Land usage analysis: A random forest approach

Abstract

Land usage analysis takes advantage of the multi-band imagery for classification and recognition. Multi-bands data contains reliable information compared to the raw image formats e.g. RGB, HIS, HSV and other color spaces. In this paper, we advocate the usage of non-parametric machine learning algorithms for land usage analysis. From the non-parametric algorithms, we propose a random forest approach for land use analysis. Our analysis is concerned with the classification of land into seven classes. We have shown that non-parametric classifier the “Random Forest” is well suited to the task of multi-band land usage analysis. In the experimentation setup, we have compared the random forest with the state-of-the-art classifiers. Based on the SPOT-5 imagery, we have shown that the random forest outperforms the state-of-the-art classifiers including Naïve Bayesian, Mutli-Layer Perceptron, Bayesian Network, SVM, Radial Basis Function Network (RBF) and Ada-boost. We further show that for the land use analysis, increasing the number of trees has no effect on the performance of the random forest and therefore the runtime of the random forest can be reduced compare to all the other classifiers. The best F-score is achieved using 4 trees and 10 Fold Cross Validation.