Research On High Efficient Fusion And Anomaly Detection Methods For Hyperspectral Remote Sensing Imagery

The spectral resolution of the hyperspectral imagery(HSI)is the order of nanometers,and the HSI can detect the undetectable objects that cannot be detected in the multispectral image and panchromatic imagery(PANI).It has been applied in the geological survey,forest supervision,disaster assessment.However,because of the constraints of imaging system energy and the long distance,the spatial resolution of HSI is low,which may affect the accuracy of application results.Multi-source image fusion is an effective method to improve the spatial resolution of HSI.The panchromatic(PAN)sensors can provide the PANI with high spatial resolution.Fusion of the HSI and PANI is able to obtain a fused HSI with high spectral and spatial resolution,which is conducive to improving the accuracy of some applications.In the HSI-based applications,the detection processing is an important research topic.According to whether the prior information of targets can be obtained,the detection is divided into target detection and anomaly detection.It is difficult to obtain the prior information in practical applications.Anomaly detection can detect targets without the prior information,which is more practical and has received increasing attention.There are some difficulties that need to be explored and solved as follows:first,how to effectively fuse the HSI and PANI,and improve the spatial information of HSI while maintaining the spectral information.Secondly,for anomaly detection,how to effectively suppress background information,solve the problem of the identical material with different spectral signatures,and make full use of the spectral and spatial characteristics.Thirdly,how to study the influences of the fusion processing on detection processing.To solve these problems,this dissertation analyzes the spectral and spatial characteristics,studies hyperspectral fusion algorithms and anomaly detection processing,and verifies their performance on multiple sets of hyperspectral images.The main contributions of this dissertation are summarized as follows:1.The component substitution(CS)-based methods mix the low-pass component of PANI in the process of replacing the spatial information of HSI with PANI,which will lead to spectral distortion.To solve this problem,this dissertation introduces guided filtering into the fusion of HSI and PANI.Based on the principle of guided filtering,the PANI and the spatial component of HSI are taken as the guidance image and the input image respectively,and the spatial information of PANI is transferred to HSI.A HSI and PANI fusion algorithm based on guided filtering is proposed,and the proposed method can effectively solve the problem that the low-pass component of PANI is mixed into the fused image.On this basis,this dissertation deeply studies the characteristics of HSI,and proposes an improved guided filtering-based HSI and PANI fusion algorithm.The difference between the PANI and each band of HSI is deeply mined by using guided filtering to accurately estimate the missing spatial information in each band of HSI.The performance of the guided filtering-based fusion method is further improved.The experimental results show that the proposed method is fast and efficient,and can achieve excellent performance in improving the spatial information of HSI and maintaining the original spectral information.2.The HSI and PANI contain different and complementary spectral and spatial information for a same scene.Existing methods such as the CS-based methods and the multiresolution analysis methods extract the spatial information from the PANI without considering the spatial information of HSI,which may cause distortion.To solve this problem,this dissertation proposes a novel HSI fusion method based on structure tensor.The proposed method considers the spatial information of both HSI and PANI,and calculates the structure tensor of each pixel of PANI to extract the spatial details of PANI.To integrate the spatial information of HSI and the spatial details of PANI,the proposed method designs an optimization strategy based on the weighted structure tensor,and ensures that the spatial information injected into the HSI is sufficient and complete.Experimental results over both simulated and real datasets demonstrate that the proposed method can effectively improve the spatial information of HSI while maintaining the spectral information of the original HSI.3.The anomaly pixels of HSI have different spectral characteristics from the background pixels,and a prior information can be obtained:the spectral values of the anomaly pixels in some bands are different from the spectral values of the background pixels in the corresponding bands.Based on this fact,using the spectral characteristics of HSI,this dissertation proposes a novel Gaussian mixture model(GMM)-based anomaly detection method for HSI,which effectively solves the problems of HSI high dimension and background information interference.The proposed method partitions the HSI into some subsets of bands,and fuses the partitioned bands to accomplish HSI dimension reduction.A GMM-based anomaly extraction approach is proposed to extract anomaly pixels of each band.The extracted abnormal pixels are fused by a GMM-based weighting method,and the anomaly detection map is constructed adaptively.Experimental results show that the dimension reduction processing described in this dissertation greatly decreases the computing time and improves the detection performance.Experimental results also show that the proposed algorithm effectively removes background information,and detect the anomalies with high accuracy.4.Most detection algorithms only use spectral information,and the low spatial resolution of HSI will lead to insufficient spatial characteristics of targets.To solve this problem,this dissertation proposes a spatial and spectral information-based anomaly detection method,which makes full use of spectral and spatial characteristics to effectively reduce the phenomenon of missed detection.This method uses the proposed fusion algorithm to enhance the spatial characteristics and enhance the separability of background pixels and anomaly pixels.The top-hat transform and the black-hat transform is utilized to detect the anomalies,and a weighted structure tensor optimization strategy is designed to obtain the anomaly detection map.The experimental results show that the proposed method detects more complete targets,less missed detection phenomena,and achieves higher detection accuracy.5.Based on the proposed fusion and anomaly detection algorithms,the influence of HSI and PANI fusion processing on anomaly detection processing is studied.On the one hand,this dissertation studies the influence of fusion processing on the spectral information-based anomaly detection processing.Analysis and experimental results show that the proposed fusion processing can effectively maintain spectral information,and the fusion processing cannot significantly improve the performance of the spectral information-based anomaly detection processing.On the other hand,the influence of fusion processing on the spatial and spectral information-bansed anomaly detection processing is studied.Analysis and experimental results show that fusion processing can improve the detection performance of the spatial and spectral information-based anomaly detection processing.Fusion processing can maintain the spectral information and enhance the spatial characteristics.It provides the sufficient spectral characteristics and more spatial characteristics for detection processing,which can detect more complete targets and obtain better detection performance.