Spatio-attention-based network to improve heavy rainfall prediction over the complex terrain of Assam

Abstract

Heavy rainfall events prediction at the local scale imposes a big challenge for meteorological agencies over the complex terrain areas in India such as Assam, Uttarakhand, and Himachal Pradesh and causes flash floods with severe consequences throughout the area causing a huge socio-economical loss over these regions. Assam is currently experiencing severe flooding in June 2023. Due to the limits of deterministic numerical weather models in accurately forecasting these events, this work investigates the incorporation of deep learning (DL) models, particularly spatial attention-based U-Net, using simulated daily collected rainfall outputs from various parametrization schemes. This is a pioneering effort to improve district-scale rainfall using the spatio-attention U-Net DL method, particularly over the orographically complex region such as Assam. The proposed model outperformed individual and ensemble Weather Research and Forecasting (WRF) model outputs over four days in June 2022, demonstrating greater abilities to forecast rainfall at the district scale with a mean absolute error of less than 10�mm. Additionally, the proposed model considerably outperformed WRF models by 51.3% in categorical rainfall prediction, achieving a high prediction accuracy of 91.9%. Furthermore, the proposed model has demonstrated improved spatial variation as compared to the WRF model by correctly predicting severe rainfall occurrences at the district scale, including Barpeta, Kamrup, Kokrajhar, and Nalbari. The WRF projections regularly underestimated rainfall intensity (< 100�mm), whereas the DL model's estimates matched actual rainfall readings from the India Meteorological Department (> 150�mm). On the quantitative estimation of rainfall thresholds using different skill scores, Equitable threat score values are more than 0.5 for all categories for the proposed model. In a nutshell, the findings of the study have direct implications for improving early warning systems and associated follow-up action in terms of developing efficient strategies toward better preparedness, mitigation, and adaptation measures over complex hilly regions to reduce loss of lives and properties. � The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.