Assessing the Temporal Transferability of Random Forest Models for Land Use and Land Cover Change Detection

Abstract. Monitoring land-use and land-cover (LULC) dynamics in rapidly urbanizing regions is critical for sustainable environmental planning. Dynamic metropolitan areas with rapid urbanization, such as Istanbul, Türkiye, are experiencing significant land-cover changes, among which deforestation is one of the most critical. This study presents a Google Earth Engine (GEE)-based framework to monitor LULC changes in Istanbul from 2016 to 2025 by fusing Sentinel-2 optical imagery, Sentinel-1 SAR and topographic data. From these datasets, a feature set—including spectral bands, vegetation indices, SAR backscatter metrics, and topographic variables—was derived and used to train a Random Forest (RF) baseline model on 2016 Land Parcel Identification System (LPIS) reference data. The baseline model was then applied across the time series to assess its temporal transferability, overcoming the limitation of up-to-date ground-truth data. The baseline model achieved an overall accuracy of 72%, calculated using a validation dataset derived from the LPIS reference data. Feature importance analysis revealed that structural variables—particularly DEM and SAR metrics—were the primary contributors to the classification, used in combination with optical features. Time-series results indicate a cumulative decline of 231 km² in agriculture and 379 km² in forest cover during the study period, inversely corresponding to urban growth. The results of the study highlight that, although applying a single-year model without independent annual validation data causes certain uncertainties—arising from methods, sensors, or topography (e.g., misclassifications)—the proposed framework is highly practical for monitoring deforestation and urbanization trends in complex landscapes.