Spatial Classification and Prediction in Hyperspectral Remote Sensing Data using Random Forest by Tuning Parameters

Abstract

Over the past decades hyperspectral remote sensing data have been emerging in identifying the geographical patterns and predicting its behaviour. A digital remote sensing data offers, many practices in learning, exploring, monitoring and understanding the behaviour of the earth surfaces. The hyperspectral remote sensing data is an advanced technique where the topological information are collected in spectral with more number of bands. Inculcating the information from the hyperspectral data relies on the data mining approaches where the data contains many spectral bands and huge temporal information. It requires an front running algorithm to yield high accuracy. The random forest technique is one of the best tree based techniques where optimal solutions are captured at high end construction of possible trees. The mission of this research paper is to give detailed analysis of tuning the parameters of random forest technique based on variable importance, conditional inference and quantile forest applied to AVIRIS Indian pine site-3 hyperspectral data to predict the class labels. The experiment result of random forest with variable importance shows high accuracy of 94.93% in predicting the class labels. Conditional inference and quantile forests are also achieved high accuracy with slight difference of 94.65% and 93.46% respectively. Keywords: Data mining, Hyperspectral data, Remote sensing, Hyperspectral classification, Random forest