Wetland feature extraction system and method based on multi-source remote sensing image and deep learning

Abstract

The invention belongs to the technical field of remote sensing image processing, and particularly relates to a wetland feature extraction system and method based on a multi-source remote sensing image and deep learning, and the system comprises a data acquisition module, a feature fusion module, a CNN model training module and a feedback control module. Dynamically optimizing the feature fusion matrix by adopting a principal component analysis method; constructing a depth feature extraction network, and performing training in combination with a gradient descent algorithm; and feature extraction parameters are fed back and adjusted in real time by using normalized dot product similarity, and feature extraction precision is improved through local consistency optimization. According to the method, the bottlenecks of insufficient precision and limited feature expression ability in wetland feature extraction of a traditional method are broken through, accurate extraction of multi-scale and multi-dimensional wetland features is realized, and high-precision feature data support is provided for applications such as wetland monitoring and ecological evaluation.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and specifically relates to a wetland feature extraction system and method based on multi-source remote sensing images and deep learning. Background Art

[0002] Wetlands, as vital ecosystems on Earth, play a key role in maintaining ecological balance and improving climate change. However, wetland environments are complex and dynamic, and traditional methods for extracting wetland information, such as manual interpretation and shallow machine learning models, suffer from low efficiency and inaccuracy. With the rapid development of remote sensing technology, high-resolution remote sensing imagery has made it possible to accurately extract wetland information.

[0003] Prior art wetland feature extraction has gradually shifted from traditional manual interpretation to automated methods based on neural networks. For example, CN119399638A proposes a neural network wetland extraction system based on multi-source remote sensing image fusion. This system fuses optical and radar features through principal component analysis (PCA) and employs a three-layer fully connected neural network for classification. Its preprocessing requires a radiometric correction error of ≤±3% and a geometric registration accuracy of 0.5 pixels. While this method improves processing efficiency, it still has significant limitations: First, feature fusion relies on a static PCA threshold (85%), which fails to account for the differences in feature sensitivity among different wetland types, resulting in information loss. Second, the network structure is simple, making it difficult to capture wetland spatial contextual features and susceptible to fragmentation noise. Third, the feedback mechanism triggers retraining solely based on the accuracy of the confusion matrix, lacking dynamic optimization of the feature extraction process, which affects model adaptability. Consequently, existing methods often employ shallow feature extraction and static fusion strategies, making it difficult to effectively capture the complex spatial structure and multi-temporal variation of wetland features, resulting in limited extraction accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a wetland feature extraction system and method based on multi-source remote sensing images and deep learning, so as to solve the problems of insufficient accuracy and insufficient feature utilization in wetland feature extraction by traditional methods.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] In the first aspect, the present invention proposes a wetland feature extraction system based on multi-source remote sensing images and deep learning, comprising:

[0007] Data acquisition module, used to collect optical remote sensing images and radar remote sensing images and perform correction and multi-view processing to form a target data set;

[0008] a feature fusion module, configured to extract optical features and radar features from the target data set, fuse the optical features and radar features using principal component analysis and preset local feature weights, and determine a fusion feature matrix based on a PCA threshold;

[0009] A CNN model training module is used to build a recognition model based on a gradient descent algorithm and a convolutional neural network, process the fused feature matrix according to a learning rate decay cycle and the recognition model, and determine a wetland extraction result;

[0010] A feedback control module is used to determine the normalized dot product similarity between the fused feature matrix and the wetland extraction result, and dynamically adjust the PCA threshold, learning rate decay period or local feature weight based on the relationship between the normalized dot product similarity and a preset threshold interval.

[0011] Furthermore, the optical remote sensing image includes blue, green, red and near-infrared bands, and the radar remote sensing image includes VV and VH polarization channels.

[0012] Furthermore, the data acquisition module includes:

[0013] The first acquisition unit is used to determine whether the acquisition time interval of the optical remote sensing image and the radar remote sensing image meets the preset overlapping interval. If so, the preset ecological period table is searched according to the target wetland type. Otherwise, the hydrological characteristics of the missing period are supplemented by linear interpolation.

[0014] The first processing unit is used to perform radiation correction, geometric correction and atmospheric correction on the optical remote sensing image in sequence;

[0015] The second processing unit is used to perform polarization matrix decomposition and multi-view processing on the radar remote sensing image.

[0016] Furthermore, the feature fusion module includes:

[0017] Optical feature extraction unit, used to extract spectral features, texture features and vegetation index from optical remote sensing images. Vegetation index includes MSAVI and NDWI to obtain optical features.

[0018] A radar feature extraction unit is used to extract backscatter coefficients and polarization decomposition features from radar images to obtain radar features;

[0019] The fusion execution unit is used to fuse the optical features and radar features using principal component analysis and preset local feature weights, dynamically adjust the PCA contribution rate threshold, and obtain the fusion feature matrix F, whose dimensions meet the following requirements:

[0020]

[0021] Where λ is the eigenvalue, k is the number of retained principal components, n is the total dimension of the original feature, and λ i is the eigenvalue corresponding to the i-th principal component, Threshold PCA It is the preset cumulative contribution rate threshold.

[0022] Furthermore, the CNN model training module includes:

[0023] A model building unit, used to build a recognition model based on a convolutional neural network;

[0024] A gradient descent algorithm unit, used to optimize the recognition model through a gradient descent algorithm;

[0025] A learning rate adjustment unit, configured to adjust the learning rate according to a preset learning rate decay period;

[0026] The feature processing unit is used to input the fused feature matrix into the recognition model for processing to determine the wetland extraction result matrix R.

[0027] Furthermore, the feedback control module includes:

[0028] The similarity calculation unit is used to calculate the normalized dot product similarity S between the fusion feature matrix F and the wetland extraction result matrix R in real time. The expression is:

[0029]

[0030] Dynamic optimization unit: When S < the first threshold, perform at least one of the following adjustments:

[0031] (1) Lower the PCA contribution rate threshold to the second threshold and increase the number of principal components;

[0032] (2) Decrease the learning rate cycle to a set number of times;

[0033] (3) Increase the weight of the MSAVI index to a set multiple.

[0034] Furthermore, the feedback control module also includes a consistency optimization unit for

[0035] Scan the wetland classification result matrix R with the preset pixels as the window, and calculate the Euclidean distance D between the features in the window and the standard template;

[0036] When D>the third threshold, the features in the window are corrected according to the following formula:

[0037] F corrected =F original +0.3×(F template -F original )

[0038] Among them, Foriginal is the original eigenvector to be corrected, F template is the standard wetland feature template, F corrected is the optimized feature vector;

[0039] The correction is repeated until the D value decrease rate is lower than the fourth threshold.

[0040] In a second aspect, the present invention proposes a wetland feature extraction method based on multi-source remote sensing images and deep learning, which is implemented based on any of the wetland feature extraction systems described above. The method includes the following steps:

[0041] S1. Obtain optical and radar remote sensing images of the wetland area, perform radiometric, geometric, and atmospheric corrections on the optical remote sensing images, and perform multi-view processing and polarization calibration on the radar remote sensing images.

[0042] S2. Extract spectral features, texture features, and vegetation indices from optical remote sensing images, and extract backscatter coefficients and polarization decomposition features from radar remote sensing images. Use principal component analysis and preset local feature weights to fuse the optical and radar features to obtain a fused feature matrix.

[0043] S3. Build a recognition model with convolutional layers and train the model using stochastic gradient descent and k-fold cross validation.

[0044] S4. Input the image to be classified into the trained model for wetland extraction, calculate the normalized dot product similarity S between the fusion feature matrix F and the extraction result matrix R, and dynamically adjust the PCA threshold, learning rate decay cycle, or local feature weight based on the relationship between the normalized dot product similarity and the preset threshold range;

[0045] S5. Optimize the local consistency of the classification results: calculate the Euclidean distance D between the feature and the standard template using a preset pixel window. When D is greater than the third threshold, correct the features in the window according to the preset formula until the decrease rate of the D value is lower than the fourth threshold.

[0046] Furthermore, in step S1, the data acquisition module further includes judging the acquisition time interval of the optical remote sensing image and the radar remote sensing image. If the preset overlapping interval is not satisfied, the hydrological characteristics of the missing period are supplemented by linear interpolation.

[0047] The beneficial effects of the present invention are:

[0048] 1. This invention innovatively fuses the spectral characteristics of optical images with the polarimetric characteristics of radar images and employs a dynamic principal component analysis algorithm. The system adaptively optimizes the feature fusion process, effectively addressing the problem of insufficient feature representation in traditional methods. By introducing a dynamic feedback mechanism, the system calculates the similarity between the fused features and the classification results in real time and intelligently adjusts key parameters such as the PCA threshold and learning rate to ensure the model maintains optimal performance. This closed-loop optimization design not only improves the accuracy of feature extraction but also significantly reduces the need for manual intervention, achieving truly intelligent processing.

[0049] 2. Through its unique local window scanning and feature correction algorithm, the system automatically detects and corrects anomalous areas in classification results, effectively eliminating the fragmented classification errors common in traditional methods. This method, particularly for wetland boundary areas, maintains the integrity and continuity of the spatial structure of the features, significantly improving the credibility and usability of the classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the wetland feature extraction method proposed in the present invention;

[0051] Figure 2 This is a framework diagram of feedback control in the wetland feature extraction system proposed by the present invention;

[0052] Figure 3 It is the overall system block diagram of the wetland feature extraction system proposed in the present invention;

[0053] Figure 4 It is a feedback control block diagram of the wetland feature extraction system proposed in the present invention. DETAILED DESCRIPTION

[0054] The present application is described in further detail below in conjunction with the accompanying drawings. It is necessary to point out that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technicians in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0055] like Figure 1-4 As shown, this embodiment proposes a wetland feature extraction system based on multi-source remote sensing images and deep learning, including:

[0056] The data acquisition module is used to collect optical remote sensing images (such as Sentinel-2 or Landsat-8 data, including blue, green, red, and near-infrared bands, with a spatial resolution better than 10 meters); and radar remote sensing images (such as Sentinel-1 VV and VH polarization data, with a spatial resolution better than 20 meters); and perform correction and multi-view processing to form the target data set.

[0057] In a specific embodiment, the data acquisition module includes:

[0058] The first acquisition unit is used to determine whether the acquisition time interval of the optical remote sensing image and the radar remote sensing image meets the preset overlapping interval (for example, there is an overlapping part of 30 to 90 days between the acquisition time intervals of the two to ensure the time consistency of the data). If it meets the requirements, the preset ecological period table is queried according to the target wetland type (such as swamps, mudflats, mangroves) to match the optimal analysis period. Otherwise, the linear interpolation method is used to fill in the hydrological characteristics of the missing period. The system will use the linear interpolation method to fill in the hydrological characteristics of the missing period based on historical hydrological data (such as precipitation and water level records).

[0059] The first processing unit is used to perform radiometric correction (using the FLAASH model, with a radiometric calibration error of ≤±3%), geometric correction (based on ground control point registration, with an accuracy of 0.5 pixels), and atmospheric correction (surface reflectance error <±2%) on the optical remote sensing image.

[0060] The second processing unit is configured to perform multi-look processing (equivalent number of views ENL ≥ 4) and polarization calibration to reduce speckle noise and eliminate systematic deviations of cross-polarization channels (HV / VH).

[0061] The system also includes a feature fusion module, which is used to extract optical features (spectral features, texture features and vegetation indices such as MSAVI and NDWI) and radar features (backscatter coefficient and polarization decomposition features) from the target data set, and fuse the optical features and radar features using principal component analysis (PCA) and preset local feature weights, and determine the fusion feature matrix based on the PCA threshold.

[0062] In a specific embodiment, the feature fusion module includes:

[0063] Optical feature extraction unit, used to extract spectral features, texture features and vegetation index from optical remote sensing images. Vegetation index includes MSAVI and NDWI to obtain optical features.

[0064] A radar feature extraction unit is used to extract backscatter coefficients and polarization decomposition features from radar images to obtain radar features;

[0065] The fusion execution unit is used to fuse the optical features and radar features using principal component analysis and preset local feature weights, dynamically adjust the PCA contribution rate threshold, and obtain the fusion feature matrix F, whose dimensions meet the following requirements:

[0066]

[0067] Where λ is the eigenvalue, k is the number of retained principal components, n is the total dimension of the original feature, and λ i is the eigenvalue corresponding to the i-th principal component, Threshold PCA It is the preset cumulative contribution rate threshold (such as set to ≥85%).

[0068] It is understandable that in the feature fusion module, the optical feature extraction unit, when specifically implemented, includes:

[0069] (1) Spectral feature extraction: extract the reflectance of blue, green, red, and near-infrared bands from optical images (such as Sentinel-2) and calculate the derived indices: NDWI (Normalized Difference Water Index) and MSAVI (Modified Soil Adjusted Vegetation Index).

[0070] (2) Texture feature extraction: Calculate the texture index of a 7×7 window based on the gray-level co-occurrence matrix (GLCM): Contrast: reflects the edge clarity of the object; Entropy: represents the complexity of the image information.

[0071] (3) Radar feature extraction, backscatter coefficient: extract σ0 (backscatter coefficient, unit dB) of VV and VH polarizations, and calculate the ratio VV / VH to enhance vegetation structure information (typical value >2 dB for forest wetlands); polarization decomposition characteristics: use H / A / α decomposition to extract the following parameters: Entropy (H): measures the randomness of scattering (0≤H≤1), and high entropy values (H>0.7) indicate complex scatterers (such as submerged vegetation); α angle: reflects the dominant scattering mechanism (0°~90°), with a typical value of 40°~60° for wetlands (dominated by volume scattering).

[0072] (4) Fusion execution, principal component analysis (PCA):

[0073] ① Combine optical features (such as NDWI, MSAVI, texture) with radar features (such as σ0 VV , H, α) are merged into an n-dimensional feature matrix X m*n (m is the number of samples).

[0074] ②Calculate the covariance matrix C = X T X / (m-1), solve for the eigenvalue λ i and eigenvectors.

[0075] ③Dynamically adjust the PCA threshold PCA (Default 85%), retain the number of principal components k that meet:

[0076]

[0077] ④ Apply 1.2 to 1.5 times the weight to the first three principal components to enhance their contribution to wetland classification.

[0078] (5) Output fusion feature matrix F

[0079] Dimension m×k (k<<n), for example, the original features n=15 (5 spectral + 4 texture + 6 radar), retaining k=5 principal components.

[0080] The system also includes a CNN model training module, which is used to build a recognition model based on the gradient descent algorithm and convolutional neural network, process the fused feature matrix according to the learning rate decay cycle and the recognition model, and determine the wetland extraction results.

[0081] In a specific embodiment, the CNN model training module includes:

[0082] A model building unit, used to build a recognition model based on a convolutional neural network;

[0083] A gradient descent algorithm unit, used to optimize the recognition model through a gradient descent algorithm;

[0084] A learning rate adjustment unit, configured to adjust the learning rate according to a preset learning rate decay period;

[0085] The feature processing unit is used to input the fused feature matrix into the recognition model for processing to determine the wetland extraction result matrix R.

[0086] It is understood that in the CNN model training module, the Adam optimizer (initial learning rate η = 0.01, β1 = 0.9, β2 = 0.999) is used during implementation. The decay strategy is: decay by 10% every 10 cycles.

[0087] In the feature processing unit, the training process includes: inputting the fused feature matrix F, outputting the predictions through the CNN forward propagation, calculating the loss L, and backpropagating the weights. The output is the wetland extraction matrix R: the same size as the input image, with each pixel representing the wetland category label (e.g., 0 = non-wetland, 1 = swamp, 2 = mangrove).

[0088] The system also includes a feedback control module for determining the normalized dot product similarity between the fused feature matrix and the wetland extraction result, and dynamically adjusting the PCA threshold, learning rate decay cycle or local feature weight based on the relationship between the normalized dot product similarity and the preset threshold interval.

[0089] In a specific embodiment, the feedback control module includes:

[0090] The similarity calculation unit is used to calculate the normalized dot product similarity S between the fusion feature matrix F and the wetland extraction result matrix R in real time. The expression is:

[0091]

[0092] S∈[0,1], the closer the value is to 1, the closer the feature matches the result.

[0093] Dynamic optimization unit: When S < the first threshold, perform at least one of the following adjustments:

[0094] (1) Lower the PCA contribution rate threshold to the second threshold and increase the number of principal components;

[0095] (2) Decrease the learning rate cycle to a set number of times;

[0096] (3) Increase the weight of the MSAVI index to a set multiple.

[0097] Combine Figure 4 ,It can be understood that, in the dynamic optimization unit, the threshold trigger mechanism includes:

[0098] Preset threshold range (needs to be calibrated through experiments):

[0099] The first threshold S low =0.7: If S<0.7, the following adjustments are triggered:

[0100] Lower the PCA threshold from 85% to 75% to retain more principal components, enhance feature expression, and reduce information loss;

[0101] Adjust the learning rate decay cycle from every 10 epochs to every 5 epochs to accelerate convergence and avoid falling into local optimality;

[0102] The MSAVI weight was increased from 1.0 to 1.5 to enhance the contribution of vegetation characteristics to wetland classification.

[0103] In a specific embodiment, the feedback control module further includes a consistency optimization unit for

[0104] The wetland classification result matrix R is scanned with a preset pixel (such as 5*5) as a window, and the eigenvector F in the window is calculated. original With the standard template F original The Euclidean distance D of

[0105] When D>the third threshold, the features in the window are corrected according to the following formula:

[0106] Fcorrected =F original +0.3×(F template -F original )

[0107] Among them, F original is the original eigenvector to be corrected, F original is the standard wetland feature template, F corrected is the optimized feature vector;

[0108] The correction is repeated until the D value decrease rate is lower than the fourth threshold.

[0109] For the above consistency optimization unit, as an example, a scenario: in the classification results of a certain swamp wetland, some pixels are misclassified as farmland (due to spectral similarity).

[0110] Action: Calculate the misjudgment window D = 1.5 (> 1.2), triggering correction;

[0111] After 3 iterations of correction, D dropped to 0.8, with a decrease rate of 4% (<5%), and the optimization was stopped.

[0112] It can be understood that the consistency optimization unit in this application significantly improves the spatial continuity and accuracy of the wetland classification results through local window scanning and feature correction. The unit calculates the Euclidean distance (D) between the classification result and the standard wetland template. When D exceeds the threshold (such as 1.2), a progressive correction formula is used to adjust the abnormal features until the D value decreases below the set threshold (such as 5%). This process effectively eliminates fragmented classification errors (such as misjudging the edge of a water body as bare land), improves the boundary accuracy of the classification results by more than 20%, and maintains the integrity of the spatial structure of the land object.

[0113] Combine Figure 1 This embodiment also proposes a wetland feature extraction method based on multi-source remote sensing images and deep learning. The method is implemented based on the above-mentioned wetland feature extraction system and includes the following steps:

[0114] S1. Obtain optical and radar remote sensing images of the wetland area, perform radiometric, geometric, and atmospheric corrections on the optical remote sensing images, and perform multi-view processing and polarization calibration on the radar remote sensing images.

[0115] S2. Extract spectral features, texture features, and vegetation indices from optical remote sensing images, and extract backscatter coefficients and polarization decomposition features from radar remote sensing images. Use principal component analysis and preset local feature weights to fuse the optical and radar features to obtain a fused feature matrix.

[0116] S3. Build a recognition model with convolutional layers and train the model using stochastic gradient descent and k-fold cross validation.

[0117] S4. Input the image to be classified into the trained model for wetland extraction, calculate the normalized dot product similarity S between the fusion feature matrix F and the extraction result matrix R, and dynamically adjust the PCA threshold, learning rate decay period, or local feature weight based on the relationship between the normalized dot product similarity and the preset threshold range;

[0118] S5. Optimize the local consistency of the classification results: calculate the Euclidean distance D between the feature and the standard template using a preset pixel window. When D is greater than the third threshold, correct the features in the window according to the preset formula until the decrease rate of the D value is lower than the fourth threshold.

[0119] In step S1, the data acquisition module further includes judging the acquisition time interval of the optical remote sensing image and the radar remote sensing image. If the preset overlapping interval is not satisfied, the linear interpolation method is used to fill in the hydrological characteristics of the missing period.

[0120] Based on the above technical solution, the present invention proposes a wetland feature extraction method based on multi-source remote sensing imagery and deep learning. Its core innovation lies in the construction of a dynamic closed-loop optimization system, achieving full-process adaptive optimization from data fusion to classification results. This method first collaboratively processes the multidimensional features of optical imagery (Sentinel-2 / Landsat-8) and radar imagery (Sentinel-1). A dynamic PCA fusion algorithm (with an adjustable threshold range of 75%-90%) combined with local feature weighting overcomes the information loss bottleneck caused by traditional static fusion. The method innovatively introduces normalized dot product similarity (S) as a feedback metric. When S < 0.7, a three-stage optimization process is triggered: the PCA threshold is automatically lowered to retain more principal components, the learning rate decay period is shortened to 5 epochs, and the MSAVI vegetation index weight is increased by 1.2-1.5 times, improving feature representation by over 30%. Furthermore, through spatial consistency optimization, the classification results are scanned with a 5×5 pixel window. A progressive feature correction formula is used to iteratively eliminate fragmentation errors, significantly improving the accuracy of wetland boundary extraction.

[0121] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A wetland feature extraction system based on multi-source remote sensing images and deep learning, characterized by: include: Data acquisition module, used to collect optical remote sensing images and radar remote sensing images and perform correction and multi-view processing to form a target data set; a feature fusion module, configured to extract optical features and radar features from the target data set, fuse the optical features and radar features using principal component analysis and preset local feature weights, and determine a fusion feature matrix based on a PCA threshold; A CNN model training module is used to build a recognition model based on a gradient descent algorithm and a convolutional neural network, process the fused feature matrix according to a learning rate decay cycle and the recognition model, and determine a wetland extraction result; A feedback control module is used to determine the normalized dot product similarity between the fused feature matrix and the wetland extraction result, and dynamically adjust the PCA threshold, learning rate decay period or local feature weight based on the relationship between the normalized dot product similarity and a preset threshold interval.

2. The wetland feature extraction system based on multi-source remote sensing images and deep learning according to claim 1 is characterized in that: The optical remote sensing image includes blue, green, red and near-infrared bands, and the radar remote sensing image includes VV and VH polarization channels.

3. The wetland feature extraction system based on multi-source remote sensing images and deep learning according to claim 1 is characterized in that: The data acquisition module includes: The first acquisition unit is used to determine whether the acquisition time interval of the optical remote sensing image and the radar remote sensing image meets the preset overlapping interval. If so, the preset ecological period table is searched according to the target wetland type. Otherwise, the hydrological characteristics of the missing period are supplemented by linear interpolation. The first processing unit is used to perform radiation correction, geometric correction and atmospheric correction on the optical remote sensing image in sequence; The second processing unit is used to perform polarization matrix decomposition and multi-view processing on the radar remote sensing image.

4. The wetland feature extraction system based on multi-source remote sensing images and deep learning according to claim 1 is characterized in that: The feature fusion module includes: Optical feature extraction unit, used to extract spectral features, texture features and vegetation index from optical remote sensing images. Vegetation index includes MSAVI and NDWI to obtain optical features. A radar feature extraction unit is used to extract backscatter coefficients and polarization decomposition features from radar images to obtain radar features; The fusion execution unit is used to fuse the optical features and radar features using principal component analysis and preset local feature weights, dynamically adjust the PCA contribution rate threshold, and obtain the fusion feature matrix F, whose dimensions meet the following requirements: Where λ is the eigenvalue, k is the number of retained principal components, n is the total dimension of the original feature, and λ i is the eigenvalue corresponding to the i-th principal component, Threshold PCA It is the preset cumulative contribution rate threshold.

5. The wetland feature extraction system based on multi-source remote sensing images and deep learning according to claim 1 is characterized in that: The CNN model training module includes: A model building unit, used to build a recognition model based on a convolutional neural network; A gradient descent algorithm unit, used to optimize the recognition model through a gradient descent algorithm; A learning rate adjustment unit, configured to adjust the learning rate according to a preset learning rate decay period; The feature processing unit is used to input the fused feature matrix into the recognition model for processing to determine the wetland extraction result matrix R.

6. The wetland feature extraction system based on multi-source remote sensing images and deep learning according to claim 1 is characterized in that: The feedback control module includes: The similarity calculation unit is used to calculate the normalized dot product similarity S between the fusion feature matrix F and the wetland extraction result matrix R in real time. The expression is: Dynamic optimization unit: When S < the first threshold, perform at least one of the following adjustments: (1) Lower the PCA contribution rate threshold to the second threshold and increase the number of principal components; (2) Decrease the learning rate cycle to a set number of times; (3) Increase the weight of the MSAVI index to a set multiple.

7. The wetland feature extraction system based on multi-source remote sensing images and deep learning according to claim 1 is characterized in that: The feedback control module also includes a consistency optimization unit for Scan the wetland classification result matrix R with the preset pixels as the window, and calculate the Euclidean distance D between the features in the window and the standard template; When D>the third threshold, the features in the window are corrected according to the following formula: F corrected =F original +0.3×(F template -F original ) Among them, F original is the original eigenvector to be corrected, F template is the standard wetland feature template, F corrected is the optimized feature vector; The correction is repeated until the D value decrease rate is lower than the fourth threshold.

8. A wetland feature extraction method based on multi-source remote sensing images and deep learning, characterized by: Based on the wetland feature extraction system according to any one of claims 1 to 7, the method comprises the following steps: S1. Obtain optical and radar remote sensing images of the wetland area, perform radiometric, geometric, and atmospheric corrections on the optical remote sensing images, and perform multi-view processing and polarization calibration on the radar remote sensing images. S2. Extract spectral features, texture features, and vegetation indices from optical remote sensing images, and extract backscatter coefficients and polarization decomposition features from radar remote sensing images. Use principal component analysis and preset local feature weights to fuse the optical and radar features to obtain a fused feature matrix. S3. Build a recognition model with convolutional layers and train the model using stochastic gradient descent and k-fold cross validation. S4. Input the image to be classified into the trained model for wetland extraction, calculate the normalized dot product similarity S between the fusion feature matrix F and the extraction result matrix R, and dynamically adjust the PCA threshold, learning rate decay cycle, or local feature weight based on the relationship between the normalized dot product similarity and the preset threshold range; S5. Optimize the local consistency of the classification results: calculate the Euclidean distance D between the feature and the standard template using a preset pixel window. When D is greater than the third threshold, correct the features in the window according to the preset formula until the decrease rate of the D value is lower than the fourth threshold.

9. The wetland feature extraction method based on multi-source remote sensing images and deep learning according to claim 8, characterized in that: In step S1, the data acquisition module further includes judging the acquisition time interval of the optical remote sensing image and the radar remote sensing image. If the preset overlapping interval is not satisfied, the linear interpolation method is used to fill in the hydrological characteristics of the missing period.

Images