Flood Prediction System: A Validation-First Machine Learning Pipeline for Flood Risk Assessment in Data-Scarce Environments

Abstract Flooding remains one of the most destructive natural hazards in Cameroon, causing significant socioeconomic losses while existing early-warning mechanisms remain reactive rather than predictive. This paper presents Flood Prediction System, an AI-powered, web-based flood risk assessment platform combining a Next.js progressive web application, a serverless Convex backend, and a Flask machine-learning microservice. Following the OSEMN data-science framework (Obtain, Scrub, Explore, Model, iNterpret), we trained and cross-validated Logistic Regression, Random Forest, and XGBoost classifiers on an eleven-feature environmental dataset associated with flood occurrence. All three models converged at chance-level performance (accuracy 49.5-50.3%, ROC-AUC approximately 0.49-0.50). Rather than treat this as an implementation failure, we designed a three-part validation protocol feature-target correlation analysis, a shuffled-label permutation control, and a cross-check against a structurally different benchmark dataset all of which confirmed that the chance-level result is a genuine property of the primary dataset's lack of learnable signal, not a defect in the modeling pipeline. On the benchmark dataset, the same pipeline recovered near-perfect performance (classification accuracy up to 100%, regression R² up to 1.000), demonstrating that the models and code correctly extract relationships when they exist. A complementary user-acceptance survey of 61 Cameroonian respondents found strong public demand for the platform (92% believed technology can help predict floods; 97% would recommend the system) despite the underlying model's current predictive limitations. We argue that this validation-first approach reporting an honest null result while proving the pipeline's correctness constitutes a defensible and underused methodology for applied machine learning in data-scarce regions, and we outline a structural- surrogate development strategy, together with concrete recommendations for institutional data sourcing, for teams facing similar constraints.