Application of Drone Images Analysis to Predict the Damage of Streets in HCMC

This study introduces a novel framework leveraging drone-based image analysis to improve urban street infrastructure management. With urbanization accelerating, effective street maintenance is critical for sustainable development. The study focuses on streets 6–8 meters wide and 1–20 years old, as defined by 22TCVN 4054-2005 standards. This study aims to develop an integrated analytical framework for urban street infrastructure management, combining drone-based image analysis, ML predictive modeling, and financial assessment. The purpose is to enhance decision-making for street maintenance and sustainability through the application of advanced technologies. The methodology involves three key phases: (1) collecting drone images of urban streets and developing a deep learning classifier to assess street conditions; (2) constructing a Markov model to predict street longevity and usage based on classified data; and (3) creating a financial model to estimate maintenance and reconstruction costs. These three models are designed to work in unison, forming a comprehensive system that transitions seamlessly through condition assessment, prediction, and cost estimation. Using a combined dataset of nearly 2,000 images from Ho Chi Minh City (HCMC) and the public RSXD dataset, deep learning classifiers were trained to evaluate road conditions. Outputs are then fed into a Markov model for usage and lifespan predictions, alongside financial projections for maintenance or reconstruction. The findings reveal a critical disparity in performance: while the models achieved moderate success in assessing the damage severity level , they failed to generalize for the more complex task of classifying the specific damage type , where validation accuracies were low despite high training accuracies. The framework features a Markov model that effectively predicts street longevity, and a cost model aiding financial planning. The framework demonstrates robust performance metrics, validating its applicability in urban street management. This study’s originality lies in its integration of advanced vision recognition, predictive modeling, and cost analysis for infrastructure planning. Limitations include a restricted geographical scope and reliance on specific street dimensions. Future research should extend the approach to broader contexts and refine the models for diverse urban conditions.