Research Scholar Indian Institute of Science, Bangalore, India
The land-atmosphere feedback system is a complex interplay between land use, climate, and other factors that can influence extreme events such as droughts, floods, and heat waves. Modeling these physical processes in a high-dimensional space is very challenging, due to incomplete understanding and limited availability of data. Deep-learning approaches may be used to analyse the complex interaction chains that lead to extreme impacts. Transformers employing deep learning architecture are explored for their ability to capture long-range dependencies in spatiotemporal data and encapsulate the complex interactions between land use, climate, and other factors that contribute to extreme events. In addition, the proposed approach incorporates attention mechanisms that enable interpretability by highlighting the important spatial and temporal features used for forecasting. The proposed approach was evaluated on a case study of the Godavari basin, India, using vegetation indices (representation of crop type and land use in the basin) and climate data from 2000 to 2020. Results showed that the proposed approach gives insights into the driving factors behind land use change and climate extremes. Such information can be useful to make informed decisions about land management, climate adaptation, and disaster risk reduction.
