case study, four UAV LiDAR flights were conducted over rotational grazing farmland, and the collected data were processed to a point cloud. A statistical model was developed to estimate CSH values (R2=0.59, RMSE = 5.9 cm) using LiDAR metrics of the point cloud data. In addition, destructive sampling of grassland facilitated the calibration process, enabling the modeling of biomass based on the CSH values, specifically expressed as above-ground herbage dry biomass (R2=0.89, RMSE=0.2669 Mg ha−1). The collected data further enabled the approximation of biomass across the entire area of interest, which covered ∼200 ha, utilizing a 2.5×2.5 m polygon grid. The data were subsequently transferred to an SMI, which operates on the same grid and complements the information, thus offering a comprehensive foundation for decision-making, optimizing grazing systems, and efficient resource allocation. We contribute to advancing precision farming and sustainable grassland management.

Utilizing multisource remote sensing images to accurately estimate tropical cyclone (TC) intensity is crucial and challenging. Traditional approaches rely on a single image for intensity estimation and lack the capability to perceive dynamic spatiotemporal information. Meanwhile, many existing deep learning methods sample from a time series of fixed length and depend on computation-intensive 3D feature extraction modules, limiting the model’s flexibility and scalability. By organically linking the genesis and dissipation mechanisms of a TC with computer vision techniques, we introduce a spatiotemporal fusion convolutional neural network that integrates three distinct improvement approaches. First, an a priori aware nonparametric fusion module is introduced to effectively fuse key features from multisource remote sensing data. Second, we design a scale-aware contraction–expansion module. This module effectively captures detailed features of the TC by connecting information from different scales through a weighted and up-sampling method. Finally, we propose a 1D–2D conditional sampling training method that balances single-step regression (for short sequences) and latent-variable-based temporal modeling (for long sequences) to achieve flexible spatiotemporal feature perception, thereby avoiding the data scale constraint imposed by fixed sequence lengths. Through qualitative and quantitative experimental comparisons, the proposed spatiotemporal fusion convolutional neural network achieved a root-mean-square error of 8.89 kt, marking a 29.7% improvement over the advanced Dvorak technique, and its efficacy in actual TC case analyses indicates its practical viability and potential for broader applications.

Reservoirs are fundamental infrastructures for the management of water resources. Constructions around them can negatively impact their water quality. Such constructions can be detected by segmenting man-made objects around reservoirs in the remote sensing (RS) images. Deep learning (DL) has attracted considerable attention in recent years as a method for segmenting the RS imagery into different land covers/uses and has achieved remarkable success. We develop an approach based on DL and image processing techniques for man-made object segmentation around the reservoirs. In order to segment man-made objects around the reservoirs in an end-to-end procedure, segmenting reservoirs and identifying the region of interest (RoI) around them are essential. In the proposed two-phase workflow, the reservoir is initially segmented using a DL model, and a postprocessing stage is proposed to remove errors, such as floating vegetation in the generated reservoir map. In the second phase, the RoI around the reservoir (RoIaR) is extracted using the proposed image processing techniques. Finally, the man-made objects in the RoIaR are segmented using a DL model. To illustrate the proposed approach, our task of interest is segmenting man-made objects around some of the most important reservoirs in Brazil. Therefore, we trained the proposed workflow using collected Google Earth images of eight reservoirs in Brazil over two different years. The U-Net-based and SegNet-based architectures are trained to segment the reservoirs. To segment man-made objects in the RoIaR, we trained and evaluated four architectures: U-Net, feature pyramid network, LinkNet, and pyramid scene parsing network. Although the collected data are highly diverse (for example, they belong to different states, seasons, resolutions, etc.), we achieved good performances in both phases. The F1-score of phase-1 and phase-2 highest performance models in segmenting test sets are 96.53% and 90.32%, respectively. Furthermore, applying the proposed postprocessing to the output of reservoir segmentation improves the precision in all studied reservoirs except two cases. We validated the prepared workflow with a reservoir dataset outside the training reservoirs. The F1-scores of the phase-1 segmentation stage, postprocessing stage, and phase-2 segmentation stage are 92.54%, 94.68%, and 88.11%, respectively, which show high generalization ability of the prepared workflow.

Rapeseed is an essential oil crop and the third major source of edible oil in the world. Accurate estimation of rapeseed phenotypic traits at field scale is important for precision agriculture to improve agronomic management and ensure edible oil supply. Unmanned aerial vehicle (UAV) remote sensing technology has been applied to estimate crop phenotypic traits at field scale. Machine learning is one of the main methods to develop estimation models for phenotypic traits based on UAV data. However, the accuracy and adaptability of machine learning estimation models are constrained by the representativeness of the training data. Here, we explored the influence of growth stage and crop conditions on the estimation of rapeseed phenotypic traits by machine learning and provided an optimized strategy to construct training data for improving the estimation accuracy. Four machine learning methods were employed, including partial least squares regression, support vector regression (SVR), random forest (RF), and artificial neural network (ANN), with SVR showing the best performance in estimating rapeseed phenotypic traits. The models established for a certain cultivar, planting site, or planting density had low estimation accuracies for other cultivars, planting sites, and planting densities during the entire growth period. The results showed that cultivar and planting site had an unquantifiable influence on phenotypic traits. Integration of stratified sampling and developing estimation models for different growth stages respectively can improve the estimation accuracy for different cultivars and planting sites during the entire growth period. Planting density exhibited a quantifiable influence on phenotypic traits, and the construction of training data with samples of both low and high planting densities could improve the estimation accuracy for different planting densities. Overall, optimization of the training data by considering the influence of crop conditions on phenotypic traits can improve the estimation accuracy of rapeseed phenotypic traits based on machine learning.