Research On Efficient Target Detection For Hyperspectral Remote Sensing Image

As an important technology for obtaining and analyzing surface information,remote sens-ing is not only widely used in various fields of the national economy such as agriculture,forestry,geology,ocean,meteorology,surveying and mapping,environmental protection,disaster prevention and relief,but also plays an increasingly important role in national de-fense and military.Hyperspectral remote sensing image（HRSI）,as an important type of remote sensing data,possesses the characteristic of”spectral-spatial integration”.It can ac-quire tens to hundreds of continuous spectral bands of ground objects,and combine them with spatial information reflecting texture,morphology and other characteristics of ground objects.As one of the important applications of HRSI processing,the target detection task has attracted extensive attention and in-depth research by scholars at home and abroad.This thesis focuses on the applications of hyperspectral remote sensing（HRS）and systemat-ically studies efficient target detection algorithms for HRS.First of all,this thesis addresses the limitations of prior information and insufficient training samples encountered in HRS.A target detection algorithm based on sample mining and background reconstruction has been proposed,which fully utilizes spectral information in the background and realizes pixel-level target detection.Furthermore,considering that the spatial information of HRSI may contain discriminant features of targets,this thesis has proposed a spatial-spectral joint target detec-tion algorithm that effectively extracts and utilizes both spectral and spatial features in data.To some extent,by further improving the accuracy of target detection,it breaks through the limitations of traditional hyperspectral target detection（HTD）that only focuses on local spectral information.In addition,with the improvement of HRS platforms,the dense spectral channels and increasing spatial resolution provide HRSI with more abundant feature infor-mation,but also bring the problem of massive data.Therefore,this thesis has proposed a distributed deep learning（DDL）framework based on data parallelism,effectively improv-ing the efficiency of model training for large-scale HRSI,and thus realizing the integration of observation and processing from local spectral information to global spatial-spectral in-formation.Specifically,the main contributions are summarized as follows:（1）To address the issue of limited prior information,insufficient training samples,and com-plex background with interference in HTD tasks,this thesis has proposed an HTD algorithm based on the generative adversarial network（GAN）.Firstly,a classical detector is used to perform coarse detection on the input image,and the”pseudo-label”of the target or back-ground category is given to some pixels in the original image based on the coarse detection result.On this basis,the training sample mining strategy can be realized by constructing a”pseudo-label”sample set to provide sufficient labelled spectral samples for model training.Subsequently,a feature extraction network based on the GAN and a background reconstruc-tion network based on the autoencoder（AE）are constructed respectively.Note that the GAN model and AE model are both constructed based on the multilayer perceptron（MLP）model.The above networks are trained using the training samples in the”pseudo-label”sample set to realize the mapping of background samples in the spectral domain and latent domain,and then to reconstruct the input image.Finally,a detection network is constructed based on the trained GAN model and AE model to achieve pixel-level reconstruction of the input image.The reconstructed image is subtracted from the original input image to obtain the residual image,which is used as the basis for obtaining the final detection result.The experimental results show that the proposed algorithm fully utilizes the spectral information in hyperspec-tral data and realizes high-precision pixel-level target detection.On one hand,it addresses the problem of insufficient labelled samples required for model training.On the other hand,it effectively suppresses background interference and obtained the lowest average₂value of 0.00233 on four real HRSIs.（2）In order to address the issue of insufficient extraction and utilization of spatial informa-tion features in the hyperspectral image（HSI）of existing methods,this thesis has proposed a spectral-spatial joint target detection algorithm based on the self-attention mechanism.By utilizing the transformer model with a global receptive field,the proposed algorithm can capture global contextual information in HSI.Firstly,based on the training sample mining method proposed in the previous chapter,a”pseudo-label”sample set can be constructed.Then,the sequence sample set consisting of spectrum sequences is constructed by region selection of input image based on the”pseudo-label”sample set.Subsequently,the dimen-sion of the sequence samples is transformed to convert the spectrum vectors into a domain with a predefined dimension.Then,the vectors are fed into the transformer model to learn the semantic and positional relationships between pixels in sequences.Finally,a detection network is constructed to input the learned features into a fully connected（FC）layer for pixel-level prediction,and the prediction results are fused with the coarse detection results at the decision level to obtain the final detection results.The experimental results show that the proposed algorithm obtained the highest average₁value of 0.99692 and the lowest average₂value of 0.00088 on four real HRSIs,indicating higher detection accuracy and lower false alarm,respectively.Therefore,it achieves utilization of both spectral and spatial information,and effectively improves the detection performance.（3）The high-resolution HRSI may bring problems of long training time,high memory oc-cupancy,and difficulty in extracting global information features,especially for data-driven algorithms based on deep learning.Based on the model proposed in the previous chapter,a framework for HTD based on DDL has been proposed in this chapter.Firstly,through data parallelism,the computational tasks of HTD in network model training are split into several parts and allocated to different workers for processing,achieving distributed training based on a decentralized architecture.On this basis,the layer-wise adaptive moments optimizer for batching training（LAMB）optimizer is adopted to improve the effect of distributed training.By increasing the total batch size under the condition of ensuring model accuracy,the training time can be effectively reduced based on large-batch learning.Finally,a gradient compres-sion algorithm is incorporated into the distributed training process to reduce the bandwidth pressure caused by gradient propagation,further improving the distributed training effect under the same communication cost.The experimental results verify the effectiveness of the data parallelism strategy,large-batch learning strategy,and gradient compression strat-egy adopted in this chapter.The proposed framework can take full advantage of the data characteristics of high-resolution HRSI,which realizes better comprehensive detection per-formance.Note that the model training acceleration ratio on four real HRSIs reached 3.42times.Approximate linear acceleration can be achieved in the case of four workers,which effectively improves the efficiency of deep learning.In conclusion,this thesis solves the problem of low performance of target detection tasks in HRS.On one hand,it improves the accuracy of target detection via the full utilization of spectral and spatial information.On the other hand,it enhances the efficiency of deep learning by expanding computational scale and improving computational efficiency.This thesis can provide theoretical support for the implementation of HRS in specific tasks such as resource exploration,disaster warning,and agricultural monitoring.