Multiscale Change Detection Method for Remote Sensing Images Based on Online Learning Framework

摘要

Change detection for remote sensing images is very important for urban planning, disaster evaluation etc. Traditional detection methods include supervised and unsupervised learning algorithm. A novel semi-supervised multiscale change detection method based on online learning framework is presented in this paper. Firstly, mean-variance classifier and SVM classifier are trained at the different scales of 2*2 pixels block and original pixel respectively. Initial training set is extracted from the ground truth. Secondly, the difference image is obtained according to two phase remote sensing images, and arranged by the unit of 16*16 pixel block. Image blocks are input into the mean-variance classifier and SVM classier to be detected one by one, it is cascade connection between two classifiers. The error correction rules are used to choose the mis-classified instances to retrain the classifiers. Experiment results show that the method in this paper can efficiently decrease the FN (false negative numbers) to improve the performance of change detection algorithm.