Search

CN-122023929-A - Hyperspectral image land coverage classification method based on spectrum-space joint modeling

CN122023929ACN 122023929 ACN122023929 ACN 122023929ACN-122023929-A

Abstract

The invention discloses a hyperspectral image land cover classification method based on spectrum-space joint modeling, which comprises the steps of 1) obtaining hyperspectral remote sensing image data to be classified, preprocessing an original hyperspectral image, cutting a space neighborhood image block with fixed size by taking pixels to be classified as the center, constructing sample data containing space information and spectrum information for subsequent model training and testing, 2) constructing a hyperspectral image land cover classification model based on spectrum-space joint modeling, performing supervised training on the model by utilizing a training set obtained in the step 1) to obtain a target model for hyperspectral image land cover classification, and 3) inputting the hyperspectral image to be classified into the classification model trained in the step 2), outputting a corresponding land cover class prediction result, and generating a land cover classification map. According to the method, the accuracy and the stability of the hyperspectral image land coverage classification are improved through combining the modeling spectral information and the spatial information, and the method can be effectively applied to the land coverage classification task under the complex ground object scene.

Inventors

  • FU CHENCHEN
  • HONG QINGQING

Assignees

  • 扬州大学

Dates

Publication Date
20260512
Application Date
20260206

Claims (7)

  1. 1. The hyperspectral image land cover classification method based on spectrum-space joint modeling is characterized by comprising the following steps of: Step 1) acquiring hyperspectral remote sensing image data to be classified, and preprocessing an original hyperspectral image, cutting a space neighborhood image block with a fixed size by taking pixels to be classified as the center, and constructing sample data containing space information and spectrum information for subsequent model training and testing; Step 2) constructing a hyperspectral image land cover classification model based on spectrum-space joint modeling, and performing supervision training on the model by utilizing the training set obtained in the step 1) to obtain a target model for hyperspectral image land cover classification; and 3) inputting the hyperspectral image to be classified into the classification model trained in the step 2), outputting a corresponding land coverage class prediction result, and generating a land coverage classification map.
  2. 2. The hyperspectral image land cover classification method based on spectrum-space joint modeling as claimed in claim 1, wherein said step 1) comprises the steps of: Step 1.1) preprocessing operation comprises the steps of carrying out normalization processing on each spectrum band, and eliminating the influence of different band differences on model training; Step 1.2) cutting out a local image block by taking a pixel to be classified in an image as a center according to a preset space neighborhood size of 9 multiplied by 9, and constructing a model input sample containing space and spectrum dimension information; Step 1.3) labeling the sample data, and dividing the sample data into a training set, a verification set and a test set according to the proportion of 1%,1% and 98%.
  3. 3. The hyperspectral image land cover classification method based on spectrum-space joint modeling as claimed in claim 1, wherein said step 2) comprises the steps of: Step 2.1) spectrum dimension downsampling and constructing initial space-spectrum characteristics; step 2.2) modeling based on spatial multi-scale context characteristics of the dynamic grouping cavity space pyramid; step 2.3) carrying out spectrum discrimination feature modeling based on a spectrum self-attention mechanism; step 2.4) parallelly fusing the spatial multi-scale context features obtained in the step 2.2) with the spectrum discrimination features obtained in the step 2.3), splicing the features in the channel dimension direction, and realizing feature compression and scale alignment through linear mapping to obtain a spatial-spectrum fusion feature representation; step 2.5) deep feature enhancement is carried out based on a cross-layer convolution transducer, so that enhanced features are obtained; Step 2.6) carrying out global average pooling on the enhanced features obtained in the step 2.5) to obtain global feature vectors, and outputting corresponding land coverage class prediction results through a classification layer.
  4. 4. A hyperspectral image land cover classification method based on spectrum-space joint modeling as claimed in claim 3 wherein said step 2.1) comprises the steps of: step 2.1.1) the hyperspectral image sample obtained in step 1) is expressed as: ; wherein B represents the sample lot size, Representing the size of the spatial domain, S representing the number of spectral bands; the input samples are convolutionally mapped along the spectral dimension direction, and the calculation process is expressed as follows: ; In which the convolution kernel is of size The method is used for fusing adjacent spectrum band information and compressing spectrum dimensions; step 2.1.2) the initial spatial-spectral joint feature representation is obtained by the spectral dimension convolution operation of step 2.1.1) on the premise of keeping the spatial structure information unchanged: ; wherein C represents the number of channels.
  5. 5. A hyperspectral image land cover classification method based on spectrum-space joint modeling as claimed in claim 3 wherein said step 2.2) comprises the steps of: Step 2.2.1) dividing the initial spatio-spectral features F 0 along the channel dimension direction into g channel subgroups: ; Wherein each channel subgroup comprises the same number of characteristic channels, and different channel subgroups can pay attention to different types of spatial structural characteristics in a channel grouping mode, so that the diversity of spatial characteristic expression is improved; step 2.2.2) each channel subgroup The space context characteristics are extracted by arranging a plurality of cavity convolution branches in parallel, the convolution kernel size is fixed to be 3 multiplied by 3, and the calculation process is expressed as follows: ; in the formula, d i represents the void ratio corresponding to the ith convolution branch, and by setting different void ratios, the model can simultaneously capture local space detail information and large-scale context structure information in the same layer; Step 2.2.3) carrying out global average pooling on the input features of each channel subgroup to obtain a global semantic description vector, and predicting weight coefficients of different hole rate branches based on the semantic description vector: ; In the formula, Representing a weight coefficient corresponding to an ith cavity convolution branch in the g-th channel subgroup; Representing a weight coefficient corresponding to a j-th cavity convolution branch in the g-th channel subgroup; Then, weighting and fusing the output of each cavity rate branch to obtain the spatial multi-scale context characteristics: ; In the formula, Representing the attention score corresponding to the ith cavity convolution branch in the g-th channel subgroup; The dynamic weighting mode enables the model to adaptively adjust the size of the space receptive field according to the complexity of the ground object space structure in the image; Step 2.2.4) the fusion result is convolved by 1×1, batch normalized and ReLU6 compressed to the original channel number.
  6. 6. A hyperspectral image land cover classification method based on spectrum-space joint modeling as claimed in claim 3 wherein said step 2.3) comprises the steps of: Step 2.3.1) expanding the initial space-spectrum characteristics in the space dimension direction, so that each space position corresponds to one spectrum characteristic sequence, and the model is gathered in the spectrum sequence without being influenced by space; Step 2.3.2) carrying out LayerNorm normalization on a spectrum dimension S to relieve numerical scale difference among bands, wherein the normalized spectrum sequence is further divided into three fields of low frequency S1, medium frequency S2 and high frequency S3; Step 2.3.3) for each field, the module introduces a shift window self-attention mechanism, generates local subsequences by utilizing sliding windows with different scales, and executes single-head self-attention in the windows, and generates a query matrix Q, a key matrix K and a value matrix V based on the self-attention mechanism, wherein the calculation process is expressed as follows: , , ; Wherein W Q ,W K ,W V represents a learnable weight matrix for mapping input features to a query space, a key space, and a value space, respectively; the spectral attention weight is calculated as: ; In the formula, d represents a characteristic dimension, and modeling of long-range dependency relationship among different spectrum bands is realized through the attention calculation; step 2.3.4) carrying out nonlinear mapping on the attention output through residual connection and a feedforward network to obtain a spectrum characteristic representation with stronger discrimination capability.
  7. 7. A hyperspectral image land cover classification method based on spectrum-space joint modeling as claimed in claim 3 wherein said step 2.5) comprises the steps of: step 2.5.1) the input features generate queries, keys and values in the attention mechanism by a 3 x 3 convolution operation to preserve spatial structure information, the computation of which is expressed as: , , ; And calculate an attention output based on the query, the key, and the value; Step 2.5.2) introducing an expansion convolution structure into a feedforward network, and gradually increasing the void ratio along with the number of network layers to model space-spectrum context information of different scales layer by layer; Step 2.5.3) fusing the previous layer of features with the current layer of features through a gating mechanism, wherein the expression is as follows: ; where σ (·) represents the gating function, Representing a gating feature for controlling a fusion ratio of a previous layer feature and a current layer feature; Representing the current layer characteristics extracted by the layer I network; Representing fusion of previous layer features through gating mechanism With current layer characteristics And (5) outputting the characteristics obtained.

Description

Hyperspectral image land coverage classification method based on spectrum-space joint modeling Technical Field The invention relates to the field of image processing, in particular to a hyperspectral image land cover classification method based on spectrum-space joint modeling. Background The hyperspectral remote sensing image is widely applied to the fields of land coverage classification, resource investigation, environment monitoring and the like because of continuous and fine spectral information. Land cover classification is a key task in hyperspectral image applications, and aims to distinguish and identify different surface types in images according to spectrum and spatial features of features. The existing land coverage classification method mainly comprises a traditional machine learning-based method and a deep learning-based method. The traditional method generally depends on manually designed spectral features or statistical features, and complex spectral correlation and spatial structure information in hyperspectral images are difficult to fully mine. The method based on the convolutional neural network extracts the spatial features through the local convolutional operation, improves the classification performance to a certain extent, but has limited receptive fields, and is difficult to simultaneously consider the global dependency relationship and the multi-scale spatial information. In recent years, models based on self-attention mechanisms have been introduced in the field of hyperspectral image processing for modeling long-range dependencies. However, the existing method focuses on modeling of single scale or single feature dimension, and has the defects that firstly correlation modeling among different spectrum bands is insufficient, cross-band feature relation is difficult to effectively describe, secondly space multi-scale feature extraction capacity is limited, and dimensional change of complex ground objects is difficult to adapt, and thirdly effective fusion mechanisms of features of different levels are lacking to influence stability and generalization capacity of classification results. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a hyperspectral image land coverage classification method based on spectrum-space joint modeling, which can effectively combine spectrum and space information and combine cross-band feature interaction, multi-scale space modeling and cross-layer feature fusion, and solves the problems of insufficient cross-band feature modeling, limited space multi-scale expression capability and poor multi-level feature fusion effect in the prior art. The invention aims to realize the hyperspectral image land cover classification method based on spectrum-space joint modeling, which comprises the following steps of: Step 1) acquiring hyperspectral remote sensing image data to be classified, and preprocessing an original hyperspectral image, cutting a space neighborhood image block with a fixed size by taking pixels to be classified as the center, and constructing sample data containing space information and spectrum information for subsequent model training and testing; Step 2) constructing a hyperspectral image land cover classification model based on spectrum-space joint modeling, and performing supervision training on the model by utilizing the training set obtained in the step 1) to obtain a target model for hyperspectral image land cover classification; and 3) inputting the hyperspectral image to be classified into the classification model trained in the step 2), outputting a corresponding land coverage class prediction result, and generating a land coverage classification map. Further, the step 1) includes the steps of: Step 1.1) preprocessing operation comprises the steps of carrying out normalization processing on each spectrum band, and eliminating the influence of different band differences on model training; Step 1.2) cutting out a local image block by taking a pixel to be classified in an image as a center according to a preset space neighborhood size of 9 multiplied by 9, and constructing a model input sample containing space and spectrum dimension information; Step 1.3) labeling the sample data, and dividing the sample data into a training set, a verification set and a test set according to the proportion of 1%,1% and 98%. Further, the step 2) includes the steps of: Step 2.1) spectrum dimension downsampling and constructing initial space-spectrum characteristics; step 2.2) modeling based on spatial multi-scale context characteristics of the dynamic grouping cavity space pyramid; step 2.3) carrying out spectrum discrimination feature modeling based on a spectrum self-attention mechanism; step 2.4) parallelly fusing the spatial multi-scale context features obtained in the step 2.2) with the spectrum discrimination features obtained in the step 2.3), splicing the features in the channel dimension direction, and