Machine Learning Enhanced GIS And AHP Modeling for High-Resolution Flood-Susceptibility Prediction, Vulnerability Assessment, and Risk Classification

Abstract

Flood-risk management often relies on generalized hazard maps that inadequately represent localized interactions among hydrological conditions, environmental characteristics, exposure, and multidimensional vulnerability, thereby limiting the accuracy and practical value of risk classification. This study aimed to develop and evaluate an integrated machine learning enhanced Geographic Information System and Analytic Hierarchy Process framework for high-resolution flood-susceptibility prediction, vulnerability assessment, and risk classification within a selected flood-prone case-study area. A quantitative, cross-sectional, explanatory, predictive, and case-based design was employed, combining structured stakeholder assessment with spatial modelling. From 340 distributed questionnaires, 312 were returned and 300 valid responses were retained from GIS specialists, hydrologists, disaster-management professionals, environmental officers, engineers, urban planners, government officials, researchers, emergency personnel, and community representatives. The principal variables included geospatial and hydrological data quality, AHP weighting effectiveness, machine-learning predictive capability, GIS-AHP-ML integration capability, socioeconomic and demographic vulnerability, physical, environmental, and infrastructural vulnerability, flood-susceptibility prediction effectiveness, vulnerability-assessment effectiveness, and integrated flood-risk classification effectiveness. Data were analyzed using descriptive statistics, reliability and validity testing, Pearson correlation, multiple regression, ANOVA, AHP consistency analysis, and comparative machine-learning evaluation. The instrument showed strong reliability, with Cronbach’s alpha values from .821 to .914, while KMO reached .891. Random Forest achieved the strongest predictive performance, with 88.7% accuracy, 89.4% recall, an F1-score of 88.6%, and an AUC of .931. High and very high susceptibility zones covered 37.2% of the study area but contained 70.1% of observed flood locations. High and very high-risk zones represented 39.4% of the area, 57.6% of the exposed population, and 60.2% of exposed infrastructure. The three regression models explained 68.2%, 59.7%, and 64.9% of the variance in susceptibility prediction, vulnerability assessment, and risk classification, respectively. The findings imply that accurate spatial data, transparent expert weighting, validated machine-learning models, and multidimensional vulnerability indicators should be integrated to improve flood planning, infrastructure protection, preparedness, and resource prioritization.