A New Structure for Binary and Multiple Hyperspectral Change Detection Based on Spectral Unmixing and Convolutional Neural Network. (December 2021)

• Record Type: Journal Article
• Authors: Seydi, Seyd Teymoor; Hasanlou, Mahdi
• Is Part Of: Measurement. Volume 186(2021)
• Publication Date: 2021-12
• Subjects: Land cover -- Change detection -- Hyperspectral -- Convolutional neural network; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110137
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Performance of multiple hyperspectral change detection based on CNN and most common methods were analyzed. The Automatic framework for multiple sample generation was proposed. The proposed CNN architecture based on multi-dimensional kernel convolution (3D, 2D, and 1D) The use of the DTW predictor for the generation of a robust binary mask. Abstract: The earth is constantly being changed by natural events and human activities that constantly threaten our environment. Therefore, accurate and timely monitoring of changes at the surface of the earth is of great importance for properly facing their consequences. This research presents a new hyperspectral change detection (HCD) framework based on a robust binary mask and convolutional neural network (CNN). The proposed method is implemented in three parts: (1) the first part provides a robust binary change map based on Otsu and dynamic time wrapping (DTW) algorithms; the DTW algorithm plays a predictor role that is a robust predictor for HCD purposes. Also, Otsu's algorithm gives an estimate about the approximate threshold for detecting change and no-change class areas. These class areas will be used in the next steps. (2) The second part generates pseudo training data based on an image differencing (ID) algorithm and spectral unmixing (SU) manner for multiple change detection. This pseudo training data will be used for training the CNN model in the next step. (3) Finally, the multiple change map is generated by training the CNN network based on pseudo training data. The result of HCD maps is compared to other robust hyperspectral change detection methods by two real bi-temporal hyperspectral image datasets. The result of HCD in multiple change map shows the proposed method can have high performance compared to other HCD methods with an overall accuracy (OA) of more than 92% and Kappa coefficient (KC) of 0.77 and higher.