Zhan proposed a change detection method based on deep Siamese convolution networks. Wang proposed region-based CNNs to extend object detection for image change detection tasks. Yang performed change detection and land cover mapping simultaneously and used land cover information to help predict changed areas. In the field of change detections, the general end-to-end 2-D convolutional neural network effectively learned distinguishing features from higher levels by a 2-D CNN and introduced a hybrid affinity matrix fused with subpixel representations to improve its generalization ability. Because of the rapid development of graphics processing units (GPUs), deep learning methods can be applied in various fields. Traditional supervised learning methods include random forests (RFs), convolutional neural networks (CNNs), etc. Supervised learning methods use labeled training data to learn which areas changed. Overall, unsupervised methods have certain limitations in change detection research. Furthermore, unsupervised methods require targeted tuning of models to adapt to different environments, which is very time-consuming and laborious. Nevertheless, unsupervised learning methods cannot make use of prior knowledge of marked data, and relay on the assumptions of some models or similar rules to distinguish changed areas. In addition, many algorithms were improved based on the CVA. It uses a simple method to perform differential operation on image data of various wavebands in different periods, in this way, it evaluates the change of each pixel, and then forms a change vector of various wavebands. Another classic method is the change vector analysis (CVA) proposed by Malila. After several iterations, the weight of each pixel would be stable when the change is less than a set threshold and the iteration will stop if it changes no more. The unchanged pixels have larger weights through calculation, and the values of final weights are the basis for determining whether a pixel changed. The core idea of the algorithm is that the initial weight of each pixel is 1, and each iteration gives a new weight to each pixel in the two images. Based on the MAD algorithm, the IR-MAD algorithm was studied and put forward in combination with expectation-maximization (EM) algorithms. However, this algorithm still cannot completely improve current multi-element remote sensing image methods. Its main mathematical essence is canonical correlation analysis (CCA) and band math in multivariate statistical analysis. The MAD is another unsupervised multitemporal image change analysis method. If the data in the changed areas and unchanged areas are unbalanced, the model performance will be seriously affected. However, the PCA relies on statistical characteristics of images. As a linear transformation technique, this method performs decorrelation on images. The PCA is one of the most famous subspace learning algorithms. Common methods include principal component analysis (PCA) based on k-means clustering, multivariate alteration detection (MAD), and the iteratively reweighted multivariate alteration detection (IR-MAD). Generally, these methods mainly focus on the generation and analysis of differential images, extracting information from either original images or differential images to detect which areas change. The unsupervised learning methods have been applied to change detection in many recent studies.
0 Comments
Leave a Reply. |