Explainable Multi-task Learning for Improved Land Use Classification in Planetary Health Monitoring

This paper introduces an innovative approach towards the development of a better understanding of planetary health monitoring techniques. Utilizing hyperspectral satellite imagery from the EuroSAT dataset, we focus on improving land use classification via multi-task learning with indice mapping, particularly emphasizing the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and Normalized Difference Water Index (NDWI) indices. Our approach employs explainable multi-task learning, integrating AI techniques such as GradCAM, SmoothGrad and saliency mapping to enhance model interpretability and provide valuable insights into the model’s environmental perception. Our model achieves a high classification accuracy of 90%, demonstrating low inaccuracy in indice prediction and mapping, thereby ensuring precise assessment of land use. The incorporation of XAI techniques aids in providing a deeper understanding of the model’s decision-making process, highlighting influential regions and features which provide valuable insight into the predictions. These results underscore the effectiveness of our approach in accurately classifying land use and assessing vegetation health and water content, contributing to improved environmental management and decision-making. This work promotes sustainable practices and advances in the field of planetary health monitoring through the use of explainable and reliable AI-driven methods.