TransCropNet: a multichannel transformer with feature-level fusion for crop classification in agricultural smallholdings using Sentinel images

Crop classification in small and fragmented agricultural lands is a challenging task. The presence of cloud cover poses additional challenges when classical techniques using multispectral images are applied. Multispectral data augmented with SAR images is capable of solving this issue. The use of robust deep neural network classifiers can further improve the classification performance. We propose a multichannel attention-based deep learning model to fuse time-series SAR and a single scene multispectral data at the feature level and classify tropical crops grown in smallholdings at pixel level with high accuracy. The attention-based transformer neural network for the fusion of Sentinel-1 and Sentinel-2 data is evaluated for crop classification in the Cauvery delta zone. The model achieves an overall classification accuracy of 92.6% and a kappa score of 0.906. The source code is available in a GitHub repository at: https://github.com/ajigeo/temporal-SAR-transformer.