Towards a ML based global crop identification model using limited SAR data - That is scalable across data-sparse geographies

Abstract

In this paper, we have proposed a machine learning based global crop identification method using limited microwave/radar data for the corn belt in the US. An attempt has been made to identify the features/crop signatures which are unique for a particular crop but common across geographies. Identified features were used to develop a robust, reliable and scalable crop identification model for corn and soy. The pre-trained model has been tested at multiple locations "as is" without any retraining, yielding best accuracy of 93% (within corn/soy belt) and 84% (at the periphery) at 20m pixel-level spatial resolution.