Utilization of Open-Source Software in Land Use and Land Cover Mapping for Agricultural Purposes: Study Case in Bogor Indonesia

Abstract

The development of digital technology changed the technique of making maps based on manual digitization into automatic digitization. A variety of software that is equipped with additional programs (plug-ins) makes the mapping process even easier because it cuts down on a long series of procedures. This research examines the effectiveness of using 4 machine learning in the Dzetsaka plug-ins on QGIS software, which is open-source software. The four machine learnings are The Gaussian Mixture Model (GMM), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN). A total of 110 polygons containing 10 training data in each class were created manually based on RGB Sentinel 2 images taken throughout 2019. Next, the training data is divided by 50% for the training model and validation test. Overall, the accuracy of the validation test with confusion matrix by Dzetsaka for 4 machine learning in 2 districts, Rumpin and Kemang (Bogor Regency) reached 90%. Virtual validation resulted in RF consistency producing the best machine learning for the classification of cropland, scarce vegetation, dense vegetation, fallow land, building area, and water body with a total accuracy of more than 74%. These results provide opportunities for the use of the open source for independent mapping of land use and land cover at the district level. The prediction maps are also adequate for describing land use and land cover at the district level.