Evaluating machine learning classifiers for irs high resolution satellite images using object-based and pixel-based classification techniques

Abstract

Satellite imagery has provided the top view for solving many engineering, agriculture, water resources, disaster response, and environmental monitoring problems. The top view requires the classification of real-world objects from satellite images. Broadly two image classification approaches are used: pixel-based and object-based. The pixel-based classification approach primarily works on spectral characteristics and ignores spatial features, whereas object-based classification operates on both spectral and spatial features. The current work examines object-based and pixel-based methods for high resolution satellite images of novel Cartosat − 2E and Cartosat-3 satellites using machine learning classifiers (Support Vector Machines (SVM), Decision Trees (DT), Random Forest (RF), k-Nearest Neighbour (kNN) and Bayes). The study identified that decision tree classifier outperformed other classifiers under object-based approach with kappa coefficient higher than 0.90. On the other hand, for pixel-based approach kNN and SVM classifiers outperformed the other classifiers for Cartosat – 2 and Cartosat – 3 images. However, for Linear Imaging Self Scanning (LISS)-4 image Bayes, SVM and kNN performed relatively better than RF and DT. The model parameters of the machine learning classifiers may be altered or fine-tuned to increase predictive performance for object-based and pixel-based image classification techniques.