Alvin Pratama, Andreas Boni Baik Simamora, Farras Ghaly

Urban Flood Susceptibility Modeling Using GIS and Machine Learning in Bandar Lampung

Introduction

Urban flood susceptibility modeling using gis and machine learning in bandar lampung. Model urban flood susceptibility in rapidly urbanizing Bandar Lampung using GIS and machine learning (LR, RF, GB, SVM). A rigorous framework identifies flood-prone areas.

Abstract

Urban flooding increasingly affects rapidly urbanizing tropical cities, where terrain, rainfall, and anthropogenic surface modification interact to shape spatial flood patterns. This study develops a GIS–machine learning framework to model urban flood susceptibility in Bandar Lampung, Indonesia, using a multi-year flood inventory (2015–2024). A balanced dataset (n = 308; 1:1 flood to pseudo-absence ratio) was constructed using buffered pseudo-absence sampling with spatial separation constraints to reduce bias. Nine environmental and infrastructure-related predictors were evaluated using Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), and Support Vector Machine (SVM). Model performance was assessed through five-fold stratified cross-validation, generalization gap analysis (Train AUC − CV AUC), learning curves, and a 20% hold-out test set. GB achieved the highest cross-validation performance (CV AUC = 0.8953), followed by RF (0.8782), SVM (0.8007), and LR (0.6925). However, ensemble models exhibited larger generalization gaps (RF = 0.1218; GB = 0.1047) compared to LR (0.0333), indicating stronger overfitting tendencies. Learning curves confirmed that LR maintained the most stable convergence between training and validation scores. On the independent test set (n = 61), GB achieved the highest predictive accuracy (ROC AUC = 0.9462), whereas LR showed lower discriminative performance (AUC = 0.7065) but greater validation stability. Flood susceptibility was concentrated in low-elevation areas, near major roads, and adjacent to river networks. By integrating learning curve diagnostics with cross-validation and hold-out testing, this study provides a rigorous framework for model selection in data-limited urban environments.

Review

This study addresses the critical and growing issue of urban flooding in rapidly urbanizing tropical cities, focusing on Bandar Lampung, Indonesia. Utilizing a robust GIS-machine learning framework, the research aims to model urban flood susceptibility by integrating a multi-year flood inventory spanning from 2015 to 2024. The authors carefully constructed a balanced dataset of 308 instances (1:1 flood to pseudo-absence ratio) using a sophisticated buffered pseudo-absence sampling technique, designed to minimize spatial bias. The analysis incorporated nine diverse environmental and infrastructure-related predictors, providing a comprehensive input for assessing flood risk. The research meticulously evaluated four prominent machine learning models: Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), and Support Vector Machine (SVM). Model performance was assessed through a rigorous combination of five-fold stratified cross-validation, generalization gap analysis, learning curves, and an independent 20% hold-out test set. While Gradient Boosting demonstrated the highest cross-validation (CV AUC = 0.8953) and independent test set performance (ROC AUC = 0.9462), the study insightfully revealed that ensemble models like RF and GB exhibited larger generalization gaps, indicating a tendency towards overfitting compared to the more stable Logistic Regression. Learning curve diagnostics further supported this observation, showing LR's consistent convergence. Spatially, the study identified flood susceptibility to be concentrated in expected areas: low-elevation zones, proximity to major roads, and areas adjacent to river networks. Overall, this study presents a highly rigorous and methodologically sound approach to urban flood susceptibility modeling, particularly valuable for data-limited environments. Its key strength lies in the comprehensive model evaluation strategy, which goes beyond mere performance metrics to include critical diagnostics like generalization gap analysis and learning curves – offering a more nuanced understanding of model stability and reliability. This framework provides significant practical implications for urban planning and disaster management in Bandar Lampung and other similar tropical cities. The findings not only identify high-risk areas but also offer valuable insights into the trade-offs between model complexity, predictive power, and generalizability, setting a high standard for future research in this vital field.

Full Text

Comments

You need to be logged in to post a comment.