Underwater Target Detection From Hyperspectral Imagery

Underwater target detection works as the basic and indispensable part of hydrology information monitoring in remote sensing.Recently,with the development of hyperspec-tral sensors,especially drone-based hyperspectral sensors,it is a mainstream tendency to detect desired underwater target based on the hyperspectral image.However,hyper-spectral underwater target detection remains plenty of challenges to be tackled,such as the interference of water background and the weakness of target signal.Furthermore,the complex mechanism of underwater environment make it difficult to construct the hyper-spectral underwater environment model,which further multiplies the difficulty to find out the desired underwater targets and then leads to a poor underwater target detection performance.On the basis of the hyperspectral bathymetric model,this paper discovers the intrin-sic characteristics of hyperspectral images with deep learning technology and builds an accurate hyperspectral underwater environment model in the data-driven method.These can finally help to address the difficulties occurred during finding underwater target de-tection with hyperspectral imagery.The main research work and achievements of the whole paper are listed as follows:Firstly,according to the high relevance between adjacent bands as well as the serious redundant information in hyperspectral image,this paper proposes a bathymetric model based band selection method to pre-process the original image before applying the detec-tion algorithm.Specifically,we propose a low-dimension latent feature space learning method based on the hyperspectral bathymetric model,for the fake of solving out desired subspace from input dataset,where the underwater target pixels and the background pix-els possess best separability.After that,the band selection rule is designed on the basis of generated subspace and spectral distance metrics.Then,an iteration-based band sub-sets generation strategy is also developed to simultaneously promote the diversity of band selection results and the utilization rate of ample spectral information of original hyper-spectral image.Moreover,in order to remove redundant information contained among adjacent bands,a specific representative band selection scheme is also conducted with the help of sparse representation learning.Finally,we can achieve the band selection re-sults by connecting all the representative bands with their initial band sequence orders.Plenty of experiments have been performed on the three hyperspectral underwater image datasets,and the corresponding results confirm the effectiveness of our proposed method in comparison with the state-of-the-art hyperspectral band selection methods qualitatively and quantitatively.Secondly,for the sake of improving the unsatisfied water inherent optical properties retrieving performance,this paper develops a hybrid sequence network for unsupervised water properties estimation from hyperspectral imagery,which can figure out the esti-mation result without any prior information and then possesses a better transferability.Specifically,to get the high-dimension nonlinear sequence features of input hyperspectral spectra,we design a special sequence feature extraction network based on one dimension convolution neural network and recurrent neural network.After that,we exploit the cal-culated sequence features to predict the inherent optical properties of hyperspectral water imagery and the prediction results are employed to reconstruct the hyperspectral water spectra.Owing to the deficiency of training labels,an unsupervised training scheme is designed to adjust the parameters of sequence feature extraction network and inherent optical properties prediction network.Then,in order to achieve a more accurate inherent optical properties prediction result,we propose a hierarchical multi-scale sequence loss as the critical term of objective function.This specific loss can describe the reconstruction er-ror more accurately and evaluate the discrepancy between different hyperspectral spectra in sequence feature aspect.Finally,adequate experiments are conducted on four different datasets to reveal the effectiveness and efficiency of our proposed method.Based on the experimental results,it is effortless to figure out that our proposed method has achieved the best performances compared with the state-of-the-art water IOPs retrieving methods.Thirdly,to overcome the unpromising detection performance and the limited appli-cation prospect of existing research works,this paper proposes a self-improving frame-work for joint depth estimation and underwater target detection from hyperspectral im-agery,which can simultaneously figure out the depth information and location informa-tion of desired underwater targets,leading to better detection performances and greater algorithm universality.Specifically,exploiting the singularity of underwater targets in their corresponding spatial neighborhoods,we propose an ensemble anomaly detector to figure high-confidence target pixels for conducting the initial training data set of depth estimation network.As for the depth estimation network,it is established based on re-gression prediction method and spectral reconstruction scheme,whose output can be uti-lized to expand the training data set of subsequent detection network by down-sampling trick.In terms of the underwater target detection network,it can be considered as a bi-nary classifier,which uses the deep sequence neural network to build the target detector.Then,according to the detection results of above target detector,more target pixels can be marked to further promote the completeness of training data set for depth estimation net-work.Therefore,it is achievable to make the continuous expansion for training datasets of depth estimation network and target detection network by iteratively repeating above updating process,meanwhile the performances of depth estimation network and target detection network are also promoted with the development of training datasets.When the all parameters of framework converge,we can get the final target detection results and depth information estimation results.With the assistance of four different hyperspectral underwater target datasets,we have confirmed that the proposed algorithm can overcome the interference of water environment and achieve satisfied detection performances and accurate depth estimation results.Lastly,to tackle the weak coupling between different modules contained in the exit-ing research works,this paper proposes a novel insight to generate the underwater target detection result.The novel insight refers to transforming the underwater target detection problem into signal separation problem and employs hyperspectral unmixing technology to solve the signal separation problem.Specifically,we regard the input hyperspectral image as a linear combination of target-background mixed pixels and pure water back-ground pixels.Therefore,a particular ensemble anomaly detector is yield by integrating various classic unmixing algorithms,which is designed to eliminate the interference of water background pixels and generate the target-background mixed pixels.Then,we ex-plore the hyperspectral bathymetric model to modify traditional deep autoencoder aiming at controlling the unmixing process,and then use the modified autoencoder to unmix the target-water mixed pixels for acquiring the target-associated abundance values and maps.Moreover,considering the physical meaningless endmembers issue,a particular spectral constraint is imposed on the objective function as a training guidance.In this way,the au-toencoder would be capable of generating specific endmembers and their corresponding abundance maps.Finally,according to the physical essence of abundance maps,we figure out the detection result by fusing the outcomes of autoencoder with weight coefficients determined by abundance values.Comparative experiments on four hyperspectral water image data sets demonstrate the feasibility of the novel ideas to detect underwater tar-get and its better detection performance compared to traditional hyperspectral underwater target detection methods.