Research On Deep Learning Based Target Detection In Hyperspectral Imagery

Since the 1980s,the advances in sensors,digital electronics,and computing capabilities have prompted the development of hyperspectral imaging technology.Hyperspectral imaging integrates images with spatial structures and spectra with radiance characteristics,which has received widespread attention in military reconnaissance,mineral surveying,archaeology and other fields,having important theoretical significance and practical value.Hyperspectral target detection（HTD）aims to perform quantitative interpretation and analysis of hyper-spectral images（HSIs）based on prior-known spectral or spatial-spectral information,and determine the presence of the interested target in each pixel.Affected by factors such as atmospheric conditions,sensor noise,material composition and high dimension,HSIs are typically with complex nonlinear structure and spectral distortion（including"foreign spectral same matter"and"same spectral foreign matter"）,making it difficult to perform effective target detection with a single spectrum.Several studies generally alleviate the aforemen-tioned issue by estimating data distributions with pre-defined models.However,the detection performance would drop dramatically when the theoretical models are insufficient to give accurate descriptions of real data.In practice,HTD can be categorized into same-scene detection and cross-scene detection according to whether the training data and test data belong to the same scene.To address the low detection performance caused by spectral distortion in these two situations,this dissertation firstly carries out researches on same-scene and cross-scene HTD by means of discriminative feature extraction,which could learn features with a larger distinction between target and background from all-bands information;then,further develops spectral consistency enhancement for same-scene HTD,which can not only reduce the intraclass spectral differences but also retain all-bands information,leading to better target extraction using specific bands.On the basis of the aforementioned studies,this dissertation builds several unsupervised and semi-supervised deep neural networks to adaptively learn data distributions using large amounts of unlabeled data,also mine hidden nonlinear and spatial-spectral joint features in the HSIs,which can improve the feature representation ability and help to separate the interested target from the complex background with high accuracy.The main contributions are shown as follows:（1）In order to address the issue of low detection performance caused by spectral dis-tortion in the same-scene HTD,this dissertation proposes a discriminative feature extraction method based on a fully connected autoencoder.To enlarge the target-background differ-ences in the feature space,a discriminative loss function using the target prior is specifically designed,which could increase the interclass distinguishability under the weighting of target-background distances in the original HSIs,thereby easing target extraction from complex background.Moreover,to address the issue of indistinguishable target and background with very serious"same spectral foreign matter"using only spectral information,this dissertation presents a multi-scale spatial-spectral feature extraction method based on a three-dimensional residual convolutional autoencoder to additionally provide spatial discrimination informa-tion,which could effectively decrease false alarms.The experimental results show that both the proposed methods can extract features with stronger discriminative ability,and the target-background differences contained in the multi-scale spatial-spectral features are more significant than that in the solely spectral features,thus greatly improving the detection performance.（2）In order to further efficiently address the issue of indistinguishability between target and background with severe"same spectral foreign matter"in the same-scene HTD,this dissertation proposes a region of interest（Ro I）feature extraction network,which could make the network focus on identifying target and hard-detected background,and improve the feature learning efficiency.Firstly,an Ro I feature transformation（RFT）layer is constructed to extract spatial Ro I features by excluding irrelevant regions without target;then,a multi-scale spectral attention（MSA）layer is designed to learn multi-scale spectral features from the spatial Ro I features and extract more discriminative spectral Ro I features by adaptively weighting across the spectral dimension.The experimental results indicate that the proposed method could effectively improve detection performance,especially on the Indian Pines data with very high similarity between target and background.The highest overall detection accuracy,AUC_OAof the proposed method can reach 1.7080,which is far better than 1.3938of the best comparing algorithm.（3）In order to address the issue of weak generalization of single model,expensive time costs of model retraining,and low utilization of existing knowledge in the cross-scene HTD,this dissertation introduces a domain adaptive few-shot learning method for cross-scene HTD,which improves the detection performance in the target domain with the ability of judging sample-wise similarity learned in the source domain.Firstly,a source network consisting of a modulated deformable spatial-spectral feature fusion and residual channel attention block is designed,which could adaptively extract useful discriminative spatial-spectral features ac-cording to the spatial structures,and learn the pixel-pair similarity-dissimilarity measurement more effectively;then,apply a weighted domain adaptation strategy to reduce the distribu-tion shift between source data and target data,and alleviate the performance loss caused by transferring model from the source domain to the target domain.Moreover,a boosted discriminative loss function is constructed to drive the indistinguishable samples closer to or farther away the target,thus weakening the degrees of spectral distortion.The experimental results show that the proposed method could save about half the time for cross-scene HTD compared to network retraining.（4）In order to address the issue of limited detection accuracy in specific-bands HTD,which is caused by the infeasibility of obtaining specific spectral information from the re-grouped discriminative features with reduced spectral dimension,this dissertation presents a hierarchical denoising autoencoder to enhance target spectral consistency and make the intraclass spectral curves more compact without losing any spectral information.Then,a same-scene detector independent of data distribution is proposed based on the unconstrained linear mixture model.It highlights target and suppresses background through a two-step subspace projection,and helps obtain steady detection performance in complex scenarios.The experimental results and further analysis demonstrate that the hierarchical denoising autoencoder could overcome the spectral distortion including"foreign spectral same mat-ter"and"same spectral foreign matter",which plays a critical role in improving detection performance.In conclusion,this dissertation effectively addresses the issue of low detection perfor-mance caused by spectral distortion in HTD,and obtains higher accuracy than the comparing methods in the same-scene and cross-scene applications.This work could provide theoretical guidance and technological support for HTD in various tasks such as city planning,mineral surveying,search and rescue,and danger prediction.