Hierarchical extraction of cropland boundaries using Sentinel-2 time-series data in fragmented agricultural landscapes.

Accurately extracting cropland field parcels from satellite data in fragmented agricultural landscapes presents huge challenges due to small field sizes and irregular field shapes. Here, we developed a hierarchical framework for agricultural field boundary extraction from Sentinel-2 satellites based on the concept of the degree in field fragmentation. Our framework focused on three tasks, including a core task for agricultural field extraction and two auxiliary tasks to address special situations in two different fragmented regions, namely the Plains-Basin scenes with relatively regular shapes and diverse crop types, and the Plateau-Hilly scenes, with irregular shapes and small field sizes. First, the field boundaries were delineated using a modified Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net), and post-processed using morphological filtering and the Douglas-Peucker algorithm (DPA). A rule-based decision fusion strategy assisted with a classification method was developed to address the under-segmentation problem in the Plains-Basin scenes, and an adaptive generalized buffer algorithm (AGBA) was used to address the incomplete-segmentation problem in the Plateau-Hilly scenes. We tested the proposed methods in Yangming County, Heilongjiang Province of China, using time series Sentinel-2 imageries with 10-m spatial resolution. Compared with the results from using R2U-Net or a single temporal image alone, the fragmented field boundaries show an outstanding performance, with an overall accuracy of 86.42% and 84.15% and F1 score of 0.85 and 0.83 in Plains-Basin scenes and Plateau-Hilly scenes, respectively. Our proposed methods provide a feasible solution to finely extract field boundaries in highly fragmented and heterogeneous agricultural landscapes.