Delineation of agricultural fields using multi-task BsiNet from high-resolution satellite images

This paper presents a new multi-task neural network, called BsiNet, to delineate agricultural fields from high-resolution satellite images. BsiNet is modified from a Psi-Net by structuring three parallel decoders into a single encoder to improve computational efficiency. BsiNet learns three tasks: a core task for agricultural field identification and two auxiliary tasks for field boundary prediction and distance estimation, corresponding to mask, boundary, and distance tasks, respectively. A spatial group-wise enhancement module is incorporated to improve the identification of small fields. We conducted experiments on a GaoFen1 and three GaoFen2 satellite images collected in Xinjiang, Fujian, Shandong, and Sichuan provinces in China, and compared BsiNet with 13 different neural networks. Our results show that the agricultural fields extracted by BsiNet have the lowest global over-classification (GOC) of 0.062, global under-classification (GUC) of 0.042, and global total errors (GTC) of 0.062 for the Xinjiang dataset. For the Fujian dataset with irregular and complex fields, BsiNet outperformed the second-best method from the Xinjiang dataset analysis, yielding the lowest GTC of 0.291. It also produced satisfactory results on the Shandong and Sichuan datasets. Moreover, BsiNet has fewer parameters and faster computation than existing multi-task models (i.e., Psi-Net and ResUNet-a D7). We conclude that BsiNet can be used successfully in extracting agricultural fields from high-resolution satellite images and can be applied to different field settings.