A Parallel Algorithm for Hyperspectral Target Detection Based on Weighted Alternating Direction Method of Multiplier

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 16; pp. 8274 - 8285
• Main Authors: Yu, Kun, Wu, Shanshan, Wu, Zebin, Sun, Jin, Zhang, Yi, Xu, Yang, Wei, Zhihui
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Alternating direction method of multiplier (ADMM); Computation; Computational modeling; Data analysis; Data processing; Deep learning; Detection; Direction; fusion; hyperspectral image (HSI); Hyperspectral imaging; Image reconstruction; Information processing; Machine learning algorithms; Multipliers; Object detection; parallel; Parallel processing; Remote sensing; Spatial data; Target detection; Task analysis
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2023.3312523

Target detection for hyperspectral images (HSIs) is one of the significant techniques in remote sensing data processing. Targets generally comprise various object categories with complex features and of varying sizes. Target detection is often used in complex application scenarios in which accurately and efficiently acquiring detection results can be challenging. The development of advanced target detection approaches is becoming increasingly necessary in both military and civilian fields. This article proposes an alternating direction method of multiplier (ADMM)-based parallel approach for hyperspectral target detection. Different from existing methods performing target detection solely on HSIs, our approach performs the fusion of hyperspectral and multispectral data to leverage both spectral and spatial information prior. For each task or data partition, the parallel processing of the computation load on multiple computing nodes can substantially reduce the computation time. In addition, we introduce a novel weighted ADMM, which takes the influence of different variables on convergence into account, to further enhance the computational efficiency of the target detection model. Experiments on real-world HSI datasets demonstrate that our proposed parallel method not only produces more accurate detection results than direct detection methods, but also achieves significant acceleration ratio compared with the serial processing flow.