COMPARISON OF MACHINE LEARNING MODELS FOR LAND COVER CLASSIFICATION

Abstract

Land cover data remain one of crucial information for public use.  With rapid human-associated land alteration, this information needs to be frequently updated. Remotely-sensed data provide the best option to construct land cover maps with numerous methods available in the literature. While disagreement exists to select the robust one, further exploration should be made to extend the understanding on the behavior of machine learners, in particular, for classification problems. This article discusses performance of pixel-based machine learning algorithms, frequently used in research or implementation. Five popular algorithms were evaluated to distinguish five rural land cover classes, i.e. built-ups, crops, mixed garden, oil palm plantations and rubber estates, from Sentinel-2 data. This research found that the benchmark, classification and regression tree, was unable to differentiate woody vegetation, although the overall accuracy was sufficiently moderate. This suggested that overall accuracy cannot be seen as the only measure for assessing the quality of the thematic output. Meanwhile, support vector machines and random forest competed to yield the highest accuracy and class detection capability, although the latter was in favor with 98% accuracy level. A newly developed model, like extreme gradient boosting, achieved a similar level of accuracy. This research implies that modern machine learning approaches would be invaluable for land cover classification; hence, access to these modeling toolkits is substantial.