A hybrid UNet based approach for crop classification using Sentinel-1B synthetic aperture radar images

With the growing popularity of deep learning, semantic segmentation using convolutional neural networks (CNNs) has proven the state of the art in the pixel-level classification of the remote sensed multi-temporal images captured by satellites such as Sentinel-1A, Sentinel-1B, Sentinel-2, and Landsat-8. Among these, the temporal Sentinel-1B data has widely been used for crop mapping. This research is entirely focused on crop classification based on Sentinel-1B synthetic aperture radar imagery. We have implemented seven popular CNN-based deep learning models and their variations for the segmentation and classification of the pre-processed Sentinel-1B SAR images. Further, we proposed an approach by collaborating the UNet and SEResNext50 as the backbone along with the custom loss function (a hybrid of dice loss and focal loss) and evaluated its performance qualitatively and quantitatively using various metrics. It is observed that the proposed approach is able to achieve an average IoU of 0.6465, average precision of 0.7371, average recall of 0.7191, and average F1-score of 0.7352. Based on the per-pixel confusion matrix the proposed approach achieves an overall accuracy of 98.69% and a kappa coefficient of 0.87. Further, the applicability in the context of Indian agriculture, as well as the current assistance provided by the Mahalanobis National Crop Forecast Centre as part of the Forecasting Agricultural output using Space, Agrometeorology, and Land-based observations programme has been discussed. We have also suggested a few proposals that can be considered by the Ministry of Agricultural and Farmer Welfare, India for the development of the application/platform to provide the ground labels/reference in formats such as GeoTiff or shapefile.