Advancing Rainfall Seasonality Detection via Bias Correction and Spatial Disaggregation (BCSD) with CHIRPS-v2 for Climate Modeling in Bangladesh

Bangladesh’s unique geographic location, situated at the junction of major rivers, Ganges-Brahmaputra-Meghna basin and bordered by the Bay of Bengal, makes it highly susceptible to heavy seasonal rainfall, often leading to widespread flooding, crop damage, and disruptions to livelihoods. Accurate rainfall monitoring is essential to manage these challenges, but the country’s rain gauge networks are deteriorating, making real-time rainfall estimation difficult. In recent years, real-time Satellite Precipitation Products (SPPs) have become more widely available, offering an alternative to traditional gauge data for forecasting stream flows and managing water resources. However, SPPs often contain biases that need correction before they can be used in hydrological applications such as flood prediction or seasonal forecasting. Our study applied Bias Correction and Spatial Disaggregation (BCSD) approach to improve daily rainfall estimates from the CHIRPS-v2 satellite data. To further refine predictions, a Long Short-Term Memory (LSTM) model was integrated with the BCSD outputs, leveraging its ability to capture temporal dependencies and improve seasonality detection. Comparative analysis of machine learning models, including Random Forest and Gradient Boosting, demonstrated the superior performance of LSTM, achieving an accuracy of 81%. The results showed that bias correction and machine learning integration significantly enhanced the accuracy of seasonal and annual rainfall patterns. This approach not only improves real-time rainfall data accuracy but also supports effective water resource management and disaster preparedness. Our findings suggest the BCSD-LSTM method could be applied in regions with similar climate variability to improve rainfall prediction and climate resilience.