Farmland segmentation from remote sensing images using deep learning methods

Abstract

Farmland segmentation is crucial and getting increasingly important role in the field of agricultural insurance and digital agriculture. To accurately achieve insured crop area and disaster loss assessment, precision smallholders' farmland segmentation and mapping are necessary. Deep learning technology has demonstrated its strength and has out-performed state-of-the-art alternatives in many fields. In this study, we aim to explore the effectiveness of five classic deep semantic segmentation models on the smallholder-wise farmland segmentation from remote sensing images. Five FCN-based segmentation models were selected including U-Net, PSPNet, DeepLabV3+, DANet, and CCNet, which were originally proposed for natural or medical image segmentation. The study area locates in Jiaxiang County in the north China. We used GF-1 at 2m spatial resolution with 4-band multispectral images (red, green, blue, and near-infrared) acquired based on image fusion. Comparative experiment results showed that DeepLabv3+ achieved the best mIoU of 89.82%. PSPNet, DANet, CCNet and U-Net obtained lower but similar mIoU with: 89.63%, 89.66%, 89.59%, and 89.15%. The OA metric of the all the five models were above 94.8%. The results indicate that the FCN-based deep semantic segmentation networks are effective for larger-scale smallholder-wise farmland segmentation with high spatial resolution multispectral satellite images.