Stochastic Gradient Boosting Algorithm For Land Use Change Detection Using Multi-temporal Landsat 5 TM In Yogyakarta City, Indonesia

Abstract

The advanced technology in remote sensing and geographic information system has facilitated a great deal to land management through land use/land cover (LULC) change detection using spatial and temporal data to analyze the dynamic conversions on the land surface in two or more periods of time. Classification is the main exercise to present this information to the decision makers. A hybrid bagging-boosting machine learning algorithm was used to generate a binary classifier for separating Urban and Non-Urban classes. A high-resolution image would be such a promising opportunity to achieve high classification accuracy, but issues of availability and feasibility have forced the analyst to settle for medium-resolution spatial data with an imbalanced dataset to detect LULC changes of urban agglomeration area in Yogyakarta city, Indonesia, in the years 1999, 2005, and 2011. The result showed that stochastic gradient boosting algorithm was succeeded in building one robust classifier model using LANDSAT-5TM 2005 with an overall accuracy of 0.76 and ROC-AUC value of 0.83. The replicability of classifier was confirmed by an agreement between the predicted class and the reference data from Statistics Indonesia, which showed root mean square error (RMSE) of 9.9% and R2 of 0.91, indicating sufficiently good accuracy for areal integrated multi-temporal urbanization monitoring.