Heterogeneous Images Change Detection Method based on Hierarchical Extreme Learning Machine Image Transformation

Due to the complementary information between different imaging mechanisms, heterogeneous image change detection is a challenging and hot topic compared to homogeneous image change detection. Its application is widespread, especially in emergency situations caused by natural disasters. To address the limitations of existing methods such as susceptibility to noise, manually selecting samples, and time-consuming computation, we propose a change detection method for heterogeneous images based on image transformation using Hierarchical Extreme Learning Machine (HELM). In our method, a HELM transformation model between heterogeneous images is constructed, which transforms the image features of one image into the feature space of the other one. Consequently, the transformed multi-temporal images could be comparable. Specifically, first, the logarithmic transformation of SAR images is carried out to obtain the same noise distribution model with the optical image. These heterogeneous images are smoothed to reduce the impact of image noise. Then, through image segmentation, the unchanged areas are selected as training samples. And a correction model for training samples is constructed to avoid manual selection of samples and improve the accuracy of image transformation. Subsequently, the corrected training samples are used to train the HELM to obtain the multi-temporal transformation images, which avoids the parameter adjustment of neural networks. Finally, the changes could be extracted by comparing the transformed multi-temporal images. To prove the effectiveness of the method, two sets of heterogeneous images (Google Earth and Sentinel 1 images) are used for experimental validation in this paper. The results show that the kappa coefficients of the method for the two data sets are improved by 6.19% and 8.94%, respectively, compared with the existing methods, which proves the effectiveness of the proposed method. © 2022, Science Press. All right reserved.