Uncertainty-Based Sample Optimization Strategies for Large Forest Samples Set

Abstract

Our study was focused on the optimization of large training samples set selected from the global forest cover change detection system. Automatically delineating training samples procedure labeled tens of millions of samples representing forests and non-forests. To improve the precision, reduce the computational complexity and avoid over-fitting, we need to select samples from the large set of tens of millions of samples that are helpful for training a classifier. In this paper, two methods were used to optimize a large sample set from the Landsat-7 ETM+ data and obtain samples for training the classifier. The first method was the traditional stratified system sampling strategy. The second was uncertainty-based sample set optimization that selects training samples based on uncertainty by examining the uncertainty measure of samples and the distribution of their feature space, and involving the subtractive clustering, KNN and support vector machine. Through precision evaluation, our experiments validated that the uncertainty-based sampling strategy can achieve better results than the stratified system sampling strategy.