Bayesian Deep Learning for Modeling Precipitation in Northern Iraq

Predicting precipitation amounts is a complex problem due to the non-linear nature of the precipitation processes, the significant interactions between factors influencing precipitation formation, and the varying levels of uncertainty resulting from climatic variability. This study aims to develop a model based on Bayesian Neural Networks (BNN) to predict monthly precipitation amounts and represent the uncertainty inherent in these predictions. The model applied to climatic data for a region from north of Baghdad to the Iraqi-Turkish border for the period 1984–2024, obtained from the NASA POWER database. The data included a set of influencing climatic variables: specific humidity at a height of 2 meters (QV2M), temperature at a height of 2 meters (T2M), wind speed at a height of 2 meters (WS2M), and surface soil moisture (GWETTOP), in addition to monthly precipitation amounts as the target variable. Because precipitation is influenced by a range of interacting and simultaneous factors, several interaction variables were derived to enhance the model's capabilities and improve its predictive flexibility. The results demonstrated the proposed Bayesian model's ability to accurately represent temporal precipitation patterns and to handle the uncertain and irregular nature of the studied climatic phenomenon. The model achieved a coefficient of determination (R²) of 0.821, which demonstrates its ability to explain a large proportion of the variance in monthly precipitation amounts. Ablation analysis performed to interpret the model's performance and determine the importance of the variables. The results showed that the Moisture_Thermal_Index variable had the greatest impact on the model's performance, followed by Evapo_Proxy, then GWETTOP and GWETTOP_lag1. This underscores the importance of interactions between moisture, temperature, evaporation processes, and soil moisture status in explaining precipitation dynamics. The findings of this study confirm that integrating BNN with physical knowledge-based climate variable engineering provides an effective methodology for precipitation forecasting. This framework can serve as a supportive tool for decision-making in water resource management, environmental planning, and climate change risk assessment.