Wheat is one of the most important cereal crops in the world. Efficiently and accurately extracting spatial distribution information of winter wheat from high-resolution remote sensing images has become crucial research and application fields in agricultural monitoring. However, the complex texture features and spatial heterogeneity in high-resolution imagery often require deep learning models to be trained with a substantial number of samples to achieve the desired outcomes. In response to the challenges posed by insufficient sample volumes, which impair the accuracy of recognition in the field of extracting spatial distribution information of winter wheat, this paper introduces a new methodology that synergizes deep learning neural networks with machine learning technologies to address these issues. The new process first applied a Grey-Level Co-occurrence Matrix (GLCM) to enhance the image texture. It then utilized a random forest (RF) model to generate a confidence map of winter wheat distribution. Next, the method recommended a supervised correction to this map. It integrated the corrected map into an improved Segformer model as a priori knowledge to increase the weight of the high-confidence regions to achieve high extraction accuracy and robustness. This study selected Shouguang City, Shandong Province, as the case study and the SuperView-1 satellite data as the input image. This study compared the proposed GLCM-RF-Segformer model with ResUnet, DeepLabV3Plus, Hrnet, FCN8S, and RF to validate the experiment's effectiveness. The experimental results have shown that the new model has achieved average metrics of 91.387% for Precision, 94.467% for Recall, 94.180% for Overall Accuracy (OA), 87.053% for mean Intersection over Union (mIoU), 85.087% for Kappa coefficient, and 92.630% for F1-Score. These accuracy statistics surpass those of the five other comparative models, highlighting the model's high accuracy and robustness in extracting the spatial distribution of winter wheat with limited samples.
|