In-Season Wall-to-Wall Crop-Type Mapping Using Ensemble of Image Segmentation Models

Abstract

Availability of high-resolution remote sensing images along with advances in deep learning has enabled timely generation of crop-type maps. However, scaling such maps to large areas is often hindered by data occlusion due to clouds. In this article, we propose an ensemble of image segmentation models, based on attention U-Net architecture and optimized using focal Trvesky loss (FTL), to scale crop-type mapping by tackling the missing data problem. We trained models on Landsat 8 and Sentinel 2 surface reflectance data and created a bagged ensemble of those trained models to generate wall-to-wall crop-type maps of the entire US Corn Belt. We also demonstrate the effectiveness of our approach in generating preharvest wall-to-wall crop-type maps. Our proposed method obtains a competitive 92% overall accuracy for crop-type mapping without requiring the availability of the entire in-season time series of remote sensing images over the entire mapping region.