Stochastic gradient boosting for urban change detection using multi-temporal LANDSAT-5TM in Yogyakarta, Indonesia

Abstract

Despite available remote sensing data, technical challenges in developing countries have hindered local urban authorities from updating periodic land cover maps. Therefore, this study proposed a practical approach for regions with insufficient ground truth data. The study implemented a machine learning algorithm using single date medium spatial resolution data to build a classifier for separating Urban and Non-Urban zones. Then, the classifier was employed on multiple dates in 1999, 2005, and 2011 to corroborate its robustness. Results showed the stochastic gradient boosting (SGB) algorithm succeeded in building a robust classifier using the digital number value of LANDSAT-5TM 2005 with an overall accuracy of 0.76 and an area under curve receiver operator characteristic (AUC-ROC) value of 0.83. Moreover, the classifier predicted that urban areas in Yogyakarta, Indonesia, reached 24,099 (hectares) ha; 26,598 ha; and 22,650 ha in 1999, 2005, and 2011, respectively. The classifier's performance in predicting multiple datasets combined with histogram matching of medium spatial resolution data showed satisfactory results comparable to reference data from Statistics Indonesia, indicating sufficient accuracy for areal-integrated multi-temporal urbanization monitoring.