Exploring the Impact of Hyperparameter Tuning on the Accuracy of Machine Learning-based Flood Susceptibility Mapping

Accurate mapping of flood-prone areas, as a non-structural approach, helps to reduce flood risk and inform flood management planning. With the recent advances in machine learning (ML) models, these algorithms, such as support vector regression (SVR), have been widely applied to generate reliable flood susceptibility maps. However, the use of ML models still faces two major research gaps: (1) the lack of a systematic evaluation of how hyperparameter tuning methods affect model performance, and (2) the lack of comparative studies aimed at identifying the most effective optimization techniques for optimizing ML model hyperparameters for flood susceptibility mapping. This study addresses these gaps by systematically benchmarking six distinct hyperparameter optimization strategies including grid search (GS), random search (RS), Bayesian optimization with Gaussian process (BOGP), Bayesian optimization with a tree-structured Parzen estimator (BOTPE), tree-based pipeline optimization tool (TPOT), and the genetic algorithm (GA) to evaluate their effectiveness in improving SVR performance for flood susceptibility mapping (FSM). The innovation of this research lies in its systematic comparison of optimization methods using a unified dataset and a consistent modeling framework. ... mehrFor this purpose, a dataset from Marvdasht County, Fars Province, Iran, was used. The results of FSM evaluation showed that the SVR-GA model achieved the highest prediction accuracy (AUC = 0.967, R² = 0.89), followed by SVR-RS and SVR-GS, confirming the superior performance of metaheuristic and stochastic optimization approaches. The findings indicate that the choice of hyperparameter optimization method significantly impacts FSM accuracy.