Using optical images to demarcate fields in L band SAR images for effective deep learning based crop classification and crop cover estimation

Abstract

This study aims at developing an end-to-end solution for deep learning-based crop classification using synthetic aperture radar (SAR) images. SAR provides all-weather, day, and night imaging capabilities, sensitive to dielectric properties. Optical images are intuitive and capture static information well, like the boundary of the field in the absence of atmospheric disturbances. In this work, the end-to-end solution to use deep learning algorithms for the crop-type classification is done using SAR images. The limitation of the SAR images is handled by using the boundary information from the optical data. For the classification of different crops in the test site, L-band ISRO L- & S-band Airborne SAR (ASAR) and Airborne Synthetic Aperture Radar (AIRSAR) images were acquired over an agricultural site near Bardoli and Flevoland respectively. Pre-trained model Inception v3 and Custom VGG like model were used for crop classification. Inception V3 enabled us to better discriminate crops, particularly banana and sugarcane, with 97% accuracy, while the Custom VGG like model achieved 95.17% accuracy for 11 classes.