Cropland plot segmentation based on deep convolutional neural networks

Cropland in mountainous areas has always been an area of focus for researchers, as it serves as the foundation for agricultural production and has profound impacts on regional economy and ecological stability. This study aims to address the challenging problem of cropland parcel segmentation in remote sensing images, driven by the increasing demand for accurate and efficient land-use monitoring. To achieve precise delineation of agricultural boundaries and overall regions, we utilized deep learning methods, including the RCF (Richer Convolutional Features) model and the UNet++ architecture. The model's accuracy was enhanced using dice and weighted cross-entropy loss functions, along with data augmentation and careful dataset preparation. Our findings showed that the RCF model had an average precision of 85.64% for detecting crop edges. Meanwhile, the UNet++ model excelled, achieving metrics such as 87.14% for intersection over union and 93.03% for the F1 score. By combining the results of both models, we obtained highly accurate cropland segmentation. Overall, our approach proved highly effective for precise land-use monitoring in remote sensing.