Bridging Basins with Algorithms: Machine Learning for Scalable Flood Prediction

Flood disasters are complex events influenced by numerous natural and anthropogenic factors, making accurate forecasting highly challenging. Machine learning (ML) techniques offer promising results by modeling these intricate processes. However, the most critical limitation of ML based flood forecasting lies in data availability. In basins lacking historical flood records, prediction capabilities are significantly constrained.This study aims to predict flood events in a data-scarce basin by utilizing records from 41 basins with historical flood data. Based on flood events recorded between 1950 and 2020, flood forecasts were conducted for the year 2021 in a separate basin located approximately 450 km away, selected due to its proximity and similarity. The primary objective is to assess the reliability of flood predictions in ungauged basins by comparing intra-basin and cross-basin prediction models.To this end, nine different machine learning algorithms were employed, and the results were spatially mapped. Six performance evaluation metrics were applied to assess model accuracy. The findings reveal that Gradient Boosting, a Hybrid Model, and the Random Forest algorithm achieved prediction accuracies exceeding 90%.Furthermore, the effectiveness of the cross-basin prediction approach was evaluated against traditional intra-basin models. The results indicate that machine learning offers substantial potential for improving flood prediction reliability in data-deficient regions.The Eastern Black Sea Region of Turkey, one of the most flood-prone areas in the country, was selected as the study area. This region has experienced severe flood events resulting in the loss of over 650 lives and significant economic damage. The study sheds light on how machine learning can support more effective flood forecasting and risk management, even in the absence of historical data.