Change detection method for hyperspectral image with sequential time series inputs

This paper proposes a novel change detection method of hyperspectral data based on convolutional feature extraction and long short term memory unit. To extract abstract nonlinear features of the input data, we choose deep learning-based method, Convolutional Neural Network, to avoid the drawbacks of linear methods. Data from different period of time are processed separately. A flatten layer helps the extracted feature maps to reconstruct as a new feature vector. Experiments show the effectiveness of the proposed model compared with traditional change detection methods. The proposed method performs well in both changed samples and unchanged samples especially for hyperspectral images.