Large Scale Land Cover Mapping in Ontario, Canada, Using a Deep Learning Framework

This study addresses the challenge of large-scale land cover mapping using advanced deep learning models. While state-of-the-art deep learning methods have demonstrated promising results in various remote sensing applications, their efficiency for large-scale semantic segmentation tasks remains underexplored. A key limitation is their reliance on extensive training datasets. To address this issue, we propose a two-stage classification approach that integrates random forest for initial land cover mapping and MobileUNetR, a lightweight hybrid convolution-transformer model, for a refined classification. Leveraging Sentinel-1 and Sentinel-2 data, the land cover map of Ontario at spatial resolution of 10 m, aligned with the North American Land Change Monitoring System Level I legend, encompassing 11 classes, is generated. The findings of this study reveal that MobileUNeTR surpasses widely used models such as UNet and PSPNet in terms of both accuracy and efficiency, underscoring its suitability for large-scale land cover mapping. In particular, an overall accuracy of 85% and a Kappa coefficient of 0.83 are achieved with MobileUNetR, which has only about one-fourth to one-fifth the number of parameters compared to other deep learning models examined in this study. As the only deep learning model, examined in this study, combining convolutional and transformer blocks, MobileUNetR demonstrates the superiority of hybrid architectures for large-scale semantic segmentation. This is due to its capability in capturing both local and global features, which are essential for semantic segmentation of heterogeneous land cover classes with varying sizes and spectral signatures.