CN-122024060-A - Wetland remote sensing data cross-domain classification method based on neural network countermeasure learning

CN122024060ACN 122024060 ACN122024060 ACN 122024060ACN-122024060-A

Abstract

The invention discloses a cross-domain classification method of wetland remote sensing data based on neural network countermeasure learning, which belongs to the technical field of image processing and comprises the following steps of S1, collecting the cross-domain wetland remote sensing data, extracting spatial features and spectral features by utilizing a spatial-spectral feature extraction network, processing to obtain final features, S2, finishing domain level distribution alignment, S3, finishing class level distribution alignment, and S4, outputting a prediction result. The invention designs the self-adaptive optimizing multi-classifier, which comprises two high-density classifiers and one low-density classifier, wherein the prediction is performed by the low-density classifier instead of the high-density classifier, so that the prediction precision of the model in a target domain can be further improved.

Inventors

SONG XIUKAI
WANG WEIYUN
JIANG XIANGYANG
SU BO
LIU FANG
Dang Junzhe
GAO YUNHAO
GAO CHEN
LI ZHENGMAO
ZHU WENBO
SUN SHAN

Assignees

山东省海洋资源与环境研究院(山东省海洋环境监测中心、山东省水产品质量检验中心)
北京理工大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (8)

1. The wetland remote sensing data cross-domain classification method based on neural network countermeasure learning is characterized by comprising the following steps of: S1, acquiring cross-domain wetland remote sensing data, extracting spatial features and spectral features by using a spatial-spectral feature extraction network, and processing to obtain final features; s2, based on final features, performing confusion proportion judgment by using a confusion estimator, and performing antagonism training with a space-spectrum feature extraction network to finish domain level distribution alignment; s3, extracting prediction differences by using a self-adaptive optimizing multi-classifier, and performing antagonism training with a space-spectrum characteristic extraction network to finish class-level distribution alignment; S4, extracting the target domain features by using the trained space-spectrum feature extraction network, classifying and predicting the target domain features by using the self-adaptive optimizing multi-classifier, and outputting a prediction result.
2. The method for cross-domain classification of wetland remote sensing data based on neural network countermeasure learning according to claim 1, wherein S1 comprises the following substeps: S11, acquiring cross-domain wetland remote sensing data, extracting shallow sub-features of the cross-domain wetland remote sensing data by using a space branch of a space-spectrum feature extraction network, and extracting shallow sub-features of the cross-domain wetland remote sensing data by using a spectrum branch of the space-spectrum feature extraction network; S12, merging the shallow layer extracted by the space branch and the shallow layer sub-feature extracted by the spectrum branch to obtain a merged shallow layer sub-feature; S13, adding bidirectional position codes to the features input into the gating attention block, and respectively carrying out feature extraction on the shallow sub-features extracted by the spatial branches and the shallow sub-features extracted by the spectral branches by using the gating attention block to obtain the spatial deep features and the spectral deep features; S14, fusing the space deep features and the spectrum deep features to obtain fused deep features; and S15, adding the fused shallow sub-features and the fused deep features to obtain final features.
3. The neural network-based learning-countermeasure-based wetland remote sensing data cross-domain classification method according to claim 2, wherein in S12, the shallow sub-features after fusion are The expression of (2) is: ; Wherein, the In order to be a gating mechanism, For shallow sub-features extracted for spatial branches, Shallow sub-features extracted for spectrum branches; In the S13, two-way position coding The expression of (2) is: ; Wherein, the For the lateral position coding, For the longitudinal position coding, For the two-way position coding, As a longitudinal dimension of the feature, As a transversal dimension of the feature, The number of channels that are features; In S14, the fused deep features The expression of (2) is: ; Wherein, the As a feature of the deep layer of the space, Is a deep spectral feature; In the S15, final characteristics The expression of (2) is: ; Wherein, the Is a summation operation.
4. The method for cross-domain classification of wetland remote sensing data based on neural network countermeasure learning according to claim 1, wherein S2 comprises the following substeps: s21, mixing the characteristics of the source domain and the characteristics of the target domain based on the final characteristics to obtain mixed characteristics; S22, performing feature coding on the mixed features by using a multi-layer perceptron, and taking a feature coding result as the input of a domain confusion estimator; s23, judging the confusion proportion by using a domain confusion estimator, and performing antagonistic training with a space-spectrum characteristic extraction network to finish domain level distribution alignment.
5. The neural network-based cross-domain classification method of wetland remote sensing data based on learning countermeasure according to claim 4, wherein in S21, the characteristics after mixing The expression of (2) is: ; Wherein, the Is a vector which is the same as the characteristic shape, As a feature of the source domain, Is a feature of the target domain; In S22, the input of the domain confusion estimator The expression of (2) is: ; Wherein, the Is a multi-layer perceptron; A first loss function of the domain confusion estimator The expression of (2) is: ; Wherein, the For the number of samples to be taken, For the number of categories to be considered, For the domain confusion estimator pair number Sample number The predicted value of the individual category is used, Is that The norm of the sample is calculated, Is a true confusion value; A second loss function of the domain confusion estimator The expression of (2) is: 。
6. The neural network challenge learning based wetland remote sensing data cross-domain classification method according to claim 1, wherein S3 comprises the sub-steps of: S31, constructing a self-adaptive optimizing multi-classifier; s32, fixing parameters of a space-spectrum characteristic extraction network, training parameters of a self-adaptive optimizing multi-classifier, increasing prediction difference, and entering S33; S33, fixing parameters of the self-adaptive optimizing multi-classifier, optimizing a space-spectrum characteristic extraction network, reducing prediction difference and completing class level distribution alignment.
7. The neural network challenge learning based wetland remote sensing data cross-domain classification method according to claim 6, wherein in S31, the adaptive optimizing multi-classifier comprises a first high-density classifier, a second high-density classifier and a low-density classifier; loss function of the self-adaptive optimizing multi-classifier The expression of (2) is: ; Wherein, the For the euclidean weighted distance between the first high density classifier and the second high density classifier prediction result, For the euclidean weighted distance between the first high density classifier and the low density prediction result, For the euclidean weighted distance between the second high density classifier and the low density classifier prediction result, Is according to a first high-density classifier The weight coefficient obtained by the calculation of the function value, Is according to a second high-density classifier A weight coefficient obtained by calculating the function value; In S32, the prediction differences of the first high-density classifier, the second high-density classifier and the low-density classifier are weighted by a first weighting factor and a second weighting factor; The first weighting factor The expression of (2) is: ; Wherein, the For the prediction result of the first high-density classifier, As a result of the prediction of the second high-density classifier, Is that A function; the second weighting factor The expression of (2) is: 。
8. The method for cross-domain classification of wetland remote sensing data based on neural network countermeasure learning according to claim 1, wherein in S4, a low-density classifier is used for classifying and predicting the target domain features, and a prediction result is output.

Description

Wetland remote sensing data cross-domain classification method based on neural network countermeasure learning Technical Field The invention belongs to the technical field of image processing, and particularly relates to a wetland remote sensing data cross-domain classification method based on neural network antagonism learning. Background The coastal wetland of yellow river in Shandong province is one of the most complete wetland ecosystems in warm temperate zone of China, has the key functions of ecological protection, water and soil conservation, biological diversity maintenance and the like, accurately classifies vegetation types (such as reed, suaeda salsa and tamarix) and non-vegetation coverage (such as water body, bare beach and cultivation area), and is a core data support for ecological monitoring, wetland restoration planning and natural resource management. However, the classification of the remote sensing images in the scene faces remarkable cross-scene challenges that on one hand, the yellow river mouth wetland is influenced by the quaternary hydrologic variation, the tidal fluctuation and the climate fluctuation, the remote sensing data acquired at different times have obvious differences in spectral characteristics and texture information, and on the other hand, even in the same time period, different areas (such as near river mouth areas, intertidal zones and land edge areas) of the wetland can also cause the remote sensing characteristics of similar ground objects to be heterogeneous, and the image data of the different areas have larger area differences. In addition, the yellow river mouth wetland marking data acquisition cost is high, and the on-site investigation and remote sensing interpretation are required to be combined, so that the time and the labor are consumed. Based on the interpretation requirement of the transregional remote sensing data of the coastal wetland of the yellow river, a method is required to be provided for solving the problems. The key technology for overcoming the regional difference and improving the generalization capability of the model is to solve the interpretation task of one region by using the knowledge learned from the other region, so that large-scale annotation data do not need to be acquired for each wetland scene, and the application floor requirement of the wetland cross-scene classification is perfectly adapted. The invention provides a wetland remote sensing data cross-domain classification method based on neural network countermeasure learning based on a transfer learning technology, which is used for the cross-domain classification of the yellow river coastal wetland and has high practical significance. Disclosure of Invention The invention provides a wetland remote sensing data cross-domain classification method based on neural network countermeasure learning, which aims to solve the problem of low accuracy of the existing wetland object cross-domain classification. The technical scheme of the invention is that the wetland remote sensing data cross-domain classification method based on neural network countermeasure learning comprises the following steps: S1, acquiring cross-domain wetland remote sensing data, extracting spatial features and spectral features by using a spatial-spectral feature extraction network, and processing to obtain final features; s2, based on final features, performing confusion proportion judgment by using a confusion estimator, and performing antagonism training with a space-spectrum feature extraction network to finish domain level distribution alignment; s3, extracting prediction differences by using a self-adaptive optimizing multi-classifier, and performing antagonism training with a space-spectrum characteristic extraction network to finish class-level distribution alignment; S4, extracting the target domain features by using the trained space-spectrum feature extraction network, classifying and predicting the target domain features by using the self-adaptive optimizing multi-classifier, and outputting a prediction result. Further, S1 comprises the following sub-steps: S11, acquiring cross-domain wetland remote sensing data, extracting shallow sub-features of the cross-domain wetland remote sensing data by using a space branch of a space-spectrum feature extraction network, and extracting shallow sub-features of the cross-domain wetland remote sensing data by using a spectrum branch of the space-spectrum feature extraction network; S12, merging the shallow layer extracted by the space branch and the shallow layer sub-feature extracted by the spectrum branch to obtain a merged shallow layer sub-feature; S13, adding bidirectional position codes to the features input into the gating attention block, and respectively carrying out feature extraction on the shallow sub-features extracted by the spatial branches and the shallow sub-features extracted by the spectral branches by using the gating attention block to obtai