A Class-balanced Cost Driven Deep Learning Approach for Efficient Land Use Land Cover Classification using Remote Sensing Data

In view of rapid climate change and its impact, it is of paramount important to identify regions of rapid change and determine the causes/drivers of the change, so that important consequences for both natural ecosystems and human systems are understood. Land Use Land Cover (LULC) must be regularly mapped and monitored to accomplish it. Automated LULC classification has turned into a fundamental task in remote sensing and geospatial analysis in recent times aided by advanced computational infrastructures and data analytical tools. Class imbalance is still a major challenge in Deep Learning (DL) based studies for remote sensing classification studies in very diverse landscapes like India. This study focuses on an efficient approach of automatic delineation of six primary land covers classes i.e., Built-Up, Agricultural Land, Woodland-Dense, Woodland-Open, Scrublands, Waterbody by applying DL architectures such as DeepLabV3+ and U-Net++ with ResNet-50 encoder for faster learning by implementing as class imbalance efficient cost function in the model. The customized models are evaluated by tracking Categorical Cross Entropy loss (CCE), Dice loss and by employing the accuracy as well as Intersection over Union (IoU) metric, a widely recognized measure of segmentation accuracy. U-Net++ architecture demonstrated a better performance in terms of mean validation IoU (85.72%) in comparison with DeepLabV3+ (83.71%) with maximum IoU for waterbody class by both the models as 96.31% and 91.31% respectively. The performance of the model is tested across diverse areas, studye demonstrated the robustness and adaptability of our approach across various landscapes and biophysical contexts.