Construction method of wetland classification model, wetland classification method and system

Abstract

The invention relates to the technical field of ecological environment protection, in particular to a construction method of a wetland classification model and a wetland classification method and system.The scheme includes the steps that historical multi-source data of different types of wetlands are obtained; according to the multispectral image, the SAR image, the DEM data and the SATITE semantic image, spectral features, radar features, topographic features and semantic features are calculated respectively; inputting the spectral features, the radar features and the topographic features into a first machine learning model for training, and outputting a first classification probability vector of each wetland type when the training is completed; inputting the semantic features into a second machine learning model for training, and when the training is completed, outputting a second classification probability vector of each wetland type; and splicing the first classification probability vector and the second classification probability vector to obtain a third classification probability vector, taking the third classification probability vector as training data of a third machine learning model, and outputting a wetland classification model when training is completed.

Description

Technical Field

[0001] This invention relates to the field of ecological and environmental protection technology, specifically to a method for constructing a wetland classification model, a wetland classification method, and a system. Background Technology

[0002] Wetlands are an important component of the Earth's ecosystem, possessing ecological functions such as hydrological regulation, material cycling, biological habitat, and climate control. Long-term, dynamic monitoring of wetlands is of great significance for ecological environmental protection and national spatial planning. Traditional ground-based monitoring methods are costly, time-consuming, and limited by geographical conditions, making it difficult to achieve efficient identification and classification of large-scale wetland distribution.

[0003] With the development of Earth observation technology, remote sensing has become an important technical means for wetland classification and dynamic monitoring. Among them, Sentinel-2 multispectral data has a spatial resolution of 10–20 meters and 13 spectral bands (including red-edge and shortwave infrared), which can provide rich vegetation spectral information and is suitable for identifying different types of wetland vegetation. Sentinel-1 radar data obtains surface scattering characteristics through synthetic aperture radar (SAR), has all-weather and all-time observation capabilities, and can effectively capture the hydrological structure information of wetlands.

[0004] Currently, Sentinel-1 and Sentinel-2 data are often used together for wetland classification. For example, a feature set is constructed using spectral features, vegetation indices (NDVI, EVI, etc.) and radar backscattering coefficients (VV, VH), and then classified using machine learning algorithms (such as random forests and support vector machines).

[0005] In existing technologies, although Sentinel-1 and Sentinel-2 series remote sensing data are widely used for wetland monitoring and classification, their classification accuracy is still insufficient. Due to the complexity of wetland types, similar spectral characteristics of land cover, and significant seasonal variations, relying solely on Sentinel-1 / 2 data often fails to accurately distinguish different wetland types, resulting in low overall accuracy and stability. Summary of the Invention

[0006] This invention provides a method for constructing a wetland classification model, a wetland classification method, and a system, which can effectively solve the problem of insufficient accuracy of wetland classification results obtained by existing technologies.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for constructing a wetland classification model, the method comprising:

[0009] Historical multi-source data of different types of wetlands are acquired; wherein, the historical multi-source data includes multispectral images, SAR images, DEM data, and satellite embedding semantic images;

[0010] Based on the multispectral images, SAR images, DEM data, and Satellite Embedding semantic images, spectral features, radar features, terrain features, and semantic features are extracted respectively.

[0011] The spectral features, radar features, and terrain features are input into the first machine learning model for training. Upon completion of training, the first classification probability vectors for each wetland type are output.

[0012] The semantic features are input into the second machine learning model for training. Upon completion of training, the second classification probability vectors for various wetland types are output.

[0013] The first classification probability vector is concatenated with the second classification probability vector to obtain the third classification probability vector. The third classification probability vector is then used as the training data for the third machine learning model. Upon completion of training, the wetland classification model is output.

[0014] In one implementation, before extracting spectral features, radar features, terrain features, and semantic features from the multispectral image, SAR image, DEM data, and SatelliteEmbedding semantic image, respectively, the method further includes:

[0015] Atmospheric correction is performed on multispectral images, radiometric calibration and topographic correction are performed on SAR images, and projection unification and hole filling are performed on DEM data to align multispectral images, SAR images, DEM data and Satellite Embedding semantic images at the pixel level.

[0016] In one implementation, before inputting the spectral features, radar features, terrain features, and semantic features into a machine learning model for training, the method further includes performing minimum and maximum normalization on the spectral features, radar features, terrain features, and semantic features.

[0017] In one implementation scheme, the first machine learning model, the second machine learning model, and the third machine learning model can all adopt any one of the four types: random forest model, support vector machine model, decision tree model, and multi-class logistic regression model.

[0018] In one implementation, the output layer of the third machine learning model employs the softmax function.

[0019] A second aspect of the present invention provides a wetland classification method, the method comprising:

[0020] Acquire multi-source data of wetlands to be classified;

[0021] Multi-source data of wetlands to be classified are input into a pre-constructed wetland classification model, and the wetland classification results are output; wherein, the wetland classification model is constructed using a method for constructing a wetland classification model as provided in the first aspect of the present invention.

[0022] In one implementation, the method further includes:

[0023] Spatial processing is performed on the wetland classification results to correct isolated pixels of each category; wherein, spatial processing includes majority filtering and minimum patch filtering.

[0024] The wetland classification results after spatial processing are corrected by combining preset auxiliary rules to generate classification raster and vector pattern;

[0025] The final wetland classification results and confidence maps are generated based on the classification raster and vector plots.

[0026] A third aspect of the present invention provides a system for constructing a wetland classification model, the system comprising:

[0027] The historical multi-source data acquisition unit is used to acquire historical multi-source data of different types of wetlands; wherein, the historical multi-source data includes multispectral images, SAR images, DEM data, and Satellite Embedding semantic images;

[0028] The feature calculation unit is used to extract spectral features, radar features, terrain features and semantic features from the multispectral image, SAR image, DEM data and SatelliteEmbedding semantic image, respectively.

[0029] The first training unit is used to input the spectral features, radar features and terrain features into the first machine learning model for training, and outputs the first classification probability vector of each wetland type when the training is completed.

[0030] The second training unit is used to input the semantic features into the second machine learning model for training, and outputs the second classification probability vectors of various wetland types when the training is completed.

[0031] The third training unit is used to concatenate the first classification probability vector with the second classification probability vector to obtain the third classification probability vector, and use the third classification probability vector as the training data of the third machine learning model. When the training is completed, the wetland classification model is output.

[0032] A fourth aspect of the present invention provides a wetland classification system, the system comprising:

[0033] The data acquisition unit is used to acquire multi-source data of the wetlands to be classified.

[0034] The wetland classification unit is used to input multi-source data of the wetland to be classified into a pre-constructed wetland classification model and output the wetland classification result; wherein, the wetland classification model is constructed using a method for constructing a wetland classification model as provided in the first aspect of the present invention.

[0035] A fifth aspect of the present invention provides an electronic device, including a memory and a processor;

[0036] A memory for storing computer programs, the computer programs including program instructions;

[0037] A processor is configured to execute the program instructions to cause the electronic device to perform the steps of a method for constructing a wetland classification model as provided in the first aspect of the invention and the steps of a wetland classification method as provided in the second aspect of the invention.

[0038] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0039] This invention combines Satellite Embedding feature vectors with Sentinel-1 / 2 multispectral and radar features into a unified feature framework, enabling deep fusion training of optical, radar, and semantic features. Satellite Embedding feature vectors provide efficient land cover feature vectors through deep learning compression, which can effectively improve the accuracy of wetland classification models. Furthermore, the fusion with Sentinel-1 / 2 data further enhances the stability and accuracy of the wetland classification model's output classification results. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0041] Figure 1 A flowchart illustrating a method for constructing a wetland classification model according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart illustrating a wetland classification method provided in an embodiment of the present invention.

[0043] Figure 3 This is a comparison chart of the humidity classification accuracy between existing technologies and the present invention;

[0044] Figure 4 A schematic diagram of a system for constructing a wetland classification model provided in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of a wetland classification system provided in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0047] It should be noted that the terms "comprising" or "may include" used in the various embodiments of this application indicate the presence of the claimed function, operation, or element, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in the various embodiments of this application, the terms "comprising," "having," and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or adding one or more combinations of the foregoing.

[0048] It should be understood that terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for constructing a wetland classification model, the method comprising:

[0051] S101, acquire historical multi-source data of different types of wetlands; wherein, the historical multi-source data includes multispectral images, SAR images, DEM data and Satellite Embedding semantic images.

[0052] In this embodiment, historical multi-source data of different types of wetlands are acquired based on the study area boundary and the target year. For historical multi-source data, Sentinel-1 and Sentinel-2 series remote sensing data are selected as the data source. As one of the technical means well known to those skilled in the art, the Sen2cor atmospheric correction model of SNAP software can also be used to perform atmospheric correction on the downloaded L1C level remote sensing images to obtain L2A level images. The data standard for wetland classification is determined, specifically including the following steps: Based on Sentinel-2 remote sensing data, the geometric features of ground features and actual images are described in combination with high-resolution Google Earth images; wherein, the primary classification is natural wetlands, artificial wetlands and non-wetlands; in the secondary classification, natural wetlands include water bodies, woody marshes, herbaceous marshes, floodplains, saline-alkali land, mangroves, salt marshes and tidal flats; artificial wetlands include paddy fields and aquaculture ponds / pits; non-wetlands include: woodlands, grasslands, drylands, built-up land, bare land and snow.

[0053] In this embodiment, the multispectral imagery is obtained from Sentinel-2 series remote sensing data, including visible, red-edge, near-infrared, and shortwave infrared bands, with a spatial resolution of 10 meters. The SAR imagery is obtained from Sentinel-1 series remote sensing data, including VV and VH polarization channels, with a spatial resolution of 10 meters. The SatelliteEmbedding semantic imagery is generated by a pre-trained satellite vision model through inference on the original satellite imagery, outputting a fixed-dimensional semantic vector imagery, which is then rasterized and projected to align spatially with the Sentinel imagery. The above data is cropped according to the study area boundary and metadata is saved for subsequent processing.

[0054] After acquiring historical multi-source data of different types of wetlands, data preprocessing is required. The data preprocessing process is as follows: atmospheric correction is performed on multispectral images, radiometric calibration and topographic correction are performed on SAR images, and projection unification and hole filling are performed on DEM data to align multispectral images, SAR images, DEM data and SatelliteEmbedding semantic images on the pixel level.

[0055] Specifically, the Fill NoData tool in the open-source software QGIS is used to perform hole filling processing on the DEM data. Based on the SAR image, the Align rasters tool is used to reduce the resolution of the DEM data to 10 meters while maintaining spatial projection consistency and pixel alignment with the SAR image.

[0056] Atmospheric correction was performed on the multispectral image using the Sen2cor atmospheric correction model of the open-source software SNAP, converting the pixel values ​​into surface reflectance images.

[0057] SNAP was used to radiometrically calibrate the SAR image, converting pixel values ​​into backscattering coefficients. The terrain correction process employed the Range-Doppler terrain correction algorithm in SNAP software, inputting DEM data to achieve SAR image terrain correction. Finally, in QGIS, the Align rasters tool was used to align the multispectral image, SAR image, DEM data, and Satellite Embedding semantic image at the pixel level, using the Satellite Embedding semantic image as a reference.

[0058] In summary, multispectral imagery undergoes atmospheric correction to generate cloud and shadow masks; SAR imagery undergoes radiometric calibration and topographic correction with speckle suppression; and DEM data undergoes projection unification and hole filling. All data are reprojected to a unified spatial reference frame and resampled to a 10-meter spatial resolution, ensuring strict pixel-level alignment of optical, radar, DEM, and satellite embeddings. Missing time-series pixels are interpolated based on temporally nearest observations to form a continuous and stable annual time-series dataset.

[0059] It should be understood that the data preprocessing process for historical multi-source data described above is existing technology, and generally speaking, the historical multi-source data obtained is open source data, which is itself preprocessed data.

[0060] S102, based on the multispectral image, SAR image, DEM data and Satellite Embedding semantic image, extract spectral features, radar features, terrain features and semantic features respectively.

[0061] In this embodiment, representative spectral indices are calculated based on the preprocessed multispectral imagery, and time-series statistical features are generated. These indices include, but are not limited to, vegetation indices (such as NDVI), water body indices (such as MNDWI), humidity / heat indices (such as NBR), and tassel cap transformation features (such as TCW), etc., with a total of 40 to 60 remote sensing index features preferred to fully characterize wetland land cover information. The vegetation index NDVI is defined as follows: ,in, and These are the surface reflectance in the near-infrared and red light bands, respectively. NDVI effectively reflects vegetation cover and photosynthetic intensity, and is the most commonly used indicator for distinguishing wetland vegetation and identifying land-water boundaries.

[0062] Specifically, vegetation indices also include normalized vegetation index, ratio vegetation index, difference vegetation index, soil-adjusted vegetation index, and soil-adjusted vegetation index.

[0063] The Water Enhancement Index (MNDWI) is defined as follows: ,in, and These represent the reflectance of the green light band and the shortwave infrared channel, respectively. MNDWI can significantly enhance water body signals and suppress interference from buildings and vegetation, and is often used for wetland water body extraction and seasonal water surface change analysis. Water body indices include the Normalized Difference Water Index (NDDI) and the Modified Normalized Difference Water Index (MNDWI).

[0064] The disturbance index NBR is defined as: ,in, and These are the near-infrared and short-wave infrared two-channel reflectances, respectively. The NBR is sensitive to vegetation moisture content and wetland dry-wet alternation, and has high discrimination in wetland degradation monitoring, grassland stress identification, and fire area assessment.

[0065] The humidity component index TCW is defined as follows:

[0066] ,in, , , , , , These represent the reflectance of blue light, green light, red light, near-infrared light, short-wave infrared channel 1, and short-wave infrared channel 2, respectively. to These are the tasseled cap transformation coefficients for blue, green, red, near-infrared, and short-wave infrared channels 1 and 2, respectively. TCW is highly sensitive to changes in surface humidity and can be used to characterize wetland moisture conditions and seasonal humidity differences.

[0067] Statistical values ​​for the above indices are calculated over an annual time window, including mean, median, standard deviation, minimum, maximum, and multiple percentile values. These values ​​are then stitched together by pixel to form spectral features and timestamp information is recorded.

[0068] The VV and VH backscattering coefficients are extracted from the preprocessed SAR images. Statistical values ​​of the VV and VH backscattering coefficients are calculated over an annual time window, including mean, median, standard deviation, minimum, maximum, and multiple percentile values. The data are then stitched together by pixel to form radar features and timestamp information is recorded.

[0069] Topographic features such as elevation, slope, and aspect can be extracted from DEM data. For example, QGIS software can be used to extract slope from DEM data using its Slope tool and aspect using the Aspect tool. DEM data itself is elevation data and does not require processing.

[0070] Satellite Embedding semantic imagery is a multi-band image with 64 bands. Each band serves as a semantic feature, resulting in 64 semantic features across the 64 bands.

[0071] It is important to understand that the radar features, terrain features, and spectral features mentioned above are organized in parallel to form a multi-source feature system that covers spectral, structural, and terrain information.

[0072] After feature extraction, to unify the dimensions of features from different sources, minimum and maximum normalization are performed on all features. The normalization formula is as follows: Where x is the original feature value, and Let x' be the minimum and maximum values ​​of the feature in the training sample set, and x′ be the normalized feature value. After normalization, the spectral features, radar features, terrain features, and Satellite Embedding semantic vector are concatenated in a preset order to form the final multi-source fusion feature vector for use in training the machine learning model.

[0073] S103, the spectral features, radar features and terrain features are input into the first machine learning model for training, and when the training is completed, the first classification probability vector of each wetland type is output.

[0074] Specifically, using spectral, index, radar, and terrain features from Sentinel1 / 2 in the multi-source fusion feature set as the main input, a random forest algorithm is employed as the base classification model for the first machine learning model. The random forest consists of several decision trees, with the number of trees ranging from 200 to 500. When splitting a node, the number of candidate features accounts for 30% to 50% of all features, and the minimum number of splits is 2 to 5. After training, a probability vector for each pixel regarding its category is output, and the model's feature importance metric is saved.

[0075] Based on Sentinel-2 remote sensing imagery, combined with high-resolution Google Earth imagery and the GWL_FCS wetland dataset, sample points were collected and manually interpreted and labeled. The primary categories include natural wetlands, artificial wetlands, and non-wetlands. The secondary categories include water bodies, woody marshes, herbaceous marshes, floodplains, saline-alkali land, mangroves, salt marshes, and tidal flats; artificial wetlands include paddy fields and aquaculture ponds / pits; and non-wetlands include woodlands, grasslands, drylands, built-up land, bare land, and snow and ice. Each category is coded from 1 to N for model training, where N represents the total number of codes. Sample points are randomly divided into training and test sets at a ratio of 3:1.

[0076] The random forest algorithm uses the `RandomForestClassifier` function from the Python scikit-learn library, with 100 trees and other parameters using the default parameters from the `RandomForestClassifier` function. It should be noted that the model training process provided in this embodiment is based on an open-source training method, which is a well-known training technique familiar to those skilled in the art and does not require further description here.

[0077] After training, the first machine learning model outputs a first classification probability vector P for each training sample. 1 =(p 11 ,p 12 ,…,p 1i ); where p 1i This represents the predicted probability that the training sample belongs to the i-th wetland type, where i = 1, 2, ..., N, and N is the total number of wetland classification categories.

[0078] It should be noted that adjusting the hyperparameters of the first machine learning model and verifying its performance during training are conventional techniques in this field, and will not be elaborated upon in this embodiment. For example, cross-validation can be used during training to adjust hyperparameters and evaluate the model's generalization performance.

[0079] S104, the semantic features are input into the second machine learning model for training, and when the training is completed, the second classification probability vectors of various wetland types are output.

[0080] Specifically, an independent random forest classification model, or second machine learning model, is constructed using the semantic features of the Satellite Embedding as input. The model parameters of this second machine learning model are consistent with those set in S103. This second machine learning model outputs probability vectors for each class of training samples, providing independent discriminative information from the semantic space for subsequent fusion. After model training is complete, the second machine learning model outputs a second classification probability vector P for each training sample. 2 =(p 21 ,p 22 ,…,p 2i ); where p 2i This represents the predicted probability that the training sample belongs to the i-th wetland type, i=1,2,…,N.

[0081] S105, the first classification probability vector and the second classification probability vector are concatenated horizontally to obtain the third classification probability vector. The third classification probability vector is used as the training data for the third machine learning model. When the training is completed, the wetland classification model is output.

[0082] Specifically, the first classification probability vector P output by the model trained in steps S103 and S104 above is... 1 With the second classification probability vector P 2 Perform horizontal concatenation to form the third classification probability vector z = (P 1 P 2 The third classification probability vector z is input into the third machine learning model for training. Upon completion of training, the final classification probability is output. When the third machine learning model uses a multi-class logistic regression model as the meta-learner, its output probability is represented by the Softmax function: Where z is the probability vector for the third classification. Let be the parameter vector of the i-th class. Let be the final classification probability for wetland type, and j represent the category code for classification, used to iterate through all categories from 1 to k, where k represents the total number of category codes. The meta-learner can also employ ensemble algorithms such as random forests, using cross-validation or K-fold stacking strategies during training to prevent information leakage.

[0083] It can be understood that, in addition to using the random forest algorithm, the first, second, and third machine learning models can also use any one of the three models: support vector machine, decision tree, and multi-class logistic regression.

[0084] In summary, this invention combines Satellite Embedding feature vectors with Sentinel-1 / 2 multispectral and radar features into a unified feature framework, enabling deep fusion training of optical, radar, and semantic features. The Satellite Embedding feature vectors provide efficient land cover feature vectors through deep learning compression, effectively improving the accuracy of wetland classification models. Furthermore, the fusion with Sentinel-1 / 2 data further enhances the stability and accuracy of the wetland classification model's output classification results.

[0085] like Figure 2 As shown in the figure, this embodiment of the invention also provides a wetland classification method, the method comprising:

[0086] S201, Obtain multi-source data of the wetland to be classified.

[0087] Specifically, the acquisition of multi-source data can be referred to the data described in step S101 of the above embodiment, and will not be described in detail in this embodiment.

[0088] S202, input the multi-source data of the wetland to be classified into a pre-constructed wetland classification model, and output the wetland classification result; wherein, the wetland classification model is based on, for example, Figure 1The wetland classification model shown is constructed using a method that...

[0089] Specifically, because the wetland classification model constructed in this invention has the ability to improve the accuracy of wetland classification models and ensure the stability and precision of classification results, the wetland classification results can be output by inputting multi-source data of the wetland to be classified into the pre-constructed wetland classification model. For example... Figure 3 As shown, the wetland classification model proposed in this invention, which integrates Satellite Embedding and Sentinel 1 / 2 multi-source features, outperforms the comparative scheme in all wetland regions. Specifically, in the Guangdong-Hong Kong-Macao Greater Bay Area and Yancheng Wetland Region, the Kappa coefficients improved from 0.893 and 0.892 to 0.924 and 0.922, respectively, representing improvements of 3.45% and 3.38%; in the Beibu Gulf Wetland Region, the Kappa coefficient improved from 0.845 to 0.896, an improvement of 5.99%; and in the Ruoergai Wetland Region, the Kappa coefficient improved from 0.929 to 0.946, an accuracy improvement of 1.83%. Overall, the average Kappa coefficient of the proposed scheme across the nine regions is 0.927, an improvement of 1.72% compared to the comparative scheme. Experimental results demonstrate that the wetland classification model proposed in this invention outperforms traditional methods in overall accuracy, thus enabling high-precision wetland classification.

[0090] In some embodiments, the wetland classification method provided by the present invention further includes: spatial processing of the wetland classification results to correct isolated pixels of the category; wherein, the spatial processing includes majority filtering and minimum patch filtering; combining preset auxiliary rules to correct the spatially processed wetland classification results to generate classification raster and vector patch; and generating the final wetland classification result and confidence map based on the classification raster and vector patch.

[0091] Specifically, for most filtering processes, a sliding window can be used to traverse the classification grid, with a window size of 3×3 or 5×5; the category that appears most frequently within the window is used as the updated category for the central pixel. When there are ties, the category with the higher corresponding classification probability can be selected first, or the original category can be kept unchanged.

[0092] Minimum patch filtering can identify connected patches of the same category based on connected component analysis (e.g., 8-neighbor connectivity). When the area of ​​a connected patch is less than a preset threshold (e.g., less than a certain number of pixels or a certain area), the patch is merged into the target category adjacent to its boundary. The merged target category can be determined by the rule of maximizing the adjacent boundary length, maximizing the average classification probability of adjacent categories, or a weighted average of both.

[0093] Auxiliary rules include, but are not limited to: threshold constraint rules based on water body indices (such as NDWI / MNDWI) and vegetation indices (such as NDVI / EVI), water-land boundary consistency rules, accessibility constraint rules based on topographic factors (elevation and slope), and spatial proximity constraint rules based on distance from rivers / lakes / coastlines. For example, when a pixel is classified as "water body" but its water body index is below the threshold and its vegetation index is above the threshold, it can be reclassified as the non-water body category with the highest probability in the neighborhood. When a pixel is classified as coastal wetland such as "tidal flat / salt marsh / mangrove", but its elevation or slope exceeds the preset range or is far from the coastline, it can be reclassified as the inland wetland or non-wetland category with the highest probability in the neighborhood.

[0094] The auxiliary rules can also use prior masks or stable region information from multi-source input data to correct the classification results. For example, the non-wetland stable region (cultivated land, forest land, construction land, etc.) mask can be used to backtrack and correct obviously misclassified wetland categories.

[0095] The classification raster is a pixel-level category-encoded raster. Vector polygons can be obtained by converting the classification raster to a vector (polygonize) and can be fused (dissolved), have their boundaries smoothed / simplified, and be filtered by minimum area constraints according to the category field. The confidence map can be calculated from the probability vectors of each category output by the wetland classification model. The confidence score can be taken as the maximum category probability, or the difference between the first and second largest probabilities, as an uncertainty measure, and output in spatial correspondence with the classification raster pixel by pixel.

[0096] like Figure 4 As shown in the figure, this embodiment of the invention also provides a system for constructing a wetland classification model, the system comprising:

[0097] The historical multi-source data acquisition unit 410 is used to acquire historical multi-source data of different types of wetlands; wherein, the historical multi-source data includes multispectral images, SAR images, DEM data, and Satellite Embedding semantic images;

[0098] The feature calculation unit 420 is used to extract spectral features, radar features, terrain features and semantic features based on the multispectral image, SAR image, DEM data and SatelliteEmbedding semantic image, respectively.

[0099] The first training unit 430 is used to input the spectral features, radar features and terrain features into the first machine learning model for training, and output the first classification probability vector of each wetland type when the training is completed.

[0100] The second training unit 440 is used to input the semantic features into the second machine learning model for training, and output the second classification probability vector of various wetland types when the training is completed.

[0101] The third training unit 450 is used to concatenate the first classification probability vector with the second classification probability vector to obtain a third classification probability vector, and use the third classification probability vector as training data for the third machine learning model. When training is completed, the wetland classification model is output.

[0102] Specifically, the construction system provided by this invention combines Satellite Embedding feature vectors with Sentinel-1 / 2 multispectral and radar features into a unified feature framework, enabling deep fusion training of optical, radar, and semantic features. Satellite Embedding feature vectors provide efficient land cover feature vectors through deep learning compression, which can effectively improve the accuracy of wetland classification models. Furthermore, the fusion with Sentinel-1 / 2 data further enhances the stability and accuracy of the classification results output by the wetland classification model.

[0103] like Figure 5 As shown, this embodiment of the invention also provides a wetland classification system, the system comprising:

[0104] Data acquisition unit 510 is used to acquire multi-source data of the wetland to be classified.

[0105] The wetland classification unit 520 is used to input multi-source data of the wetland to be classified into a pre-constructed wetland classification model and output the wetland classification result; wherein, the wetland classification model is constructed using a method for constructing a wetland classification model as provided in the first aspect of the present invention.

[0106] Specifically, the wetland classification model constructed in this invention has the ability to improve the accuracy of wetland classification models and ensure the stability and precision of classification results. Therefore, by inputting multi-source data of the wetland to be classified into the pre-constructed wetland classification model, wetland classification results can be output.

[0107] This invention also provides an electronic device. The electronic device includes a processor, a memory, a communication interface, and at least one communication bus for connecting the processor, the memory, and the communication interface. The memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (PROM), or portable read-only memory (CD-ROM), and is used for related instructions and data.

[0108] The communication interface is used to receive and send data. The processor can be one or more CPUs; if the processor is a single CPU, it can be a single-core CPU or a multi-core CPU. The processor in the electronic device is used to read one or more programs stored in the memory and perform the following operations: acquire historical multi-source data on different types of wetlands; wherein the historical multi-source data includes multispectral images, SAR images, DEM data, and Satellite Embedding semantic images; extract spectral features, radar features, topographic features, and semantic features based on the multispectral images, SAR images, DEM data, and Satellite Embedding semantic images; input the spectral features, radar features, and topographic features into a first machine learning model for training, and output a first classification probability vector for each wetland type upon completion of training; input the semantic features into a second machine learning model for training, and output a second classification probability vector for each wetland type upon completion of training; concatenate the first and second classification probability vectors to obtain a third classification probability vector, and use the third classification probability vector as training data for a third machine learning model, outputting a wetland classification model upon completion of training.

[0109] It should be noted that the specific implementation of each operation can be described above. Figure 1 and Figure 2 The corresponding description of the method embodiments shown indicates that the electronic device can be used to execute a method for constructing a wetland classification model and a wetland classification method according to the above method embodiments of this application, which will not be described in detail here.

[0110] This invention also provides a computer-readable storage medium, which is a memory device in a computer device for storing programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for constructing a wetland classification model and a wetland classification method in the above embodiments. Those skilled in the art should understand that the embodiments of this invention can be provided as methods, systems, or computer program products. Therefore, this invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD ROM, optical storage, etc.) containing computer-usable program code.

[0111] This invention also provides a computer program product containing program instructions. The computer program product may be software or program products containing program instructions, capable of running on a computing device or stored on any usable medium. When the computer program product is run on at least one electronic device, it causes the at least one electronic device to execute a method for constructing a wetland classification model and a wetland classification method.

[0112] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a wetland classification model, characterized in that the method... The method comprises the following steps: acquiring historical multi-source data of different types of wetlands; wherein the historical multi-source data comprises multispectral images, SAR images, DEM data, and Satellite Embedding semantic images; calculating spectral features, radar features, terrain features, and semantic features according to the multispectral images, SAR images, DEM data, and Satellite Embedding semantic images; inputting the spectral features, radar features, and terrain features into a first machine learning model for training, and outputting first classification probability vectors of various types of wetlands when the training is completed; inputting the semantic features into a second machine learning model for training, and outputting second classification probability vectors of various types of wetlands when the training is completed; splicing the first classification probability vectors and the second classification probability vectors to obtain third classification probability vectors, and taking the third classification probability vectors as training data of a third machine learning model, and outputting a wetland classification model when the training is completed.

2. The method of claim 1, wherein, Before the spectral features, radar features, terrain features, and semantic features are calculated according to the multispectral images, SAR images, DEM data, and Satellite Embedding semantic images, the method further comprises: performing atmospheric correction on the multispectral images, performing radiation calibration and terrain correction on the SAR images, and performing projection unification and hole filling processing on the DEM data, so that the multispectral images, SAR images, DEM data, and Satellite Embedding semantic images are aligned on the pixel.

3. The method of claim 1, wherein, Before the spectral features, radar features, terrain features, and semantic features are inputted into the machine learning model for training, the method further comprises performing minimum and maximum normalization processing on the spectral features, radar features, terrain features, and semantic features.

4. The method of claim 1, wherein, The first machine learning model, the second machine learning model, and the third machine learning model can adopt any one of the following four types: random forest model, support vector machine model, decision tree model, and multi-classification logistic regression model.

5. The method of claim 1, wherein, The output layer of the third machine learning model adopts a softmax function.

6. A method of wetland classification, the method comprising: The method comprises the following steps: acquiring multi-source data of a wetland to be classified; inputting the multi-source data of the wetland to be classified into a pre-constructed wetland classification model, and outputting a wetland classification result; wherein the wetland classification model is constructed by using the construction method of the wetland classification model according to any one of claims 1 to 5.

7. The method of claim 6, wherein, The method further comprises: performing spatial processing on the wetland classification result to correct isolated class pixels; wherein the spatial processing comprises majority filtering processing and minimum patch filtering processing; combining a preset auxiliary rule to modify the wetland classification result after the spatial processing, to generate a classification raster and a vector patch; generating a final wetland classification result and a confidence map according to the classification raster and the vector patch.

8. A system for constructing a wetland classification model, the system comprising: The system comprises: a historical multi-source data acquisition unit configured to acquire historical multi-source data of different types of wetlands; wherein the historical multi-source data comprises multispectral images, SAR images, DEM data, and Satellite Embedding semantic images; a feature calculation unit configured to calculate spectral features, radar features, terrain features, and semantic features respectively according to the multi-spectral image, the SAR image, the DEM data, and the SatelliteEmbedding semantic image; a first training unit configured to input the spectral features, the radar features, and the terrain features into a first machine learning model for training, and output a first classification probability vector of each type of wetland when the training is completed; a second training unit configured to input the semantic features into a second machine learning model for training, and output a second classification probability vector of each type of wetland when the training is completed; a third training unit configured to splice the first classification probability vector and the second classification probability vector to obtain a third classification probability vector, use the third classification probability vector as training data of a third machine learning model, and output a wetland classification model when the training is completed.

9. A wetland classification system, the system comprising: a data acquisition unit configured to acquire multi-source data of a wetland to be classified; a wetland classification unit configured to input the multi-source data of the wetland to be classified into a pre-constructed wetland classification model, and output a wetland classification result; wherein the wetland classification model is constructed by using the construction method of a wetland classification model according to any one of claims 1 to 5.

10. An electronic device, comprising: comprising a memory and a processor; the memory is configured to store a computer program, and the computer program comprises program instructions; the processor is configured to execute the program instructions to enable the electronic device to perform the steps of the construction method of a wetland classification model according to any one of claims 1 to 5 and the steps of the wetland classification method according to any one of claims 6 to 7.

Images