Comparison of two approaches for land cover classification from ICESat/GLAS waveform data

Abstract

The Ice, Cloud, and land Elevation Satellite/Geoscience Laser Altimeter System (ICESat/GLAS) as a full-waveform satellite LiDAR enables land cover classification estimation. Although it has already been retired, its successor satellites will provide valuable information in climate change and icecap glaciology studies. In order to study the way of waveform-based land cover classification using machine learning methods, we utilized decision tree (DT) and Gaussian process (GP) methods to analyze the accuracy of land cover classification. DT classification is a convenient and practical method used in previous reports. GP classification can provide the state-of-the-art recognition performance using an elegant Bayesian framework. To generate the true labels of machine learning classifiers, a solution of the time-matching and location-matching between ICESat/GLAS laser footprints and LANDSAT images has been achieved. Plenty of experiments have been implemented in the Jakobshavn Glacier by evaluating classifiers features selection, training data ratio and classification confusion matrix. Experimental results show that the overall accuracy of the GP classification is solidly higher than DT classification but GP classification consumes more time. At the best training data ratio of 70%, GP classification accuracy is 92.22% which is higher than DT classification accuracy of 87.78%.