Extraction of Cultivated Land from GF-2 Images Based on Level Wise Semantic Segmentation and Edge Detection
A remote sensing image farmland information extraction method based on hierarchical semantic segmentation and edge detection models is proposed,using GF-2 images as the data source,to address the problem of unclear boundaries and difficult classification extraction of sloping farmland and small area farmland fragments in mountainous and hilly areas.Firstly,select a cascading approach based on the characteristics of different types of cultivated land.Secondly,the edge of cultivated land is treated as an independent category,and combined with improved U-Net,DeeplabV3+,and DexiNed models to integrate surface and line features,so that the edge features of cultivated land can complement semantic features,thereby improving the accuracy of cultivated land extraction,suppressing complex terrain background noise,and extracting different types of cultivated land.The experimental results show that the overall accuracy,Kappa coefficient,and F1 score of the cascaded model for extracting farmland information have been improved to different degrees compared with the single model DeeplabV3+and U-Net.The cultivated land results extracted from the cascade model for different types of cultivated land are closer to the real cultivated land label,and the areas of missing and false extraction are far lower than that of the single model.
cultivated land informationsemantic segmentationedge detectionGF-2 imagehilly and mountainous area
万方数据
维普
