Integrating remote sensing and machine learning for flood modelling: A systematic literature review

Abstract

This paper provides a systematic review that synthesizes recent advancements in the integration of machine learning (ML) and remote sensing techniques for flood modelling in developing regions between 2010 and 2025. To achieve the main objective the study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) to search for articles in key databases such as Google Scholar, Web of Science and Science Direct. A total of 50197 articles were obtained and screened based on specific set criterion and a total of 126 articles were obtained after screening and were used in this study. These articles were then subjected to bibliometric analysis which revealed an exponential growth in the number of articles obtained with a sharp rise in publications post-2018. Further analysis revealed that most studies concentrated in South and East Asia, highlighting regional bias and underrepresentation of data-scarce areas such as Africa and Latin America. The results further indicated a widespread increase in the use of freely available remote sensing data (e.g., Sentinel-1/2, Landsat, MODIS) which was driven by accessibility and resolution advantages, while advanced but cost-prohibitive platforms (e.g., RADARSAT, UAVs) remain underutilized. Comparative analysis of model performance showed that traditional hydrological and hydraulic models remain relevant but often suffer from oversimplification and high data and computational demands. In contrast, ML models like CNNs, RF, and SVM demonstrated robust performance with AUC values frequently exceeding 0.90. However, the lack of consistent benchmarking, standardized evaluation metrics, and open-source codebases limits model comparability and reproducibility across studies. Furthermore, most reviewed studies overlook uncertainty quantification, compound event interactions, and tail dependence. To address these gaps, the review recommends integrating uncertainty-aware techniques such as Bayesian deep learning (e.g., MC-Dropout) and copula-based bivariate extreme value models. Moreover, emphasis should be placed on the ethical deployment of ML in flood-prone regions, advocating for transparency in model assumptions, fairness assessments, and participatory model design. Future research should prioritize scalable, interpretable, and equitable modelling approaches, particularly in underrepresented and high-risk regions.