Agricultural Field Boundary Delineation From Multi-Temporal IRS P-6 LISS IV Images Using Multi-Task Learning

Farm scale information is critical for large-scale agricultural monitoring and policymaking. Manual digitization of the field map or field boundaries using optical satellite images during each growing season is a challenging task. Deep Learning approaches have been successfully implemented for delineating farm parcel boundaries. However, small-scale parcel delineation in such approaches either requires a huge training dataset and high-resolution satellite imagery or gives malformed edges. To overcome these challenges, we developed a multi-task deep learning-based method that generalizes well to unseen data sets. The potential of the proposed algorithm was demonstrated on temporal high resolution (5.8m) IRS P-6 LISS IV images. The deep learning network takes linear directional features and the temporal standard deviation as input to the model in a multi-task learning framework. The model was allowed to simultaneously learn the tasks such as crop-field extent, boundary and distance. We show that our results compare favourably to existing approaches for boundary delineation.