Land Cover Segmentation Using 3-D FCN-Based Architecture With Coordinate Attention

This letter presents a land cover segmentation approach by analyzing multispectral, multitemporal Sentinel-2 satellite images through deep-learning-based fully convolutional networks (FCNs). Existing segmentation methods face problems in generating accurate segmentation masks for satellite scenes with bright pixels and those containing cloud cover and incur a high computational cost. To tackle these problems, first, we generated cloud cover masks using a trained U-Net++-based architecture with ResNet-50 backbone and a lightweight attention mechanism to remove too cloudy satellite scenes as they hide the ground view. Second, to incorporate the benefits of attention mechanisms while keeping the computational cost low, a 3-D FCN-based satellite image segmentation architecture with lightweight coordinate attention (CA) mechanism is proposed for generating land cover segmentation masks. Third, a novel dataset comprising multitemporal, multispectral Sentinel-2 satellite images for the year 2020 along with their ground-truth mask was composed for the chosen study site which is a semi-arid region spanning over an area of 4978 km2 of Ludhiana district located in the state of Punjab, India. Understanding the patterns of land use and land cover in the context of such agriculturally rich regions experiencing significant urban population increase and cropland loss makes this study crucial. Mean intersection over union (mIoU) and $F1$ scores of 92.13% and 96.76%, respectively, were achieved.