Agricultural Field Boundary Delineation Using a Cascaded Deep Network Model from Polarized SAR and Multispectral Images

Abstract

The accurate acquisition of farmland boundary information is an important means for agricultural production data statistics. Due to the extreme imbalance of categories in farmland boundary extraction tasks, the direct use of neural networks for boundary pixel prediction in remote sensing images often results in poor performance. This article investigates the beneficial effects of multisource remote sensing data fusion and the use of semantic segmentation to assist in boundary extraction. The research area of this work covers an area of approximately 600 km2 of farmland located in northern Xinjiang. And the obtained Sentinel-2 multispectral images covered the area, as well as Sentinel-1 dual-polarized synthetic aperture radar (SAR) images with multiple temporal phases. A novel cascaded model of semantic segmentation and edge detection is proposed for the boundary extraction task of farmland. Using the fused data of Sentinel-1 and Sentinel-2 as the raw data input of the model for semantic segmentation of farmland and nonfarmland. Then, the semantically segmented feature map is overlaid with the original data to detect the edge and extract the farmland boundary. The results show that after adding the SAR data, the extraction effect of farmland boundaries has been significantly improved. Moreover, compared with directly using neural network models for boundary extraction, using semantic segmentation information to assist in boundary extraction can achieve higher extraction accuracy. The experimental results validate the effectiveness of the proposed model and the feasibility of combining SAR remote sensing information to optimize the extraction of farmland boundaries.