Determining design dimensions for reusable code in built-up area change detection from satellite imagery

This study aims to detect settlement area changes from Sentinel-2 satellite imagery using a reusable and adaptable code framework. To this end, an integrated analysis workflow was developed, comprising data preprocessing, seasonal composite generation, variable-set optimization, grid-based training–testing separation to reduce spatial leakage, automatic selection of Random Forest parameters, and data-driven threshold determination. The performance of the model was evaluated using different variable sets constructed from spectral bands, derived indices, and texture features. The results showed that the contribution of these variables depends on the spectral characteristics of the region, indicating that model dimensionality must be optimized in a data-driven manner for each study area. Instead of random sampling, the study employed a grid-based validation strategy to control spatial leakage, which produced more realistic accuracy estimates. Additionally, the automatic selection of the number of trees and decision thresholds in the Random Forest classifier improved both model stability and classification performance. The generated building-change maps indicate a net settlement increase of approximately 3.3 km² in the Ömerli Basin and 0.33 km² in the Çubuk-1 Basin between 2018 and 2024. These findings demonstrate the applicability of the proposed method for watershed management, land-use monitoring, and urban growth analyses. Overall, the study presents a reusable analysis workflow that can be applied to different regions with minimal adjustments, automatically optimizes model parameters, and produces spatially reliable results.