A deep learning method for individual arable field (IAF) extraction with cross-domain adversarial capability

Abstract

The monitoring of agricultural fields from remote sensing imagery allows for the effective management of agricultural resources at large spatial scales. Deep learning-based methods have shown great potential in the extraction of individual arable fields (IAF) from high-resolution imagery; however, accurate boundary localization and collection of adequate training samples remain a challenge. This study proposes a novel deep learning network (FieldSeg-DA) to extract IAFs from Chinese high-resolution satellite imagery (Gaofen-2). FieldSeg-DA adopts a parallel network structure consisting of two branch networks (UNet and DeepNetV3+) to independently extract the boundary and extent of IAFs. A post-processing module, connecting boundaries and filling field (CB-FF), was used to integrate the extracted boundary and extent and promote the integrity of each extracted IAF. The parallel network coupled with CB-FF improved the boundary and extent accuracy of IAF extraction. Moreover, we used fine-grain adversarial domain adaption (FADA) in the training stage to promote the transferability of the trained network from source domains with labeled samples to target domains without any labeled samples. We found that FieldSeg-DA outperformed IAFRes, a state-of-the-art method for IAF extraction, in both the source and target domains, with improvements of 0.016 and 0.069 in the F1-score, respectively. Therefore, the proposed FieldSeg-DA method has the potential to extract IAFs accurately across diverse farming areas without training samples.