What question did this study set out to answer?

To create a machine learning framework that accurately predicts flood risk at ward level in Nyando Basin.

July 1, 2026Open Access

Nyando Basin Flood Risk Prediction System: A Production-Grade, Explainable Machine Learning Framework for Ward-Level Flood Susceptibility Mapping in Kisumu County, Kenya — Revised Version 2.1

JKJames KoeroGreat Lakes University of Kisumu

Key Points

To create a machine learning framework that accurately predicts flood risk at ward level in Nyando Basin.
Used 2,308 Google Earth Engine satellite observation points for model training.
Implemented GradientBoosting model achieving AUC-ROC of 0.9717 and F1 score of 0.9022.
Developed system in Android with CI/CD for testing using GitHub Actions.
Achieved spatial CV of 0.9727 +/- 0.004, indicating high prediction accuracy.
Generated real-time flood susceptibility scores for five wards in Nyando sub-county.
Expanded test suite to 41 unit tests for enhanced reliability.

Abstract

Revised and corrected Version 2.1 of the Nyando Basin Flood Risk Prediction System. Corrections in this version: population updated to 161,000 residents of Nyando sub-county, training dataset confirmed as 2,308 real Google Earth Engine satellite observation points, primary model confirmed as GradientBoosting achieving AUC-ROC 0.9717, F1 0.9022, spatial CV 0.9727 plus or minus 0.004, test suite expanded to 41 pytest unit tests with GitHub Actions CI/CD. The system delivers real-time 72-hour ward-level flood susceptibility scores for five wards of Nyando sub-county. Built entirely on Android using Termux and Google Colab from Kisumu, Kenya. Live API: https://nyando-flood-api.onrender.com/docs Dashboard: https://nyando-flood-ai.vercel.app Repository: https://github.com/jameskoero/nyando-flood-ai

Demander à l'IA

Bookmark

View Full Paper