Main content area

Learning to match multitemporal optical satellite images using multi-support-patches Siamese networks

He, Haiqing, Chen, Min, Chen, Ting, Li, Dajun, Cheng, Penggen
Remote sensing letters 2019 v.10 no.6 pp. 516-525
algorithms, data collection, land cover, learning, least squares, prediction, remote sensing
In this study, multi-support-patches Siamese networks are proposed to match multitemporal optical satellite images under land cover changes. To adequately use spatial and spectral information, a multi-support-patches extraction block is exploited to extract multispectral central-surround regions (including visible and near-infrared bands) as inputs, and a multi-support-patches convolutional layer block (including two weight-sharing channels) is used to extract deep features. Then, the learned deep features are concatenated and output with Sigmoid probabilistic labeling predictions for similarity measure. Furthermore, matching and non-matching probabilistic maps are used to determine matches by integrating an iteratively reweighted least squares algorithm. The proposed method is compared with four other algorithms by using three multitemporal satellite image datasets (i.e., Landsat-5/8, ZY-3, and GF-1) in terms of significant land cover changes. Experimental results demonstrate that the performance of the proposed method, particularly in the number of correct matches, is 10 times improved over manually designed methods.