Hyperspectral image target detection method and system based on spectral dimension and spatial cooperation neighborhood attention

Abstract

The invention discloses a hyperspectral image target detection method and system based on spectral dimension and spatial cooperation neighborhood attention. The method comprises the following steps: generating a 3D cube set; respectively taking a bidirectional recurrent neural network of spectral dimension neighborhood attention mechanism based on target identification feature self-adaptive extraction and convolutional neural network of three-dimensional neighborhood attention mechanism based on spatial structure self-adaptive extraction as spectral branch and spatial branch to respectively extract spectral features and spatial features of hyperspectral image for cascade generation; forming spatial-spectral cooperation characteristics to obtain an optimal network model; and obtaining a target detection result of the network to the data set through an activation function according to the spatial-spectral cooperation characteristics. through a neighborhood attention mechanism of spectrumdimension and space cooperation, the neural network can adaptively learn and acquire space-spectrum cooperation features, the interdependence relationship between discriminative spectrum features andsimilar space features is better mined, the generalization ability is high, and high target detection precision can be obtained.

Description

technical field [0001] The invention relates to the technical field of hyperspectral image processing, in particular to a hyperspectral image target detection method and system based on spectral dimension and spatial collaborative neighborhood attention. Background technique [0002] Hyperspectral imaging technology can not only use the imaging system to obtain the spatial dimension information of the measured object in the continuous space in the visible light to infrared wavelength range, but also capture the narrow and continuous spectral dimension information reflected or emitted by electromagnetic energy. Therefore, hyperspectral images acquired by surface observation aircraft or satellites contain a large amount of spectral and spatial information, and are often combined into a three-dimensional data cube for processing and analysis. The cube size is (x, y, b), where x and y represent space dimension, and b represents the spectral dimension (including the wavelength ra...

AI Technical Summary

Problems solved by technology

[0004] The main difficulties in hyperspectral image processing are as follows: First, hyperspectral data has many bands in the spectral dimension, and there is high information redundancy, resulting in high data dimensionality.

However, the traditional dimensionality reduction method discards the detailed information on the spectral dimension in the process of processing, so it is not desirable

Second, compared with the nanoscale spectral resolution, the spatial resolution of hyperspectral images is low, and there are a large number of mixed pixels, that is, there are multiple objects of different categories in one pixel, which need to be unmixed

In addition, due to the influence of imaging noise, atmospheric turbulence and spectral mixing factors, the hyperspectral data acquired in the actual environment often have strong nonlinearity and non-Gaussianity, which brings great difficulties to feature extraction.

Early hyperspectral target detection methods include Adaptive Cosine Estimation Detector (ACE) and Matched Filter (MF) based on the assumption of normal distribution of spectral data, as well as Constrained Energy Minimization Algorithm (CEM) based on probabilistic statistical models. These classic detection algorithms have strong generalization ability, but weak identification ability, and are not suitable for detection problems in the case of strong nonlinear distribution of spectral data.

In the past two years, the recurrent neural network (RNN) has also been gradually applied to hyperspectral image problems, filling the gap in the lack of sequence-based algorithms in hyperspectral image processing, and the hyperspectral target detection technology based on RNN machine learning methods is being generalized It is impossible to have both the ability and nonlinear feature recognition ability. For different hyperspectral data sets, different feature extraction methods need to be designed, which greatly reduces the efficiency of target detection.

Images