Beyond the Extremes: Interpretable insights based on Attention framework

Extreme weather events significantly impact the economy, agriculture, infrastructure, and ecosystems of a region. According to the Center for Science and Environment (CSE), extreme weather events caused the loss of nearly 3000 lives, 2 million hectares of crops, and the death of 90,000 cattle in India in 2023 alone. Effective mitigation and adaptation strategies for the region necessitate a reliable forecasting system. The spatiotemporal interactions of several hydroclimatic components at different scales make it difficult to provide reliable forecasts for a region with multiple climate zones. The present research proposes an encoder-decoder-based deep learning framework with an attention mechanism to develop a reliable forecasting model. Attention frameworks have exhibited considerable potential in learning contextual awareness within the time series domain which will help in identifying the temporal dependencies between the meteorological variables and extreme events. The proposed architecture of the forecasting model is made interpretable as it is crucial to comprehend the underlying mechanism of climatic extremes. It recognizes the contributing variables influencing the intensity and frequency of extreme events. The study employed 0.12° × 0.12° high-resolution IMDAA (Indian Monsoon Data Assimilation and Analysis) dataset encompassing climatic variables like precipitation and temperature.

Various studies have supported the association of the ENSO parameters with the anomalous climatic conditions over India. The present study also aims to ascertain the distinct contributions of ENSO variables through the implementation of an interpretable framework. Explainability results underscore the significance of precipitation patterns while forecasting drought conditions in the region. Moreover, the results highlight the complex interaction of climatic variables that affect the intensity of the extremes.