| dc.description.abstract | Accurate water-level forecasting plays a crucial role in proper water resource management, flood control and agricultural planning especially in climate sensitive areas like Bangladesh. Nonlinear interactions and dynamic variability in river systems that exist in increasing levels of drought conditions is usually not well represented in traditional hydrological models. This study suggests a drought-sensitive machine learning system that can be used to enhance water-level prediction in the Nabaganga and Chitra river basin to overcome these issues. The suggested methodology incorporates multi-river data of Bangladesh Water Development Board (BWDB) stations during a 25-year span (20002025) in which the crucial hydrological and climatic variables are involved such as water level, rainfall, lag features, and seasonal indicators. Various machine learning models such as Random Forest, Gradient Boosting, Extra Trees, xgboost, Artificial Neural Networks (ANN), lightgbm and catboost were created and tested to determine the best to use in water-level prediction. The experimental findings indicate that ensemble and boosting-based models are much better than neural network-based approaches. Gradient Boosting has the best performance in the drought-sensitive multi-river framework, with an R 2 of 0.993 and an RMSE of 0.081, and xgboost has similar accuracy, which indicates that it is highly suitable in terms of discerning cross-basin interactions, seasonal variability, and drought-related low-flow conditions. Further validation on single-river data further supported the dominance of catboost, with an R 2 of 0.9982, which is almost perfect predictive accuracy. The results also show the role of time-dependent factors with lagged water-level factors being the most significant predictors, then rainfall and seasonal factors. The long-term trend analysis shows that water levels have been decreasing gradually over the past years, which demonstrates a growing drought stress over the region of study. The multi-river modeling approach has been successful in capturing spatial variability and hydrological connectivity, which can be more generally applicable and robust than single-river models. Overall, the study indicates that multi-river modeling using machine learning in conjunction with drought sensitivity, can be more effective in terms of predictive accuracy, robustness and practicality. The suggested framework offers a flexible and dependable solution to water-level forecasting and aids in making well-informed decisions in the management of water resources, especially in climate-sensitive and drought-affected areas. | en_US |
