Accurate extraction of fragmented field boundaries using classification-assisted and CNN-based semantic segmentation methods.

The accurate extraction of cropland field parcels in fragmented agricultural landscapes poses significant challenges due to the small sizes and irregular shapes of the fields. Here, we proposed a hierarchical framework for delineating agricultural field boundaries based on the level of field fragmentation. Firstly, a modified Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) was applied to delineate the agricultural field boundaries. Then, a rule-based decision fusion strategy assisted with a classification method was developed to address the under-segmentation problem in Plains-Basin scenes, and an adaptive generalized buffer algorithm (AGBA) was used to address the incomplete-segmentation problem in Plateau-Hilly scenes. We conducted tests on the proposed methods in Yangming County, Heilongjiang Province, China, using time series of Sentinel-2 images with a spatial resolution of 10-m. The results demonstrate the outstanding performance of the proposed method in delineating fragmented field boundaries, with an overall accuracy of 85.29% and F1 score of 0.84 in fragmented areas. The methods proposed in this study provide a practical solution for accurately extracting field boundaries in highly fragmented and heterogeneous agricultural landscapes.