Long-term prediction of floods with a TSVD-LSTM based on the century-scale rainfall data: A case study on Kerala, India

Floods are among the most destructive natural disasters because they create life-threatening conditions and damage essential infrastructure worldwide. The 2018 Kerala flood was one of the most severe floods in recent history, caused by heavy rainfall. The lack of an early warning system leaves authorities and communities unprepared. The situation requires flood prediction systems that provide both precise results and prompt operational capabilities. This research presents a hybrid Truncated Singular Value Decomposition–Long Short-Term Memory (TSVD-LSTM) model for regional flood prediction, trained on the Kerala region dataset. A multivariate time-series dataset is created utilizing rainfall records (1901–2025) from IMD (India Meteorological Department) and flood-related information sourced from Kerala State Disasters Management Authority (KSDMA), as well as open-source archives like Kaggle datasets, news articles from the Malayala Manorama portal, and social media platforms, especially Facebook, to track real-time and community reports of flood events. In this research, TSVD is used to reduce dimensions, handle multicollinearity, and extract dominant features. While the LSTM algorithm captured long-term temporal dependencies in flood patterns. The TSVD-LSTM model achieves superior predictive performance, with an accuracy of 0.958 compared to the baseline LSTM model (0.833), demonstrating that combining dimensionality reduction with deep learning techniques enhances forecasting flood events.