Sharpened Cosine Similarity U-Net for Deforestation Mapping

Forests cover about 30% of the Earth’s surface, having a significant global impact on climate and atmosphere. A significant change (i.e., forest loss) in the world’s forests has been brought on by factors, such as rising population, increased urbanization, and environmental pollution due to economic activities. Consequently, forest cover and forest loss mapping and monitoring are vital. Convolutional Neural Networks (CNNs) are among the most utilized segmentation algorithms for deforestation mapping. However, CNNs may be more prone to increased model variance, over-sensitivity, and a lack of generalizability. Thus, new concepts, such as Cosine Similarity can be investigated as an alternative approach to the current extensively utilized CNNs. Thus, in this study, we develop and propose the SCS-UNet for precise deforestation mapping utilizing satellite imagery of Sentinel-2 in South America. The segmentation results illustrated that the proposed SCS-UNet segmentation algorithm exhibited the least training time complexity as compared to other implemented segmentation models, such as UNet, Attention UNet, R2UNet, ResUNet, Swin UNet, UNet+++, TransUNet, and TransUNet++, while resulting in comparable statistical results as compared to the U-Net segmentation model.