An effective approach for improving the accuracy of a random forest classifier in the classification of Hyperion data

Random forest (RF) is one of the most powerful ensemble classifiers often used in machine learning applications. It has been found successful on many benchmarked data. However, the performance of an RF model is highly affected by the calibration of the model parameters. It requires optimization of two parameters—(i) size of RF and (ii) number of features. RF is based on the principle of bagging and random selection of relevant features. This paper conveys an effective method in improving classification accuracy of RF. The principal component analysis (PCA) technique was used for dimension reduction of spectral bands whereas correlation-based feature selection (CFS) was used to identify the optimal set of features. RF was initialized by 10 random trees with an increment of 10, with a variable number of features till the model achieved its highest accuracy. The model was tested with variable sample sizes in order to observe the effectiveness. An investigation was carried out on Hyperion sensor data of the Earth Observing-1 (EO-1) satellite. The performance of RF was observed to be significantly enhanced in terms of predictive ability and computational expenses with the optimized set of features and number of random trees as base classifiers. While comparing with the other advanced classifiers like a support vector machine (SVM), multilayer perceptron (MLP) and maximum likelihood classifier (MLC), the optimized RF outperformed all the other classifiers.