Generating pre-harvest crop maps by applying convolutional neural network on multi-temporal Sentinel-1 data

Abstract

ABSTRACT Pre-harvest crop mapping, the fundamental requirement for many of the crop management decisions, continues to be challenging either due to cloud cover in satellite images or due to spectral separability issues. These limitations are overcome in this study by using Synthetic Aperture Radar (SAR) data and deep learning technique. Two-dimensional convolutional neural network (2D-CNN) architecture is applied on multi-temporal SAR data of Sentinel-1 to classify soybean, jowar, cotton and sugarcane crops in a large geographic area located in a Central Indian state. Classification experiments are conducted with full season data from nine overpasses in three modes namely VH alone, VH and VH/VV ratio, and VH and VV, and these experimental results reveal that VH and VV combination performed better with an overall accuracy of 91.75% as compared to 84.96% and 88.75% respectively by others. Classification performances with three sets of temporal data, covering part of the season, reveal that crop map can be generated as early as 27th August (i.e. roughly 45 days prior harvesting) with an accuracy of 89.15% slightly less than the mapping accuracy achieved with full season data (i.e. till 14th October). 2D-CNN algorithm has performed better than SVM and RF techniques. This methodology can be extended to similar agro-ecological regions of the country. Richly available SAR data from Sentinel-1 and the potential of deep learning techniques for recognising complex phenological patterns offer immense opportunities for early-season crop mapping even during monsoon season in tropical countries like India.