A Research Of Hyperspectral Target Detection Algorithm Based On Spatial Spectrum Joint

Hyperspectral remote sensing images,while acquiring rich and detailed spectral information of ground features,also include spatial structure information.HSIs can find many targets that are difficult to detect in traditional single-band images and multispectral images,which play an important role in the military and civilian fields.However,due to the imaging environment and acquisition platform,the hyperspectral image has problems such as lack of prior information,high-dimensional small samples,same-object heterospectrum,and strong noise interference.It is difficult to achieve an adaptive balance between detection performance and sample size.For the aforementioned problems,this paper addresses the needs of pixel-level small target detection in ground observation,and designs a hyperspectral target detection technology combining semantic segmentation,transfer learning,multi-scale and cascade structure.The algorithm proposed in this paper makes full use of the spatial and spectral relationship between the target and the background of the hyperspectral image,overcomes the shortcomings of target detection on a single space or spectrum,and alleviates the problem of low target detection accuracy caused by scarce samples and complex backgrounds.The main research contents and innovations of this article are as follows:(1)In order to improve the accuracy of spatial domain object detection and alleviate the lack of spatial information in hyperspectral remote sensing images,a spatial domain detection method based on semantic segmentation is proposed.This method constructs AIAS datasets with approximately homogeneous spatial distribution and trains neural network parameters based on invariant transfer learning based on spatial features.The method can extract the deep features of the target object and improve the problem of insufficient training samples of hyperspectral images.In addition,this paper builds an HSN network based on the encoder-decoder structure,and fuses the residual structure and the channel spatial attention mechanism,thereby enhancing the network information flow and highlighting the spatial characteristics and location information of the target,and also solves the problems of insufficient prior information of the target and difficulty in extracting the edge and texture information of the image.The experimental results show that the proposed HSN network can improve the ACC value of 7.93%,and the introduction of deep space feature detection method can increase the AUC value of 1.26% compared with the traditional method.(2)In order to enhance the stability of the detector to spectral changes,and to solve the problems of high-dimensional redundancy of high-spectral images and the phenomenon of homologous heterospectrality leading to high detection complexity and low accuracy,a multiscale feature extraction method based on spectral dimensions is proposed.The proposed method uses multi-scale technology to extract spectral information features in different waveband ranges,and by concatenating different fine-grained spectral features.On the one hand,it can overcome the problem of spectral distortion caused by atmospheric interference in the original pixel and enhance the robustness of the detection structure.On the other hand,the extracted features can reduce the dimension by 9.33 times and retain the distinguishing spectral features of the target and the background,thereby improving the efficiency of target detection.(3)In order to enhance the target while suppressing the background information and solving the problem of noise interference,a cascade detection method based on spatial-spectrum joint is proposed.This method combines the spatial location of adjacent pixels in the spatial domain with the spectral characteristics of a single pixel in the spectrum.A five-level detection structure is formed by cascading a regularized detector and a nonlinear suppression function.Thereby enhancing the non-linear discriminating ability of the model.In the five real hyperspectral image data sets,the average detection time of the proposed algorithm is0.1793 s,which can achieve a detection accuracy of 99.55% and a false alarm probability of1.42%.