A Study of Target Detection in Remote Sensing Images

• Main Authors: Tang, Zay-Shing, 唐再興
• Other Authors: Chang, Lena
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/bnrk6n

博士 === 國立臺灣海洋大學 === 電機工程學系 === 102 === In recent years, although the advances in sensor technology make remote sensing images have significantly improved spatial and spectral resolutions, many unknown and undefined targets, referred to as interferences, are also unexpectedly acquired by remote sensing sensors. Therefore, the main objective of the research is to improve detection capability of desired targets effectively. In this dissertation, we propose target detection methods for synthetic aperture radar (SAR) images and hyperspectral images, respectively. The proposed methods can improve target detection accuracy which can improve quality of further analysis results in remote sensing images. In SAR images, we propose a region-based oil spill detection method with data modeling incorporated, which can improve the speckle noise problem encountered in pixel-based detection methods. We apply moment-preserving method to partition SAR images into some proper regions. Then, according to these segmentation results which contain oil spills and sea area, we build the data models for oil spills and sea area, respectively. Next, based on the built data models, we propose a region-based oil detection method by the use of the generalized likelihood ratio test (GLRT) decision rule. Under the condition of constant false alarm rate (CFAR), we may determine a threshold automatically to discriminate oil spills and sea regions. Finally, the experimental results from SAR images validate that the proposed method exhibits excellent detection performance both for trained and untrained images to automatically detect oil spills. Because hyperspectral images have many dimensionalities, it is difficult to build a specific target model. Furthermore, target signature is uncertain due to atmosphere interference or other random noise and target detection performance is seriously affected by the uncertainty of target signatures. To overcome the aforementioned problems, we propose a target detection method, which incorporates signal subspace projection (SSP) with the spatial filter based on linearly constrained minimum variance (LCMV) principle. Instead of using a single constraint on target detection, we first design an optimal filter with multiple constraints by using SSP. Then, by projecting the weights of the detection filter on the signal subspace, the proposed SSP can reduce estimation errors in target signatures and alleviate the performance degradation caused by the uncertainty of target signatures. Furthermore, the SSP approach can simultaneously detect desired targets, suppress undesired targets and minimize the interference effects. In the experiments, we provide three schemes in selecting multiple constraints of the desired target: K-means, principal eigenvectors and endmember extracting techniques. Simulation and experimental results show that the proposed SSP with K-means schemes has the best detection performance, as compared to some existing methods. Furthermore, the proposed SSP with multiple constraints is less sensitive to the uncertainty of target signatures.