Refining land cover classification and change detection for urban water management using comparative machine learning approach

Abstract

Accurate land cover (LC) maps are essential for urban water balance modeling, particularly in rapidly urbanizing cities like Debrecen, Hungary, where industrial expansion has intensified since 2019. However, LC classification remains challenging due to limited studies evaluating the optimal combination of classifiers and satellite data. This study builds upon previous research by introducing a comparative analysis of three machine learning classifiers—Support Vector Machine (SVM), Maximum Likelihood Classification (MLC), and Random Forest (RF)—in LC classification using Sentinel-2 and Landsat 8 imagery from 2018, 2020, and 2022.

Results show that SVM on Sentinel-2 achieved the highest accuracy (F1 score: 0.84 ± 0.11, overall accuracy: 88 ± 2.1 %, kappa: 0.84 ± 0.03) with the lowest total disagreement values (D% = 12.6 in 2020, 13.1 in 2022). Consequently, SVM with Sentinel-2 was selected for LC change detection, employing trajectory analysis to assess urban development dynamics. The quantity gain component accounted for 5 % of the study area, representing net urban expansion, while the exchange component (10 %) indicated bidirectional shifts between developed and non-developed classes. Given Debrecen's rapid industrialization and the lack of a robust LC classification strategy for hydrological applications, this research refines LC change detection methods. It improves water balance calculations by LC type, strengthening the hydrological framework. By demonstrating the value of satellite imagery and GIS in monitoring urbanization, the findings support future urban water balance assessments, sustainable planning, and resource management, providing local authorities with a robust tool to adapt spatial strategies to an evolving landscape.