A Multi-Year, Multi-Location Sampling Method to Improve the Robustness of Forest Cover Classification in Google Earth Engine Shomat F, Radinal, Permana J, Pratomo T D A, Budihandoko, Y
Fauna & Flora International Indonesia Programme
Natural Resources Conservation Center (KSDA) of West Kalimantan
Abstract
Monitoring forest land cover change is essential for evaluating conservation impacts, with remote sensing providing a powerful tool for this purpose. However, conventional supervised classification methods often yield inconsistent results year-on-year, primarily because training samples are typically collected from the same year and location as the imagery under analysis, limiting their temporal transferability. This study investigates an alternative approach to enhance consistency by developing more robust and generalized spectral signatures. We tested a sampling strategy that generates training data from multiple years and diverse geographical locations. Utilizing the Google Earth Engine platform, we employed a Random Forest classifier on a multitemporal stack of Landsat 8 imagery. A comparative analysis was conducted, contrasting our proposed multi-year, multi-location sampling method with the traditional single-year, single-location approach. The performance of both methods was evaluated using Kappa Analysis. The results reveal that the multi-year, multi-location sampling strategy yielded a marginally higher Kappa coefficient compared to the conventional method. These findings suggest that developing a more generalized training dataset can improve classification accuracy and, crucially, enhance the consistency of long-term land cover monitoring efforts, leading to more reliable assessments of conservation outcomes over time
Keywords: Forest Cover Monitoring, Google Earth Engine, Landsat 8, Temporal Consistency, Training Sample Generalization
