An Investigation of Pixel-Based and Object-Based Image Classification in Remote Sensing

Abstract

This research evaluates pixel-based and object-based image classification techniques for extracting three land-use categories (buildings, roads, and vegetation areas) from six satellite images. The performance of eight supervised machine learning classifiers with 5-fold cross validation are also compared. Experimental validation found that using 'Bagged Tree' for object-based classification algorithms provides maximum overall accuracy when tested on 10,000 objects produced by the SLIC segmentation method, and improves upon an existing RGB-based approach. Our aforementioned proposed approach takes about 12 times less total runtime than the pixel-based method, demonstrating the power of the combined approach.