Automatic Land Cover Classification of Geotagged Images using ID3, Naive Bayes and Random Forest

Land cover represents characteristics of earth surface. By utilizing the abundance of geotagged images from online crowdsource images like Geotagged photo library (http://eomf.ou.edu/photos) from the University of Oklahoma, prediction of land cover types will be established by using machine learning techniques. RGB Histogram, Edge Orientation and Vegetation Indices were used to obtain 8 features that representing images, therefore several classifiers were performed to observe which of classifiers produce best accuracy. Best classifier then used to predict unclassified images. The result, Random Forest classifier produces 82% in overall validation accuracy and 89% of 74 unclassified images was successfully predicted comparing with expert prediction result. The last, 74 of successful predicted images were mapped into Geographic Information System (GIS) to show land cover in GIS. This model was measured by using precision, recall, F-Test and Kappa Coefficient. The performance of each measurement reaches 89.8%, 88.1%, 88.6%, 85.6% respectively.