Generating Land Cover Maps in Semi-arid Regions Based on a 3D Semantic Segmentation Architecture Using Multi-temporal Sentinel-2 Satellite Images: A Case Study of Ludhiana District in Punjab, India

Using satellite imagery for land cover mapping is of utmost importance due to its role played in environmental management and protection, assessment of natural resources and human activities and planning for sustainable development. The objective of this research study is to generate high-resolution land use/land cover maps for the semi-arid regions. First, we introduce a novel district-scale land cover dataset for a semi-arid region spanning over an area of 4978 km2 of Ludhiana district located in the state of Punjab, India. The dataset is composed of multi-temporal, multi-spectral Sentinel-2 satellite images for the year 2020 along with their ground truth mask. Eight major land use/land cover classes were mapped at the pixel level. Second, in order to exploit the spatio-temporal nature of satellite imagery, we employ a 3D encoder-decoder-based semantic segmentation architecture to produce the land cover maps. Third, we show the relevance of incorporating temporal features of satellite data for land use/land cover mapping through experimental results. Mean Intersection over Union (mIoU) and F1 score of 89.30% and 94.35%, respectively, were achieved. This study is vital from the perspective of understanding the dynamics of land use/land cover patterns in the context of agriculturally rich areas experiencing rapid urban population growth and shrinkage of cropland.