Research On Key Technologies For Target Detection In Aerial Hyperspectral Remote Sensing Imagery

With rich spatial and spectral information,hyperspectral images not only acquire the topographic features and spatial distributions but also detect their radiance and spectral properties,which have the characteristics and advantages of spectral-spatial integration.It has a wide range of applications in marine scientific research,precision agriculture,environmental monitoring,mineral exploration and military target detection.Since aerial remote sensing targets often show characteristics such as relatively small object occupation,diversified shapes and scales,complex image backgrounds and certain similarities with the backgrounds,it brings some challenges to realize target detection in aerial hyperspectral remote sensing images.The aviation environment is complex and volatile.Under the impacts of the spectral imager and external environment,the hyperspectral images are susceptible to distortion in different degrees.The pre-processing can effectively recover the original image information and maximize the accuracy and completeness of data,which is a prerequisite for achieving high-precision hyperspectral remote sensing target detection.The feature extraction technology can fully explore the spectral and spatial features of the image in the massive and high-dimensional data,enhance the difference between the target and the background,and improve the ability to distinguish between the same objects with different spectra and the different objects with same spectra.As a key means to improve feature extraction utilization and enhance detection performance,multi-feature fusion detection technology can comprehensively compare the extracted each spectral-spatial feature.It realizes the complementary advantages between multiple features,and then target location and scale information can be accurately detected.In order to effectively improve the target detection performance of aerial hyperspectral remote sensing images,this dissertation focuses on three key technologies of pre-processing,feature extraction and multi-feature fusion detection,and the main research contents of the dissertation are as follows:1.According to the imaging characteristics of the aerial hyperspectral imager and the aerial hyperspectral remote sensing data collected by the laboratory in the early stage,the hyperspectral image pre-processing method was studied.The radiation correction is performed on the hyperspectral images to eliminate the image defects caused by the non-uniformity of detector response and the radiation errors caused by atmospheric absorption and scattering so that the detector response to radiation values in the same intensity is consistent.The spectral bands greatly affected by noise are removed to prevent the impact on the detection performance.The correction effect is evaluated according to the image quality evaluation index.The experimental results show that the pre-processing method can effectively remove the strip noise in the image,reconstruct a pure hyperspectral image,and accurately retain the image texture details.2.In view of the characteristics of aerial remote sensing targets with different scales and irregular shapes,a spectral-spatial feature extraction method based on cross-subspace combination and differential attribute filtering is proposed.A subset of spectral bands with rich information and low correlation is selected to solve the information redundancy problem.Four morphological attributes of area,diagonal length,moment of inertia and standard deviation are selected to filter the subset of spectral bands.And the corresponding differential processing methods and threshold selection rules are designed according to the attribute types to fully characterize the target spatial structure features,and accomplish the effectiveness of background suppression and target highlighting while reducing the effects of noise,shadows and scene brightness variations on the hyperspectral images.3.To address the problems of weak characterization ability of single image features and low detection accuracy of similar spectral features,a multi-feature fusion target detection method based on kernel principal component analysis is proposed.Taking advantage of the complementary information of multi-scale and multi-type spectral-spatial features,the implicit mapping transformation is performed by kernel principal component analysis to retain as much information of each spectral-spatial feature as possible,and fully exploit the higher-order correlation between features to capture the non-linear relationship between multiple pixels such as image edges and contour curves,optimizing the optimal parameter selection process efficiently.Combining subjective visual and objective indicators with six typical target detection algorithms for comparison,the experimental results show that this method has high stability and robustness,and its average detection accuracy is improved by 4.3%compared with the GTVLRR method.4.To address the challenge of how to balance feature description completeness and redundancy,a multi-feature preferential target detection method based on genetic algorithm is proposed.A strategy of selection before fusion is used.With the technical advantages of efficient search,simple process and strong fault tolerance,the genetic algorithm is combined with band selection by cross-subspace combination and differential attribute filtering.The extracted spectral-spatial features are classified by clustering ideology to reduce the intra-class differences and increase the inter-class differences.The most representative subset of spectral-spatial features is found for weighted fusion through repeated iterations,which effectively suppresses the false alarm phenomenon and improves the detection capability of various size targets.Five aerial hyperspectral datasets containing objects of different shapes,sizes,and materials were used for performance validation,and the algorithm excelled in detection performance,algorithm stability and background-target separation.The average detection accuracy is up to 0.9991,and the detection accuracy is higher than0.9990 in beach,city and street scenes.