Search

CN-121999256-A - Hyperspectral image classification method

CN121999256ACN 121999256 ACN121999256 ACN 121999256ACN-121999256-A

Abstract

The invention provides a hyperspectral image classification method. The method comprises the steps of extracting spatial spectrum combined features of images by means of two layers of 3D convolution, constructing a spatial spectrum feature extraction module based on a pre-activated residual error network by means of combining 2D and 3D convolution, achieving high-level spatial semantic feature extraction capability enhancement of the images and accelerating convergence rate of models, fully excavating features of different forms extracted by each convolution layer by means of connection of a plurality of residual error modules, achieving feature complementation by means of multi-feature fusion among blocks, achieving effective fusion of shallow and deep features by means of long-distance residual error connection, further enhancing expression capability of the features, and achieving image classification by means of a Softmax classifier. The invention can effectively realize the feature reuse of hyperspectral images, obtain deep high-level spatial spectrum features with better identification performance and robustness, and improve the classification precision.

Inventors

  • LV HUANHUAN
  • SUN YULE
  • ZHANG HUI

Assignees

  • 湖州师范学院

Dates

Publication Date
20260508
Application Date
20240110

Claims (10)

  1. 1. A hyperspectral image classification method, comprising: Establishing a hyperspectral image classification model comprising a shallow feature extraction layer, a deep feature extraction layer and a fusion layer; The method comprises the steps of inputting hyperspectral images to be classified into a hyperspectral image classification model, extracting shallow features of the hyperspectral images to be classified by a shallow feature extraction layer, carrying out mixed convolution calculation on the shallow features by a mixed convolution residual error module HCRM in a deep feature extraction layer through a 3D convolution layer and a 2D convolution layer which are connected in a crossing mode through a pre-activation residual error to obtain deep features, inputting the deep features into a fusion module, inputting the shallow features into the fusion module through the pre-activation residual error to be fused with the deep features to obtain fusion features, and classifying based on the fusion features to obtain classification results of the hyperspectral images to be classified.
  2. 2. The hyperspectral image classification method as claimed in claim 1, wherein the pre-activation residual comprises: Pre-activation of the residual input layer; two sequentially connected BN-ReLU-Conv layers, wherein the first BN-ReLU-Conv layer is connected with the output end of the preactivation residual error input layer; And the pre-activation residual output layer is connected with the output end of the second BN-ReLU-Conv layer and is connected with the output end of the pre-activation residual input layer in a short connection mode.
  3. 3. The hyperspectral image classification method as claimed in claim 2, wherein the deep feature extraction layer comprises an HCRM1 module, an HCRM2 module, an HCRM3 module, and a first 3D convolution layer connected in sequence; and splicing the outputs of the HCRM1 module, the HCRM2 module and the HCRM3 module, inputting the spliced outputs into the first 3D convolution layer for convolution operation, and outputting deep features.
  4. 4. The hyperspectral image classification method according to claim 3, wherein the HCRM1 module, the HCRM2 module, and the HCRM3 module have the same structure, and each include a second 3D convolution layer, a third 3D convolution layer, a first 2D convolution layer, a second 2D convolution layer, a first reshape layer, a second reshape layer, a third reshape layer, a fourth reshape layer, a first pre-activation residual, and a second pre-activation residual; the second 3D convolution layer, the third 3D convolution layer 2, the first reshape layers, the first 2D convolution layer, the second 2D convolution layer and the second reshape layers are sequentially connected; The output of the second 3D convolution layer is input into a first pre-activation residual after being recombined by a third reshape layer, and is input into the first 2D convolution layer after being added pixel by pixel with the output of the first reshape layer; the output of the third 3D convolution layer is recombined by a fourth reshape layer and then is input into a second pre-activation residual, the output of the second pre-activation residual and the output of the first 2D convolution layer are added pixel by pixel and then are input into the second 2D convolution layer .
  5. 5. The hyperspectral image classification method as claimed in claim 1, wherein the shallow feature extraction layer comprises two fourth and fifth 3D convolution layers connected in sequence.
  6. 6. The hyperspectral image classification method of claim 1 wherein the fusion layer comprises a sixth 3D convolution layer consisting of 32 convolution kernels of size 3 x 3.
  7. 7. The hyperspectral image classification method as claimed in claim 1, wherein the hyperspectral image classification model further comprises an output layer connected to the output end of the fusion layer, the output layer comprising: A third 2D convolution layer; A maximum pooling layer connected with the output end of the third 2D convolution layer; the Softmax function connected to the output of the max pooling layer.
  8. 8. The hyperspectral image classification method as claimed in claim 1, wherein the training step of the hyperspectral image classification model includes: constructing a hyperspectral image dataset; randomly initializing the weight of a hyperspectral image classification model; Inputting the hyperspectral image dataset into a hyperspectral image classification model, and outputting a classification result of the dataset by initialized weights; Selecting the difference between the classification result of the cross entropy loss function metric high-spectrum image classification model and the real classification mark; and (3) iteratively training a hyperspectral image classification model on the data set, calculating the value of the cross entropy loss function through forward propagation, updating the weight of the hyperspectral image classification model through backward propagation, and stopping iteration when the cross entropy loss function converges to obtain a trained hyperspectral image classification model.
  9. 9. The hyperspectral image classification method as claimed in claim 8, wherein the constructing a hyperspectral image dataset comprises: Acquiring hyperspectral image data containing spectral information of a plurality of wave bands from a satellite, an airplane or other sensors; Denoising hyperspectral image data through atmospheric correction and geometric correction; Marking the denoised hyperspectral image data, and determining the category of each pixel; Enhancing the denoised hyperspectral image data through rotation, overturning and scaling operations to obtain an enhanced data set; the enhanced data set is divided into a training set, a validation set and a test set.
  10. 10. The hyperspectral image classification method as claimed in claim 9, wherein the optimizer used in the hyperspectral image classification model training process is Adam optimizer.

Description

Hyperspectral image classification method Technical Field The invention relates to the technical field of hyperspectral images, in particular to a hyperspectral image classification method. Background The hyperspectral image HSI has the characteristics of integrated patterns, rich spatial information, wide spectral band range, high resolution and the like, enhances the remote sensing earth observation capability and the ground object identification capability, and is widely applied to a plurality of fields such as military exploration, environment monitoring, precise agriculture, medical diagnosis and the like. The hyperspectral image classification HSIC is one of the basic problems in remote sensing image processing, is also the basis and key of remote sensing image analysis and interpretation, and mainly aims to identify actual ground objects from images, namely, assign unique categories to each pixel in the images. In early HSIC, traditional machine learning algorithms such as support vector machines, random forests, K-neighbors, and the like were mostly focused on spectral information only. However, the same feature may have spectral differences in different spaces, and different features may also have similar spectral characteristics. Therefore, the method ignores abundant space structure characteristics, so that a great amount of noise is often generated in the classification result, and accurate classification of complex ground features is difficult to realize. Integrating spectral and spatial information is a method that effectively improves the classification effect. Considering that related information often exists between spatially adjacent pixels, the methods such as a Markov random field and a morphological attribute profile are used for extracting the spatial information of the image, and a better effect is obtained. However, the above-mentioned spatial spectrum features based on hand-made are severely dependent on rich expertise, and the extracted shallow features have limited improvement effect on classification accuracy. In recent years, the image classification method based on deep learning DL can automatically extract abstract features from low-level to high-level semantics in images, and the features have better representation performance and more accurate subsequent classification results, so that the method has been widely applied to HSICs. As one of the most representative deep learning models, convolutional neural network CNN exhibits better performance in feature extraction and classification. The early one-dimensional convolutional neural network 1DCNN can only extract the spectral characteristics of the image for pixel-level classification of the hyperspectral image, the abundant spatial information of the image is not considered, the phenomenon of 'homospectral foreign matters' and 'foreign matters homospectrum' exist in the hyperspectral image, a better classification effect is difficult to obtain by utilizing the spectral characteristics, and in order to fully utilize the spectral and spatial characteristics of the image, the 2D convolutional neural network 2DCNN and the 3D convolutional neural network 3DCNN are sequentially proposed. The 2DCNN can extract spectrum and space characteristics, but ignores the internal relation between the two characteristics, and the 3DCNN can directly extract space spectrum characteristics, but has the problems of more network parameters, long training time and the like. Deep networks can extract higher level semantic features that are more abstract at a deeper level than shallow networks. However, in the deepening of the network hierarchy, the performance of the network gradually decreases and a model degradation phenomenon easily occurs. The residual connection of the residual network ResNet in DL is achieved by adding an identity mapping layer to the shallow network, deepening the network depth and optimizing the model structure. However, the above method cannot obtain a characteristic representation with stronger discrimination due to the insufficient utilization rate of the extracted characteristic caused by directly carrying out convolution operation on the data, so that the effect of hyperspectral image classification is poor. Disclosure of Invention The invention provides a hyperspectral image classification method which can solve the technical problem that the hyperspectral image classification effect is poor because deep characteristic representation with stronger discrimination cannot be obtained in the prior art. The invention provides a hyperspectral image classification method, which comprises the following steps: Establishing a hyperspectral image classification model comprising a shallow feature extraction layer, a deep feature extraction layer and a fusion layer; The method comprises the steps of inputting hyperspectral images to be classified into a hyperspectral image classification model, extracting shallow features of the hy