A large-scale wetland dynamic mapping method combining knowledge guidance and machine learning

Abstract

The application discloses a large-range wetland dynamic mapping method combining knowledge guidance and machine learning, comprising the following steps: step one, image acquisition and pretreatment; step two, time sequence training sample preparation and optimization; step three, wetland potential distribution estimation; step four, multi-dimensional space-time feature extraction; step five, machine learning classification; and step six, wetland space-time dynamic characteristic analysis. The application provides a large-range wetland dynamic mapping method combining knowledge guidance and machine learning. Compared with a traditional large-range wetland dynamic monitoring method, the method has the following technical advantages: 1. Training sample dependence is greatly reduced. The traditional machine learning needs to arrange a large number of uniformly distributed training samples in the whole research area, while the method needs to select samples only in the high-probability wetland area by means of wetland potential distribution range constraint sample collection, and the sample amount is reduced by more than 40% compared with similar research, and the classification accuracy can be maintained or even improved.

Description

Technical Field

[0001] This invention relates to the field of wetland remote sensing extraction and dynamic monitoring technology, specifically a method for large-scale dynamic mapping of wetlands that combines knowledge guidance and machine learning. Background Technology

[0002] Wetlands, as key ecosystems maintaining biodiversity, regulating climate, and controlling hydrological cycles, play an irreplaceable role in global sustainable development. Over the past century, affected by both global warming and human activities, global wetlands face widespread degradation and area loss. Although my country has continuously promoted wetland protection and restoration projects, high-resolution, long-term, and continuous wetland dynamic datasets at the national scale are still lacking. Traditional monitoring methods are insufficient to accurately identify the complex spatiotemporal heterogeneity and highly dynamic hydrological evolution processes of wetlands. Against this backdrop, accurate wetland mapping is a core prerequisite for supporting wetland ecological protection planning, policy formulation, and project effectiveness evaluation, while also providing crucial data support for the United Nations Sustainable Development Goals (SDGs). Currently, wetland dynamic monitoring and mapping technologies have formed a technical system supported by remote sensing data, centered on machine learning methods, and outputting multi-scale products. However, existing technical approaches still have multi-dimensional technical shortcomings in adapting to the needs of accurate wetland mapping.

[0003] First, in terms of remote sensing data, Landsat satellite imagery has become the mainstream for large-scale long-term monitoring due to its global coverage and rich historical archives. However, the mainstream spatial resolution of 30m is difficult to capture the fine spatial heterogeneity of wetlands. At the same time, due to the satellite revisit cycle, cloud pollution and insufficient early data coverage, the temporal continuity and data integrity are lacking.

[0004] Second, the fusion of optical and radar data from Sentinel-1 / 2 has the potential to achieve fine classification of wetlands with high spatiotemporal resolution because it can effectively avoid cloud pollution. However, the relevant technologies still need to be optimized and adapted to large-scale application scenarios.

[0005] Third, in terms of machine learning methods, traditional algorithms such as Random Forest are widely used for large-scale, high-resolution wetland dynamic monitoring due to their low computational cost and high interpretability. However, the training of such algorithms depends on massive samples. If the samples lack representativeness or have labeling errors, they are prone to problems such as high misclassification rate and insufficient computational efficiency.

[0006] Fourth, deep learning methods, represented by architectures such as Transformer, have shown potential in improving classification accuracy, but their models have poor interpretability and require huge amounts of computing power.

[0007] Therefore, developing a method that can quickly and accurately map and dynamically monitor large-scale wetlands is of great significance for the rational utilization and protection of wetland resources. Combining machine learning algorithms guided by prior knowledge to reduce dependence on massive samples provides a practical and feasible technical path to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide a method for large-scale dynamic mapping of wetlands that combines knowledge-guided learning and machine learning, in order to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a large-scale dynamic wetland mapping method combining knowledge guidance and machine learning, comprising: step one, image acquisition and preprocessing; step two, preparation and optimization of time-series training samples; step three, estimation of potential wetland distribution; step four, extraction of multi-dimensional spatiotemporal features; step five, machine learning classification; and step six, analysis of spatiotemporal dynamic features of wetlands.

[0010] In step one above, Sentinel-1 and Sentinel-2 images are acquired and preprocessed based on the Google Earth Engine platform, including denoising, cloud removal, radiometric calibration, terrain correction, and median image synthesis, in order to generate high-quality remote sensing images.

[0011] In step two above, multi-regional field survey data and multi-source geospatial products are integrated, and the sample size is expanded through visual interpretation, dynamic annotation and anomaly removal, and spatial consistency verification to construct a spatially continuous annual training dataset.

[0012] In step three above, spectral index features are extracted based on multi-temporal remote sensing image data, a temporal feature image set is constructed by combining topographic information, potential wetland range is generated through multi-condition threshold constraints, and morphological methods are used to remove noise.

[0013] In step four above, the spectral, texture, polarization and temporal dynamic features of the image are extracted, and key spectral indices are reconstructed to extract periodicity, extrema and variability features.

[0014] In step five above, a potential wetland spatial mask is used as prior knowledge. Through steps such as superpixel segmentation, feature and knowledge fusion, model training and classification, the model is guided to focus on high-probability wetland areas to generate a wetland type classification dataset.

[0015] In step six above, based on the generated annual wetland dataset, the trend of wetland category area change and spatial transfer characteristics are quantified, hotspots of wetland degradation or restoration are identified, and refined monitoring of wetland dynamics is achieved.

[0016] As a further technical solution of the present invention, in step one, Sentinel-1 and Sentinel-2 image products are acquired based on the Google Earth Engine platform, and a filtering algorithm is used to remove thermal noise in the images, remove invalid observations such as clouds and cloud shadows in the images, perform radiometric calibration on the images to eliminate sensor errors, and perform terrain correction to eliminate terrain effects. Through the time aggregation function of the Google Earth Engine platform, median image synthesis is completed to fill observation gaps and synthesize high-quality remote sensing images.

[0017] As a further technical solution of the present invention, in step two, sample data integration is carried out by combining multi-regional field survey data with multi-source geospatial products, expanding the sample scale through visual interpretation, merging and standardizing the initial samples according to the wetland classification system, preparing a time-series training sample library, and dynamically labeling and removing anomalies.

[0018] As a further technical solution of the present invention, in step two, the time series analysis module of the Collect Earth platform is used to generate high-resolution time series images of each sample location. The images are dynamically labeled in combination with the annual changes in vegetation cover, surface water range, and surface humidity. The platform's sampling and review tools are used to analyze the sample clustering patterns, remove inconsistent abnormal samples, perform spatial consistency checks, check the geographical distribution uniformity of the samples, construct a spatially continuous annual training dataset, keep the total number of samples stable, and dynamically update the category labels only according to changes in surface conditions, so as to provide reliable sample support for annual classification.

[0019] As a further technical solution of the present invention, in step three, based on multi-temporal remote sensing image data, spectral index features reflecting the characteristics of water bodies, vegetation and surface humidity are extracted, and a temporal feature image set is constructed in combination with topographic information. Key temporal statistical feature parameters are calculated, and statistical analysis is performed on the temporal feature image set to calculate key parameters reflecting the frequency of occurrence of water bodies (WOF), the frequency of occurrence of vegetation (VOF) and the degree of surface moisture (HMW), which are used to characterize the spatiotemporal characteristics of wetlands.

[0020] As a further technical solution of the present invention, in step three, the potential wetland range (WOF>0 or VOF>0, HMW>T1, altitude<T2, slope<T3) is generated by multi-condition threshold constraints. The key thresholds are determined by multi-region sample experiments. Among them, T2 (altitude threshold) and T3 (slope threshold) are calibrated according to ecological zones to adapt to regional topographic differences. Morphological methods are used to remove noise and generate a smooth and continuous wetland potential mask and output it.

[0021] As a further technical solution of the present invention, in step four, the spectral, texture, polarization, and temporal dynamic features of the integrated image are extracted, mainly including the following features: radar backscattering features: including the backscattering intensity of different polarization modes and their derived polarization indices; spectral features and transformation features: including the original spectral bands and vegetation index, water index, and red edge index, while simultaneously performing tassel transformation to obtain brightness, greenness, and humidity components; texture features: based on five indicators of the gray-level co-occurrence matrix (GLCM) (angular second moment ASM, contrast, entropy, correlation, and dissimilarity).

[0022] As a further technical solution of the present invention, in step four, a time series fitting algorithm is used to reconstruct key spectral indices, and statistical parameters that can quantify periodicity, extreme values ​​and degree of variation are extracted from them. Based on the reconstructed curve, the frequency characteristics of vegetation water bodies are extracted using the Otsu algorithm.

[0023] As a further technical solution of the present invention, in step five, the generated potential wetland spatial mask is used as prior knowledge to guide the model to focus on high-probability wetland areas. The specific execution process is as follows: (1) Superpixel segmentation: The annual synthetic image is segmented using a simple non-iterative clustering algorithm to optimize compactness and scale parameters, balance computational efficiency and segmentation accuracy, and use superpixels as the smallest classification unit; (2) Feature and knowledge fusion: For each superpixel, the multidimensional spatiotemporal features of all pixels inside it are averaged to form a superpixel feature vector; the potential distribution mask of wetlands is used as prior knowledge to constrain the classification range to high-probability wetland areas; (3) Model training and classification: The annual training samples and their feature vectors are used as input to train a random forest (RF) classifier, and the trained model is used to classify the wetland types of superpixels, and finally generate a dataset.

[0024] As a further technical solution of the present invention, in step six, based on the generated annual wetland dataset, a spatial overlay analysis method is used to quantify the area change trend and spatial transfer characteristics of each wetland category, calculate the frequency and magnitude of wetland changes, identify hotspots of wetland degradation or restoration, and realize refined monitoring of large-scale wetland dynamics.

[0025] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a large-scale wetland dynamic mapping method that combines knowledge-guided learning and machine learning. Compared with traditional large-scale wetland dynamic monitoring methods, this method has significant technical advantages: First, it greatly reduces the dependence on training samples: Traditional machine learning requires a large number of uniformly distributed training samples throughout the entire study area, while this method constrains sample collection by the potential distribution range of wetlands, requiring only the selection of samples in high-probability wetland areas. The sample size is reduced by more than 40% compared with similar studies, while maintaining the same or even better classification accuracy, thereby reducing the cost of acquiring training samples in the process of multi-temporal large-scale wetland mapping; Second, it significantly improves computational efficiency and classification accuracy: By using the probability of wetland occurrence as prior knowledge and limiting it through spatial constraints... The classification scope extends to high-probability wetland areas, avoiding invalid calculations for non-wetland areas and significantly improving overall processing efficiency. Simultaneously, ecological prior knowledge effectively reduces category confusion and the risk of misclassification. Accuracy is balanced across various wetland types and different ecological zones, with more precise boundary delineation. Third, it enhances model robustness and adaptability: This method constructs prior knowledge based on the ecological mechanisms of wetland formation, including hydrological dynamics, topographic constraints, and the temporal characteristics of vegetation and water bodies, rather than relying solely on data-driven approaches. This results in stronger adaptability to complex ecological environments. In areas with a high proportion of non-wetlands, it effectively eliminates interference from low-probability areas, making the classification results more consistent with actual wetland distribution patterns. Its robustness is significantly better than traditional unconstrained machine learning methods. Fourth, it supports large-scale annual continuous mapping: Relying on the Google Earth Engine (GEE) cloud computing platform and a knowledge-driven, efficient classification framework, it achieves rapid processing of massive multi-source time-series data, breaking through the computational bottleneck of traditional methods in large-scale data processing. It can stably output 10-meter resolution annual wetland datasets, providing a scalable technical paradigm for long-term, continuous wetland dynamic monitoring. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention;

[0027] Figure 2 Flowcharts for image acquisition, preprocessing, and time series analysis are constructed.

[0028] Figure 3 A diagram illustrating the process of estimating the potential wetland extent;

[0029] Figure 4 Flowchart for multidimensional spatiotemporal feature extraction;

[0030] Figure 5 A comparison of the information extraction framework and classification results for a knowledge-driven random forest classifier.

[0031] Figure 6 This is a map showing the dynamic monitoring results of wetlands in hotspot areas. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] Please see the appendix Figure 1 -Appendix Figure 6 The present invention provides an embodiment of a large-scale dynamic wetland mapping method combining knowledge guidance and machine learning, comprising: step one, image acquisition and preprocessing; step two, preparation and optimization of time-series training samples; step three, estimation of potential wetland distribution; step four, extraction of multi-dimensional spatiotemporal features; step five, machine learning classification; and step six, analysis of spatiotemporal dynamic features of wetlands.

[0034] In step one above, Sentinel-1 and Sentinel-2 image products are acquired based on the Google Earth Engine platform. Filtering algorithms are used to remove thermal noise from the images, remove invalid observations such as clouds and cloud shadows, perform radiometric calibration on the images to eliminate sensor errors, and perform terrain correction to eliminate terrain effects. Through the time aggregation function of the Google Earth Engine platform, median image synthesis is completed to fill observation gaps and synthesize high-quality remote sensing images.

[0035] In step two above, sample data integration is carried out by combining multi-regional field survey data with multi-source geospatial products. The sample size is expanded through visual interpretation. The initial samples are merged and standardized according to the wetland classification system to prepare a time-series training sample library. Dynamic annotation and anomaly removal are performed. The time-series analysis module of the Collect Earth platform is used to generate high-resolution time-series images of each sample location. Dynamic annotation is performed by combining annual changes in vegetation cover, surface water range, and surface humidity. The platform's sampling and review tools are used to analyze sample clustering patterns, remove inconsistent abnormal samples, and perform spatial consistency verification. The samples are checked for geographical distribution uniformity to construct a spatially continuous annual training dataset. The total number of samples remains stable, and the category labels are dynamically updated only according to changes in surface conditions to provide reliable sample support for annual classification.

[0036] In step three above, based on multi-temporal remote sensing image data, spectral index features reflecting the characteristics of water bodies, vegetation, and surface humidity are extracted. A temporal feature image set is constructed by combining topographic information, key temporal statistical feature parameters are calculated, and statistical analysis is performed on the temporal feature image set to calculate key parameters reflecting water body occurrence frequency (WOF), vegetation occurrence frequency (VOF), and surface moisture degree (HMW), which are used to characterize the spatiotemporal characteristics of wetlands. Potential wetland ranges are generated through multi-condition threshold constraints (WOF>0 or VOF>0, HMW>T1, altitude<T2, slope<T3). The key thresholds are determined by multi-regional sample experiments, where T2 (altitude threshold) and T3 (slope threshold) are calibrated separately according to ecological zones to adapt to regional topographic differences. Morphological methods are used to remove noise, and a smooth and continuous wetland potential mask is generated and output.

[0037] In step four above, the spectral, texture, polarization, and temporal dynamic features of the integrated image are extracted, mainly including the following features: radar backscattering features: including the backscattering intensity of different polarization modes and their derived polarization indices; spectral features and transformation features: including the original spectral bands and vegetation indices, water body indices, and red edge indices, while simultaneously performing tassel transformation to obtain brightness, greenness, and humidity components; texture features: based on five indicators of the gray-level co-occurrence matrix (GLCM) (angular second moment ASM, contrast, entropy, correlation, and dissimilarity), a time series fitting algorithm is used to reconstruct key spectral indices, from which statistical parameters that can quantify periodicity, extreme values, and degree of variation are extracted. Based on the reconstructed curves, the Otsu algorithm is used to extract the frequency characteristics of vegetation and water bodies.

[0038] In step five above, the generated potential wetland spatial mask is used as prior knowledge to guide the model to focus on high-probability wetland areas. The specific execution process is as follows: (1) Superpixel segmentation: The annual synthetic image is segmented using a simple non-iterative clustering algorithm to optimize compactness and scale parameters, balance computational efficiency and segmentation accuracy, and use superpixels as the smallest classification unit; (2) Feature and knowledge fusion: For each superpixel, the multidimensional spatiotemporal features of all pixels inside it are averaged to form a superpixel feature vector; the potential distribution mask of wetlands is used as prior knowledge to constrain the classification range to high-probability wetland areas; (3) Model training and classification: The annual training samples and their feature vectors are used as input to train a random forest (RF) classifier, and the trained model is used to classify the wetland types of superpixels to finally generate a dataset;

[0039] In step six above, based on the generated annual wetland dataset, the spatial overlay analysis method is used to quantify the area change trend and spatial transfer characteristics of each wetland category, calculate the frequency and magnitude of wetland changes, identify hotspots of wetland degradation or restoration, and achieve refined monitoring of large-scale wetland dynamics.

[0040] Among them, the appendix Figure 2 This demonstrates the process of acquiring Sentinel-2 imagery products using the Google Earth Engine platform, and then performing preprocessing steps such as thermal noise removal, radiometric calibration, terrain correction, Refine-Lee filtering, and cloud and cloud shadow removal. Finally, it showcases the process of synthesizing high-quality remote sensing images by supplementing observational gaps with data from adjacent years. (Attached) Figure 3 This describes the process of extracting NDWI, SDWI, NDVI indices and DEM-derived elevation and slope topographic features from preprocessed time-series images, calculating water body frequency (WOF), vegetation frequency (VOF), and maximum humidity (HMW), and delineating the spatial distribution range of potential wetlands through regional threshold calibration and screening, integrating buffer analysis and morphological dilation algorithms. Figure 4 This demonstrates the process of extracting multidimensional initial features (including polarization, spectral, and texture features) from Sentinel-1 / 2 imagery, obtaining statistical features reflecting dynamic changes in land cover through time-series reconstruction, and constructing a multidimensional spatiotemporal feature set by combining the calculated frequencies of vegetation and water bodies. (Attached) Figure 5 This paper describes an information extraction framework based on spatiotemporal feature sets and a knowledge-driven random forest classifier (incorporating latent wetland distribution constraints), and presents a comparison of the classification results with existing datasets. Figure 6 Taking Poyang Lake and Changzhuang Reservoir as examples, the paper demonstrates how data fusion and feature construction are carried out based on high-resolution multi-source remote sensing images of the areas to be measured. The time series harmonic analysis algorithm is used to extract temporal patterns, a prior knowledge model is introduced as a constraint, and a superpixel segmentation algorithm and a random forest classifier are combined to classify and predict multidimensional spatiotemporal features, ultimately generating an annual dynamic distribution map of wetlands.

[0041] Based on the above, the advantages of this invention are as follows: Traditional machine learning methods typically require a large number of training samples to be evenly distributed throughout the entire study area. However, this invention constrains the sample collection area by estimating the potential distribution range of wetlands, requiring only the selection of samples from high-probability wetland areas. This method reduces the sample size by more than 40% compared to similar studies, while maintaining the same or even higher classification accuracy. This significantly reduces the cost and time of acquiring training samples in multi-temporal, large-scale wetland mapping. By using the potential distribution of wetlands as prior knowledge, this invention can spatially constrain the classification range to high-probability wetland areas, avoiding invalid computations in non-wetland areas, thus greatly improving accuracy. This invention improves overall processing efficiency, while the introduction of ecological prior knowledge effectively reduces category confusion and the risk of misclassification, resulting in more balanced classification accuracy for various wetlands and different ecological zones, and more precise boundary characterization. Based on the ecological mechanisms of wetland formation, including hydrological dynamics, topographic constraints, and temporal characteristics of vegetation and water bodies, a prior knowledge system is constructed. Compared with traditional machine learning methods that rely solely on data-driven approaches, this invention is more adaptable to complex ecological environments. In areas where non-wetlands account for a very high proportion, this invention can effectively eliminate interference from low-probability areas, making the classification results more consistent with the actual wetland distribution patterns and significantly enhancing the robustness of the model. Relying on the Google Earth Engine (GEE) cloud computing platform and a knowledge-driven, efficient classification framework, this invention enables rapid processing of massive multi-source time-series data, solving the computational bottleneck of traditional methods in large-scale data processing. It can stably output high-resolution annual wetland datasets, providing a scalable technical paradigm for long-term, continuous wetland dynamic monitoring, helping to grasp the dynamic changes of wetland resources in a timely manner, and providing a scientific basis for wetland protection and management.

[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for large-scale dynamic wetland mapping combining knowledge-guided learning and machine learning, comprising: Step 1, image acquisition and preprocessing; Step 2, preparation and optimization of time-series training samples; Step 3, estimation of potential wetland distribution; Step 4, extraction of multi-dimensional spatiotemporal features; Step 5, machine learning classification; Step 6, analysis of spatiotemporal dynamic features of wetlands; characterized in that: In step one above, Sentinel-1 and Sentinel-2 images are acquired and preprocessed based on the Google Earth Engine platform, including denoising, cloud removal, radiometric calibration, terrain correction, and median image synthesis, in order to generate high-quality remote sensing images. In step two above, multi-regional field survey data and multi-source geospatial products are integrated, and the sample size is expanded through visual interpretation, dynamic annotation and anomaly removal, and spatial consistency verification to construct a spatially continuous annual training dataset. In step three above, spectral index features are extracted based on multi-temporal remote sensing image data, a temporal feature image set is constructed by combining topographic information, potential wetland range is generated through multi-condition threshold constraints, and morphological methods are used to remove noise. In step four above, the spectral, texture, polarization and temporal dynamic features of the image are extracted, and key spectral indices are reconstructed to extract periodicity, extrema and variability features. In step five above, a potential wetland spatial mask is used as prior knowledge. Through steps such as superpixel segmentation, feature and knowledge fusion, model training and classification, the model is guided to focus on high-probability wetland areas to generate a wetland type classification dataset. In step six above, based on the generated annual wetland dataset, the trend of wetland category area change and spatial transfer characteristics are quantified, hotspots of wetland degradation or restoration are identified, and refined monitoring of wetland dynamics is achieved.

2. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 1, characterized in that: In step one, Sentinel-1 and Sentinel-2 image products are acquired based on the Google Earth Engine platform. Filtering algorithms are used to remove thermal noise from the images, remove invalid observations such as clouds and cloud shadows, perform radiometric calibration on the images to eliminate sensor errors, and perform terrain correction to eliminate terrain effects. Through the time aggregation function of the Google Earth Engine platform, median image synthesis is completed to fill observation gaps and synthesize high-quality remote sensing images.

3. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 1, characterized in that: In step two, sample data integration is carried out by combining field survey data from multiple regions with multi-source geospatial products, expanding the sample size through visual interpretation, merging and standardizing the initial samples according to the wetland classification system, preparing a time-series training sample library, and dynamically labeling and removing anomalies.

4. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 3, characterized in that: In step two, the time series analysis module of the Collect Earth platform is used to generate high-resolution time series images of each sample location. These images are dynamically labeled in conjunction with annual changes in vegetation cover, surface water range, and surface humidity. The platform's sampling and review tools are used to analyze sample clustering patterns, remove inconsistent abnormal samples, perform spatial consistency checks, and check the geographical distribution uniformity of the samples. A spatially continuous annual training dataset is constructed, with the total number of samples remaining stable. The category labels are dynamically updated only based on changes in surface conditions, providing reliable sample support for annual classification.

5. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 1, characterized in that: In step three, based on multi-temporal remote sensing image data, spectral index features reflecting the characteristics of water bodies, vegetation and surface humidity are extracted. A temporal feature image set is constructed by combining topographic information, key temporal statistical feature parameters are calculated, and statistical analysis is performed on the temporal feature image set to calculate key parameters such as water body occurrence frequency (WOF), vegetation occurrence frequency (VOF) and surface moisture degree (HMW), which are used to characterize the spatiotemporal characteristics of wetlands.

6. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 5, characterized in that: In step three, the potential wetland range (WOF>0 or VOF>0, HMW>T1, altitude<T2, slope<T3) is generated by multi-condition threshold constraints. The key thresholds are determined by multi-region sample experiments. Among them, T2 (altitude threshold) and T3 (slope threshold) are calibrated according to ecological zones to adapt to regional topographic differences. Morphological methods are used to remove noise and generate a smooth and continuous wetland potential mask and output it.

7. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 1, characterized in that: In step four, the spectral, texture, polarization, and temporal dynamic features of the integrated image are extracted, mainly... The features include: radar backscattering characteristics: including backscattering intensity of different polarization modes and their derived polarization indices; spectral and transformation characteristics: including the original spectral bands and vegetation index, water index and red edge index, and simultaneously performing tassel transformation to obtain brightness, greenness and humidity components; texture characteristics: based on five indicators of gray-level co-occurrence matrix (GLCM) (angular second moment ASM, contrast, entropy, correlation and dissimilarity).

8. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 7, characterized in that: In step four, a time series fitting algorithm is used to reconstruct key spectral indices, from which statistical parameters that can quantify periodicity, extreme values ​​and degree of variation are extracted. Based on the reconstructed curves, the frequency characteristics of vegetation and water bodies are extracted using the Otsu algorithm.

9. The method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 1, characterized in that: In step five, the generated potential wetland spatial mask is used as prior knowledge to guide the model to focus on high-probability wetland areas. The specific execution process is as follows: (1) Superpixel segmentation: The annual synthetic image is segmented using a simple non-iterative clustering algorithm to optimize compactness and scale parameters, balance computational efficiency and segmentation accuracy, and use superpixels as the smallest classification unit; (2) Feature and knowledge fusion: For each superpixel, the multidimensional spatiotemporal features of all pixels inside it are averaged to form a superpixel feature vector; the potential distribution mask of wetlands is used as prior knowledge to constrain the classification range to high-probability wetland areas; (3) Model training and classification: The annual training samples and their feature vectors are used as input to train a random forest (RF) classifier, and the trained model is used to classify the wetland types of superpixels, and finally generate a dataset.

10. A method for large-scale dynamic wetland mapping combining knowledge guidance and machine learning according to claim 1, characterized in that: In step six, based on the generated annual wetland dataset, a spatial overlay analysis method is used to quantify the area change trends and spatial transfer characteristics of each wetland category, calculate the frequency and magnitude of wetland changes, identify hotspots of wetland degradation or restoration, and achieve refined monitoring of large-scale wetland dynamics.

Images