Unsupervised ensemble change detection using kernel PCA

In this paper, we present a novel approach for unsupervised change detection on multi-spectral satellite images. The advantage of unsupervised approach over the supervised one is that the generation of an appropriated ground truth is not required. Especially, when the ground truth is not available, the unsupervised approach is the fundamental one. The unsupervised change detection method used in this paper is based on the concept of kernel Principal Component Analysis (PCA). The advantage of using kernel PCA over standard PCA is that it can handle non-linear relationships between data by projecting the data into higher dimension using kernel function. Assuming that relationship between data point after projecting data into higher dimensional space is linear, the principal basis vectors can then be extracted. A pixel is classified as unchanged if the absolute value of its second principal component is larger than a threshold. To improve the performance of kernel PCA, the ensemble approach is applied i.e., bootstrapped aggregating, which is also known as the bagging. The concept of bagging is to combine the change detection results from weak change detectors to improve overall performance. The proposed method begins with generating bootstrap samples. That is, the image pixels are random sampled with replacement. Each bootstrap sample is then used to construct a change detector using kernel PCA; a set of change detector is hence obtained. As a result, to classify a pixel as change or un-change, the change detection results from those change detectors are then fused by majority voting. By using random samples instead of using all image pixels at once for detecting change, the computation time is reduced. In other words, a large amount of computational resource is not required. To evaluated, the performance of the proposed methods, it was tested with multispectral satellite images i.e., Landsat 8. The experimental results will be illustrated as a change map. The analysis on principal basis vector extracted from the data using kernel PCA will be performed.