Small Target Detection And Recognition For Hyperspectral Remote Sensing Imagery

The hyperspectral imagery data obtains rich spectral feature information while obtaining the image information of the ground,and records the fine spectral features of different materials.Because different materials have different spectral characteristics,a more efficient spatial-spectral model can be established to achieve more accurate target detection and recognition.In the hyperspectral military small target detection system,the target detection and recognition technologies are the keys to accurately discover the target and obtain the enemy information.They are the key and difficult problems of information processing system and have important research value.In this paper,the research on small target detection and recognition technologies of hyperspectral imagery are carried out,and the following research contributions are obtained.(1)Aiming at the problem of great randomness caused by image simulation,a theoretical analysis model for hyperspectral target detection performance and a theoretical analysis model for hyperspectral target recognition performance were established.Based on the statistical distribution characteristics of Mahalanobis distance,the two models are mathematically derived rigorously,and the relationship between each influencing factor and detection performance or recognition performance is clarified.This model significantly improves the convenience of target detectability and identifiability analysis under different conditions.(2)Aiming at detecting subpixel-level targets with dense distribution,a hyperspectral anomaly detection method based on constrained sparse representation is proposed.By adding the sum-to-one constraint and non-negativity constraint to the weight vector,the method ensures that the model has physical meaning.By removing the upper bound constraint on sparsity level,the test pixel can be better reconstructed and the sparsity level is not needed to set.By removing the anomalous pixels from the local background,it builds a pure background dictionary.This method significantly improves the anomaly detection performance when the local background is contaminated by target signals.(3)Aiming at detecting subpixel-level targets with dense distribution,a hyperspectral target detection method based on binary hypothesis model and constrained sparse representation is proposed.By adding the non-negativity constraint to the weight vector,the method ensures that the model has physical meaning.By adding an upper bound constraint to the weight vector,the target signals in the local background are suppressed.This method fully utilizes the abundance difference and residual difference between the two hypotheses,which significantly improves the target detection performance when the local background is contaminated by target signals,and has strong robustness to the spectral radiation intensity variation.(4)Aiming at recognizing subpixel-level targets with a few training samples,a hyperspectral target recognition method based on multiscale capsule network is proposed.This method effectively extracts multiscale spatial-spectral features of target through a multiscale feature extraction network.By partially recovering the original image,the decoder network is simplified and can reduce the parameters of the network.By adding weights to the sample of each class in the loss function,the class with a large number of training samples is prevented from dominating the training process.This method significantly improves the target recognition performance under different conditions when the training samples are few and unbalanced.