Mapping built-up land and settlements: A comparison of machine learning algorithms in Google Earth engine

Monitoring the growth of built-up land finds its challenge as a never-ending mapping process. Since the built-up maps are taken into account in development planning and measuring the achievement of SDGs objectives (Goal 11-indicator 11.3.1), the best mapping method should be attempted. Therefore, all efforts were worked on to speed up the mapping process without compromising accuracy. Various methods have been proposed, but numerous difficulties remain in accurate and efficient built-up and settlement extraction. With the combination of passive and active sensors images, we try several machine learning methods using Google Earth Engine (GEE) platforms to map built-up land and settlements in Purwokerto, Banyumas, Central Java. Around 369 samples were occupied to distinguish four classes of land covers: settlements, built-up land, waters, and others. The decision tree-based algorithms give the best performances, scilicet Random Forest (RF) and Gradient Tree Boost (GTB). Random forest is a collection of many decision trees, while Gradient Boosting is a machine learning algorithm that uses an ensemble of decision trees to predict values. Thus, the algorithms can handle complex patterns and data when linear models cannot. On the whole, RF and GTB classifiers can distinguish between settlements and non-settlement with an overall accuracy of 80. The Support Vector Machine (SVM) classifier produces 71.43 accuracy with Kappa = 0.61, and the Minimum Distance (Mahalanobis) classifier gain overall accuracy of 74.29 (Kappa = 0.64). © 2021 SPIE.