High and cold wetland dynamic monitoring and health evaluation method and system based on multi-source remote sensing

Abstract

The invention relates to a high and cold wetland dynamic monitoring and health evaluation method and system based on multi-source remote sensing. The method comprises the following steps: step 1, data acquisition and preprocessing; step 2, multi-modal data feature extraction; step 3, performing multi-modal data fusion; step 4, dynamic monitoring; step 5, evaluating wetland health; according to the invention, the calculation method of the SAR polarization decomposition model and the optical wetland index is optimized due to the unique seasonal change and earth surface complexity of the high and cold wetland, a scientific basis is provided for the management of the wetland protection area, and the ecological restoration efficiency is improved; alpine region water resource management and ecological protection policy making are supported; and data support is provided for researching the wetland carbon sink function and the climate change influence.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing monitoring and ecological assessment, and specifically relates to a method of combining multimodal SAR data and multispectral remote sensing data and applying a deep learning algorithm to conduct full-time and space dynamic monitoring and ecological health evaluation of alpine wetlands. Background Art

[0002] Alpine wetlands play an irreplaceable role in water conservation, carbon sink balance and climate regulation. However, global climate change and intensified human activities have led to increasingly serious problems such as degradation, area reduction and biodiversity loss of alpine wetlands.

[0003] Existing technologies usually use single optical or SAR data for wetland monitoring, which has the following main shortcomings: (1) Data uniqueness Optical remote sensing data are easily restricted by cloud cover, sunshine conditions, etc., and it is difficult to ensure multi-temporal and full coverage acquisition; although SAR data can penetrate clouds, it has limitations in extracting vegetation details and differences in wetland types.

[0004] (2) Insufficient dynamic monitoring Most studies focus on static monitoring of short-term wetland status and lack a comprehensive evaluation of the hydrological and ecological evolution of wetlands over multiple years and seasonal scales; the monitoring indicators are single and cannot fully reflect the dynamic changes of multiple factors such as wetland water bodies, vegetation and soil moisture.

[0005] (3) Single health assessment model Traditional wetland health assessment mostly relies on a small number of indicators or a simple single index, which makes it difficult to fully utilize the advantages of SAR and optical multi-source data in the comprehensive analysis of wetland ecology; the model accuracy and applicability are difficult to meet the special environmental needs of alpine wetlands.

[0006] To this end, the present invention deeply integrates SAR and optical remote sensing data, combines spatiotemporal convolutional neural network (ST-CNN) and attention mechanism to conduct long-term and dynamic monitoring of hydrology, vegetation and soil moisture in alpine wetlands, and adopts a multi-index ecological health evaluation system (AHP) to quantify the health status of wetlands, so as to achieve refined management and protection of alpine wetlands. Summary of the invention

[0007] The present invention aims to overcome the problems of incomplete monitoring of alpine wetlands by single remote sensing data, insufficient dynamic monitoring and health evaluation, and provide a method and system for dynamic monitoring and health evaluation of alpine wetlands based on multi-source remote sensing. By integrating the complementary advantages of SAR and optical data, combining deep learning multimodal feature extraction and AHP health evaluation model, comprehensive monitoring and health classification of soil moisture, hydrological range, vegetation coverage, etc. of alpine wetlands at different seasons and different annual scales can be achieved, providing a scientific decision-making basis for the ecological protection and restoration of wetlands in alpine areas.

[0008] In order to solve the above-mentioned technical problems, the present invention adopts the following solutions: 1. Methods The dynamic monitoring and health assessment method of alpine wetlands based on SAR and optical remote sensing includes the following steps: Step 1: Data collection and preprocessing; Acquire SAR image data, optical image data, auxiliary data and other data, and pre-process them respectively.

[0009] Step 2: Multimodal feature extraction; Through SAR features, optical features are extracted.

[0010] Step 3: Multimodal data fusion; By adopting a spatiotemporal convolutional neural network (ST-CNN), the SAR and optical multi-temporal data are input and combined with an attention mechanism to improve the sensitivity to wetland characteristics.

[0011] Step 4: Dynamic monitoring; Through the fused feature data, based on the random forest (RF) or support vector machine (SVM) classification algorithm, wetlands are divided into permanent wetlands, seasonal wetlands, peatlands and degraded wetlands; the dynamic change trends of wetlands, wetland area, vegetation coverage and water body expansion and contraction are analyzed through time series data.

[0012] Step 5: Wetland health assessment; Construct health indicators: Water Health Index (WHI): based on changes in water area and water connectivity; Vegetation Health Index (VHI): combines NDVI and surface temperature changes to quantify the health of vegetation growth; Soil Moisture Index (SMI): estimates the soil moisture status of wetlands based on SAR texture features.

[0013] Comprehensive evaluation model: Construct a health evaluation model based on weight distribution (AHP hierarchical analysis method) to integrate various indicators into a comprehensive score of wetland ecological health.

[0014] Furthermore, the step 1 inputs the data source: SAR image: Use VV / VH polarization data from the European Space Agency (ESA) Sentinel-1 satellite or ALOS PALSAR data from the Japan Aerospace Exploration Agency (JAXA). The repetition period of Sentinel-1 is 6 days, with a spatial resolution of about 5 meters; the spatial resolution of ALOS PALSAR is about 10 meters, and the repetition period is about 14 days. Focus on obtaining the scattering characteristics of wetland water and soil moisture, and use multi-polarization and multi-temporal characteristics to analyze the hydrological dynamics of the wetland.

[0015] Optical imagery: Use the 10-meter resolution multispectral imagery (with a repetition period of 5 days) from the European Space Agency (ESA) Sentinel-2 satellite or the 30-meter resolution imagery (with a repetition period of 16 days) from the United States Geological Survey (USGS) Landsat series to obtain information on wetland vegetation coverage, surface water distribution, and surface temperature, and to support the calculation of vegetation index and water index.

[0016] Auxiliary data: The 30-meter resolution DEM data provided by Shuttle Radar Topography Mission (SRTM) or ASTGTM V2 (Advanced Spaceborne Thermal Emission and ReflectionRadiometer Global Digital Elevation Model) are used to extract the topographic features of wetlands, such as identification of low-lying areas and analysis of water body convergence points. The ERA5 reanalysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF) or the regional meteorological data provided by the China Meteorological Administration (ERA5 provides data with a global grid resolution of about 31 kilometers per hour) are used to analyze the impact of meteorological factors such as precipitation and temperature on wetland hydrological dynamics and support the calibration and verification of wetland hydrological models. The Global Land Cover by National Mapping Organizations (GLCNMO) dataset provided by the United States Geological Survey (USGS) or the LULC data provided by the State Forestry and Grassland Administration are used for wetland type classification, land use change analysis, and to assist in wetland classification and health assessment. The SSURGO (Soil Survey Geographic) database or the China Soil Database provided by the United States Department of Agriculture (USDA) is used to calibrate the wetland soil moisture inversion model to improve the accuracy of soil moisture estimation.

[0017] The data preprocessing process in step 1 is as follows: 1) SAR data preprocessing method steps: S11, speckle noise filtering: use Lee filter or IDAN filter to remove speckle noise in SAR images and improve the signal-to-noise ratio of subsequent feature extraction; S12, Geometric correction and radiation correction: Geometric correction is performed based on precise orbit data to ensure the spatial consistency of SAR images and optical images; radiation correction is used to correct the amplitude value of SAR images and unify data obtained at different times or by different sensors; S13. Polarization decomposition: The Freeman-Durden decomposition method is used to decompose the dual-polarization (VV / VH) SAR data into surface scattering, vegetation scattering and volume scattering components to extract the surface scattering characteristics of the wetland.

[0018] 2) Optical data preprocessing method steps: S21, cloud and cloud shadow removal: Fmask algorithm is used to identify and remove cloud and cloud shadow areas in optical images to ensure accurate extraction of vegetation and water features; S22, Atmospheric Correction: Use 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) or LEDAPS (Landsat Ecosystem Disturbance Adaptive Processing System) method to perform atmospheric correction and obtain surface reflectance; S23. Time series interpolation completion: Based on spatial and temporal interpolation algorithms such as STARFM (Spatial and Temporal Adaptive Reflectance Fusion Model), missing data in multi-temporal optical images caused by cloud cover or other factors are completed.

[0019] 3) Data alignment method Using pixel-level registration algorithms, such as the registration method based on the cross-correlation method or the registration method based on feature point matching (such as SIFT, SURF, etc.), the SAR image and the optical image are accurately aligned in the same coordinate system to ensure the spatial consistency of multi-source data and provide a basis for subsequent data fusion processing.

[0020] Furthermore, the multimodal features in step 2 mainly include the following features: 1) SAR feature extraction: Scattering characteristics: Utilize the VV (vertical transmission and vertical reception) and VH (vertical transmission and horizontal reception) polarization information of Sentinel-1 or ALOS PALSAR data. The reason for choosing these two polarization modes is that they are highly sensitive and recognizable to different types of land objects in alpine wetland environments. Due to its vertical propagation characteristics, VV polarized waves are particularly sensitive to the specular reflection of the water surface and can effectively distinguish high-scattering water areas. In addition, due to its horizontal reception characteristics, VH polarized waves can better capture the volume scattering and multi-directional scattering effects of vegetation, thereby distinguishing medium-scattering vegetation areas. By calculating the VV / VH ratio and the Polarization Difference Index (PDI), the contrast between water and vegetation, and vegetation and bare land is further enhanced, thereby improving the classification accuracy. In addition, combined with additional scattering features such as dual polarization ratios and polarization difference indices, the present invention can more comprehensively describe different types of land objects in wetlands, ensuring high-precision classification and monitoring in complex alpine environments.

[0021] Texture features: The Gray-Level Co-occurrence Matrix (GLCM) method is used to quantify the roughness and moisture state of wetland soil from SAR images. GLCM can capture the spatial distribution characteristics of the surface microstructure and evaluate the texture complexity and moisture changes of the soil surface by calculating statistical indicators such as contrast, homogeneity, energy and correlation. The surface of alpine wetlands often has freeze-thaw cycles and seasonal wetness changes, and texture features can effectively reflect these dynamic processes. For example, an increase in soil roughness is usually related to a reduction in vegetation cover or human interference, while changes in moisture state directly affect the reflective properties of soil texture. Combined with multi-scale texture features (such as scale-invariant feature transform SIFT and local binary pattern LBP), the present invention can capture the detailed changes of wetland surfaces at different spatial scales, improve the detail and accuracy of wetland type distinction, and has significant application effects under complex surface conditions in particularly alpine environments.

[0022] 2) Optical feature extraction: NDVI (Normalized Difference Vegetation Index): By calculating the difference in reflectance between the near-infrared band (NIR) and the red band (Red), the formula is: NDVI = (NIR - Red) / (NIR + Red), which is used to quantify the density and health of vegetation. In alpine wetland environments, vegetation coverage has a key impact on ecological health. NDVI can effectively reflect the density and growth status of vegetation, and is suitable for capturing seasonal growth changes in vegetation and assessing its health status. Due to the short growth period of alpine wetland vegetation, NDVI can provide highly sensitive vegetation monitoring data in a short period of time, supporting timely assessment of dynamic changes in vegetation.

[0023] MNDWI (Modified Normalized Difference Water Index): Calculated through the reflectance of the green band (Green) and the shortwave infrared band (SWIR), the formula is: MNDWI = (Green - SWIR) / (Green + SWIR), which is used to enhance the recognition of water body boundaries. MNDWI is highly sensitive to the recognition of water body boundaries and is suitable for complex wetland environments, especially in areas with multiple vegetation and multiple water bodies. Compared with traditional water body indices, MNDWI can reduce the interference of buildings and shadows on water body identification and improve the accuracy and stability of water body extraction. In addition, when used in combination with NDVI, MNDWI can provide more comprehensive vegetation and water body information, enhance the accuracy of wetland classification and monitoring, and support a comprehensive assessment of dynamic changes in wetland hydrology.

[0024] LST (Land Surface Temperature): The temperature distribution of the surface is estimated through remote sensing data in the thermal infrared band. Usually, a single-band or multi-band radiation temperature calculation method is used to reflect the thermal state of the surface. In the alpine wetland environment, the surface temperature is closely related to vegetation health and the thermal characteristics of water bodies. LST can reflect the transpiration of vegetation and the heat capacity characteristics of water bodies, help identify areas where vegetation is stressed, and monitor the dynamic changes in the thermal state of the surface. Alpine wetlands are significantly affected by changes in temperature. LST provides real-time monitoring of the thermal environment of the wetland surface and supports dynamic response assessments to environmental changes.

[0025] 3) Dynamic features: Use Sentinel-1 (SAR) and Sentinel-2 / Landsat (optical) image data acquired in multiple temporal phases (monthly or quarterly) to construct a time series dataset of wetlands. The data at each time node include key parameters such as scattering characteristics (VV / VH), spectral indices (NDVI, MNDWI) and land surface temperature (LST). By analyzing the characteristic changes in different seasons, the growth peak and decline process of wetland vegetation, as well as the seasonal fluctuations in water area, are identified. For example, the seasonal growth peak and winter degradation trend of vegetation are detected using time series NDVI data; the expansion and contraction of water bodies in rainy and dry seasons are monitored through MNDWI time series. Combined with multi-year (such as 5 or 10 years) time series data, the long-term change trend of wetland hydrological dynamics and vegetation coverage is evaluated. Statistical analysis methods (such as linear regression, moving average) are used to detect the growth or decrease trend of wetland water area, as well as the long-term increase or decrease trend of vegetation index. And use anomaly detection algorithms in time series data (such as Z-score, outlier detection models) to identify sudden changes in wetland ecosystems, such as abnormal droughts, floods, or vegetation mutations caused by human interference.

[0026] Furthermore, the step 3 uses a spatio-temporal convolutional neural network (ST-CNN) combined with an attention mechanism to perform deep fusion processing on SAR and optical multi-temporal data.

[0027] 1) Space-time convolutional neural network (ST-CNN) architecture design: ST-CNN is designed to capture feature changes in both spatial and temporal dimensions and adapt to the complex spatiotemporal dynamic environment of alpine wetlands. Its main architecture includes the following components: Input Layer: SAR feature input: including scattering features (VV, VH), texture features (GLCM indicators) and dynamic features (time series change curve).

[0028] Optical characteristic input: including spectral index (NDVI, MNDWI, LST) and time series change characteristics (such as EVI, SAVI, GNDVI).

[0029] Convolutional Layers: Spatial convolution layer: Use a two-dimensional convolution kernel (such as 3×3 or 5×5) to extract the spatial features of the input data and capture the spatial distribution and structural information of the objects.

[0030] Temporal convolution layer: A one-dimensional convolution kernel (such as 1×3 or 1×5) is used to process time series features along the time dimension to capture the temporal dependence of wetland dynamic changes.

[0031] Pooling Layers: Spatial pooling: Use maximum pooling or average pooling operations to reduce the spatial dimension of the feature map and extract high-level spatial features.

[0032] Time pooling: Use the pooling operation in the time dimension to compress time series features and retain key dynamic information.

[0033] Attention Mechanism Layers: Channel Attention: Adaptively adjust the importance weights of different feature channels to highlight key features and suppress irrelevant or redundant features. Global average pooling and global maximum pooling operations are used to generate channel weight vectors, and the attention weight of each channel is calculated through a fully connected layer and an activation function (such as Sigmoid).

[0034] Spatial Attention: By adjusting the importance of features in different spatial regions, the characteristic response of the core area of ​​the wetland is highlighted and the interference of background or noise areas is suppressed. The spatial information of different feature maps is combined to generate a spatial weight map, and the attention weight of each spatial position is calculated through convolution operations.

[0035] Fusion Layers: Feature Concatenation: Concatenate the SAR and optical feature maps processed by the channel and spatial attention mechanisms to form a comprehensive feature representation.

[0036] Weighted Fusion: Through the learned weights, feature maps from different sources are weighted summed to fuse multi-source information.

[0037] Fully Connected Layers: Function: Map the fused high-dimensional feature vector to the output space of the classification or regression task to achieve wetland type classification and dynamic change trend prediction.

[0038] Activation function: The Softmax activation function is used to output the classification probability for multi-category classification of wetland types.

[0039] Output Layers: Wetland classification results: Output classification probability maps of various types of wetlands (permanent wetlands, seasonal wetlands, peatlands, degraded wetlands).

[0040] Dynamic change trend chart: Output dynamic change trend prediction chart of key indicators such as wetland water area, vegetation coverage and soil moisture.

[0041] 2) Integration method of attention mechanism The attention mechanism plays a key role in this invention. By adaptively adjusting the importance weights of features, the model's sensitivity to key features of alpine wetlands is improved. The specific integration method is as follows: Channel Attention implementation process: Perform global average pooling and global maximum pooling on the input feature map to generate two different context description vectors.

[0042] These two vectors are input into the fully connected layer, after passing through the ReLU activation function, the channel weight vector is generated through the Sigmoid activation function.

[0043] The channel weight vector is multiplied by the original feature map channel by channel to adjust the feature response strength of each channel.

[0044] Spatial Attention implementation process: The input feature map is average pooled and max pooled in the channel dimension to generate two single-channel feature maps.

[0045] The two single-channel feature maps are concatenated in the channel dimension to form a dual-channel feature map.

[0046] Generate a spatial weight map through a convolution layer (such as a 7x7 convolution kernel) and a Sigmoid activation function.

[0047] The spatial weight map is multiplied pixel by pixel with the original feature map to adjust the feature response intensity at each spatial position.

[0048] 3) Data fusion process Feature extraction: Extract scattering features, texture features, dynamic features and spectral index from SAR and optical data respectively.

[0049] Feature preprocessing: Standardize the extracted features to ensure the consistency of the numerical range of different features and improve the fusion effect.

[0050] Feature input: The processed SAR features and optical features are input into different network branches of ST-CNN respectively.

[0051] Feature fusion: In ST-CNN, spatial and temporal features are extracted through convolutional layers and pooling layers. Channel attention and spatial attention mechanisms are integrated to optimize feature weight distribution and highlight key features. SAR and optical features are deeply fused through feature concatenation or weighted fusion layers to form a comprehensive feature representation.

[0052] Classification and prediction: Through the fully connected layer and Softmax activation function, multi-category classification of wetland types and prediction of dynamic change trends are achieved.

[0053] 4) Training and verification process Model training and verification are key links to ensure data fusion effect and model performance. The specific process is as follows: Sample labeling: Wetland type labels are obtained through field surveys or high-resolution image data to form training sets and validation sets. Ensure that the samples cover different wetland types and different time nodes to enhance the generalization ability of the model.

[0054] Loss function: Cross-Entropy Loss is used to optimize the accuracy of classification tasks. Temporal Consistency Regularization is introduced to ensure that the prediction results of the model in the time dimension are smooth and consistent, reducing mutations and noise interference.

[0055] Optimization algorithm: Use Adam Optimizer, set the learning rate to 0.001, and combine momentum and adaptive learning rate adjustment strategies to accelerate model convergence. Set training epochs and batch size, and dynamically adjust according to the accuracy of the validation set to avoid overfitting.

[0056] Model validation: Use an independent validation set to evaluate model performance, and use indicators such as accuracy, precision, recall, and F1 score to measure classification results. Cross-validation is used to further verify the robustness and generalization ability of the model.

[0057] Further, in step 4, after completing the multimodal data fusion, the dynamic monitoring step accurately divides the types of alpine wetlands by applying classification algorithms and time series analysis, and monitors their dynamic change trends. The specific implementation of this step includes wetland type classification, calculation and analysis of key monitoring indicators, and output and application of monitoring results. The specific contents are as follows: 1) Classification of wetland types Alpine wetlands are divided into different types according to their hydrological characteristics and ecological status, including permanent wetlands, seasonal wetlands, peatlands and degraded wetlands, so as to carry out targeted ecological management and protection.

[0058] Classification algorithm selection: Random Forest (RF): Utilizes its advantages in processing high-dimensional features and resisting overfitting to perform multi-category classification of wetland types.

[0059] Support Vector Machine (SVM): Suitable for data classification in high-dimensional space, especially when the sample size is small and the category distribution is uneven.

[0060] Classification features: Fusion features: Based on the comprehensive features fused in step 3, including SAR scattering features, optical spectral index, texture features and dynamic features.

[0061] Time series characteristics: Combine the characteristic change information of wetlands at different time nodes to improve the temporal and spatial accuracy of classification.

[0062] Classification standards and judgment basis: Permanent wetlands: defined as areas where the water coverage time and humidity state remain stable for most of the time phases (e.g. more than 9 months a year) for at least three consecutive years, with seasonal fluctuations in water levels less than 5%. The water coverage time accounts for more than 75% of the year, and the soil moisture index (SMI) trend is stable as the basis for judgment.

[0063] Seasonal wetlands: defined as wetlands with obvious seasonal fluctuations in water level or water area within the year, and the seasonal extreme differences in water area exceeding 30%.

[0064] Peatlands: Combining SAR scattering characteristics (low scattering coefficient, usually less than 0.4), optical vegetation index (low NDVI value, usually <0.3) and soil type data, it can identify wetland types containing large amounts of organic matter, showing special polarization responses and textural characteristics.

[0065] Degraded wetlands: Based on multi-source indicators (such as an annual average NDVI value below 0.3 and an interannual decline rate of more than 5%), combined with a continued downward trend in the soil moisture index (SMI) or a significant increase in human disturbance indices (such as land use change, grazing intensity), they are judged to be degraded wetlands.

[0066] 2) Calculation and analysis of key monitoring indicators Monitoring indicators: Wetland area: Based on the classification results, the geographic information system (GIS) tools of remote sensing images are used to calculate the areas of different wetland types and generate annual wetland area change curves.

[0067] Vegetation coverage: Use vegetation indices such as NDVI, EVI, and SAVI to calculate vegetation coverage area and vegetation health change trends, and generate a vegetation coverage time series curve.

[0068] Water body expansion and contraction: Based on the MNDWI time series data, the annual increase and decrease of water body area is analyzed to generate a dynamic change curve of water body expansion and contraction.

[0069] Soil moisture: The soil moisture index (SMI) extracted by SAR is combined with meteorological data to calculate the annual average soil moisture and the moisture change trend during key periods.

[0070] 3) Output and application of monitoring results Output: Annual wetland change map: shows the spatial distribution and area changes of different wetland types in different years. It can visually show the expansion or contraction trend of wetland area and assist in identifying ecological hot spots and degradation risk areas.

[0071] Wetland change curve: including time series change curves of indicators such as wetland area, vegetation coverage, water body expansion and contraction, and soil moisture. Analyze the long-term trend and seasonal fluctuations of wetland dynamics to support continuous assessment and trend prediction of wetland health status.

[0072] Health level zoning map: Based on the comprehensive health score, wetlands are divided into four health levels: excellent, good, medium, and poor, and the areas of each level are marked on the map. This provides an intuitive spatial distribution reference for the delineation of regional wetland protection areas and the priority restoration of degraded areas.

[0073] Further, in step 5, after completing the dynamic monitoring, the wetland health assessment step comprehensively evaluates the ecological health status of the alpine wetland by constructing and calculating multiple ecological health indicators. This step includes the calculation of health indicators, the comprehensive evaluation of the analytic hierarchy process (AHP), and the formulation of health zoning and restoration recommendations. The specific contents are as follows: 1) Calculation of health indicators By quantifying key ecological elements such as wetland water bodies, vegetation and soil moisture, the ecological health of wetlands can be comprehensively assessed.

[0074] Health indicators: Water Health Index (WHI): WHI reflects the ecological health of wetland water bodies by comprehensively analyzing the rate of change of water area and water connectivity. The annual water area growth rate or decrease rate is calculated using the time series data of water area obtained through dynamic monitoring. The density and connectivity of the water body network are analyzed based on the digital elevation model (DEM) and water body distribution data. The WHI score is obtained by weighted summation of the water area change rate and connectivity index.

[0075] Vegetation Health Index (VHI): VHI quantifies the health of wetland vegetation by combining the rate of change of vegetation coverage, peak vegetation growth and surface temperature. Using NDVI time series data, calculate the annual rate of increase or decrease in vegetation coverage. Identify the NDVI peak during the vegetation growing season to reflect the growth intensity of vegetation. Combined with LST data, evaluate the impact of high temperature on vegetation health, and calculate the VHI score comprehensively through a weighted formula.

[0076] Soil Moisture Index (SMI): SMI reflects the health of wetland soil moisture by inverting soil moisture conditions. Using SAR texture features (such as roughness index extracted by GLCM) and meteorological data, a soil moisture inversion model is established to calculate the annual average soil moisture or key period value. Multi-temporal SMI data are analyzed to evaluate the long-term trend of soil moisture (such as the annual average moisture change rate). The current state of soil moisture is compared with historical data, and the SMI score is obtained through standardization.

[0077] 2) Comprehensive evaluation using the analytic hierarchy process (AHP) The Analytic Hierarchy Process (AHP) was used to assign weights and comprehensively evaluate multiple health indicators to generate a comprehensive score for wetland ecological health.

[0078] Implementation method: Hierarchical structure construction: the comprehensive score of wetland ecological health is taken as the target layer, the water health index (WHI), vegetation health index (VHI) and soil moisture index (SMI) are taken as the criterion layer, and the specific sub-indicators under each criterion layer (such as water area change rate, vegetation coverage growth rate, soil moisture mean, etc.) are taken as the indicator layer.

[0079] Judgment matrix construction: Through expert scoring or comparison based on historical data, a judgment matrix at the criterion level and indicator level is constructed to reflect the importance relationship between the indicators. The weights of each level are calculated to ensure the consistency of the judgment matrix (CI < 0.1).

[0080] Weight allocation: The weights calculated by the analytic hierarchy process are allocated to each health indicator to reflect its relative importance in the comprehensive evaluation. For example: WHI: 0.4; VHI: 0.4; SMI: 0.2.

[0081] Calculation of comprehensive score: The scores of each health indicator are weighted and summed according to their weights to obtain the comprehensive score of wetland ecological health. For example, the calculation formula is: comprehensive score = (WHI × 0.4) + (VHI × 0.4) + (SMI × 0.2).

[0082] 3) Healthy Partitions and Repair Suggestions Based on the comprehensive score, wetlands are divided into different health levels, and corresponding ecological restoration and management suggestions are put forward for different level areas.

[0083] Implementation method: Score range division: Excellent: comprehensive score > 0.8; Good: 0.6 ≤ comprehensive score ≤ 0.8; Fair: 0.4 ≤ comprehensive score < 0.6; Poor: comprehensive score < 0.4.

[0084] Generation of health grade zoning map: Based on the comprehensive score, the wetland space is divided into four health grade areas: excellent, good, medium and poor, and a health zoning map is generated.

[0085] (II) System part The alpine wetland dynamic monitoring and health assessment system based on SAR and optical remote sensing includes the following modules: (1) Data acquisition and preprocessing module Responsible for collecting SAR images, optical images and auxiliary data (DEM, meteorology, LULC, soil type, etc.), and performing pre-processing steps such as denoising, geometric correction, radiation correction, cloud shadow removal, atmospheric correction, time series interpolation and completion, and registration to ensure the consistency of multi-source data in space and time.

[0086] (2) Multimodal feature extraction module SAR feature extraction: Scattering, texture and dynamic features are extracted from preprocessed SAR images, and the soil moisture and roughness are quantified using polarimetric decomposition and GLCM methods. Optical feature extraction: Extract spectral indices such as NDVI, MNDWI, LST, and dynamic characteristics of vegetation growth on a monthly or quarterly basis from preprocessed optical images; Output a multimodal feature grid or feature vector that satisfies subsequent deep learning fusion.

[0087] (3) Multimodal data fusion module The spatiotemporal convolutional neural network (ST-CNN) is used in combination with the attention mechanism to deeply fuse SAR and optical multi-temporal data; Network branch design: Branch 1 (SAR subnet): Processes SAR scattering and texture features to extract hydrological and soil moisture characteristics of wetlands; Branch 2 (optical subnet): Processing optical vegetation and water characteristics, extracting wetland vegetation coverage and water distribution characteristics; Attention Mechanism: Channel Attention: Assign adaptive weights to different polarization / band features to highlight important features; Spatial Attention: Enhance the characteristic response of the core area of ​​the wetland and suppress the interference of irrelevant areas; Fusion and output: Through feature splicing or weighted fusion layer, the features of branch one and branch two are fused to output wetland classification results, dynamic change trend chart and key feature distribution map.

[0088] (4) Wetland dynamic monitoring module Using the fused features, wetlands are divided into permanent wetlands, seasonal wetlands, peatlands, and degraded wetlands through classification algorithms such as random forest (RF) or support vector machine (SVM).

[0089] Combined with time series images, the dynamic changing trends of wetlands are analyzed, including key indicators such as wetland area, vegetation coverage, and water body expansion and contraction.

[0090] Output monitoring results such as annual wetland change maps and area change curves.

[0091] (5) Wetland health assessment module Health index calculation: Calculate the Water Health Index (WHI), Vegetation Health Index (VHI), and Soil Moisture Index (SMI).

[0092] Comprehensive evaluation: The analytic hierarchy process (AHP) was used to assign weights to each indicator (e.g., WHI 0.4, VHI 0.4, SMI 0.2) and calculate the comprehensive score of wetland ecological health.

[0093] Wetland health levels are divided according to score ranges (excellent, good, medium, and poor).

[0094] (6) Output and decision support module Generate wetland health grade zoning maps and annual wetland change reports.

[0095] Provide wetland protection and restoration decision-making recommendations, including restoration priorities for degraded areas and protected area demarcation strategies.

[0096] Provide dynamic monitoring data and health assessment reports to alpine wetland ecological management departments to support scientific decision-making.

[0097] The present invention provides a method and system for dynamic monitoring and health assessment of alpine wetlands based on multi-source remote sensing, which has the following beneficial effects: (1) The technical solution of the present invention overcomes the limitation of optical images being obscured by clouds by fusing SAR and optical data, and analyzes the dynamic changes of wetland soil moisture and water bodies through SAR data. This combination of multi-source data improves the ability to jointly analyze wetland hydrology, vegetation and soil moisture.

[0098] (2) The present invention integrates dynamic monitoring and health assessment. Through the spatiotemporal convolutional neural network (ST-CNN) combined with the attention mechanism, it can simultaneously process the spatiotemporal characteristics of SAR and optical data, and realize long-term monitoring of the dynamic changes of wetlands. Especially under the dual influence of climate change and human activities, the system can provide timely ecological health assessment.

[0099] (3) The multi-task deep learning network of the present invention designs a multi-task network based on spatiotemporal convolution and attention mechanism, which improves the accuracy of wetland classification and change detection while meeting the output requirements of multi-task results.

[0100] (4) The present invention optimizes the characteristics of alpine wetlands. In view of the unique seasonal changes and surface complexity of alpine wetlands, the SAR polarization decomposition model and the calculation method of the optical wetland index are optimized to make the model more suitable for actual application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 : The alpine wetland dynamic monitoring and health assessment method and system flow chart based on multi-source remote sensing in the present invention.

[0102] Figure 2 :The time series curve of the continuous water area in the Ruoergai wetland area of ​​the present invention.

[0103] Figure 3 :The seasonal water area time series curve of the Ruoergai wetland in the present invention.

[0104] Figure 4 : The distribution range of surface water bodies in the typical area of ​​Ruoergai of the present invention.

[0105] Figure 5 : Distribution map of surface water bodies in typical lake areas of the Yangtze River source in the present invention.

[0106] Figure 6 : The present invention shows the distribution map of wetland health index of the source area of ​​the Yangtze River in 2018.

[0107] Figure 7 : The present invention shows the distribution map of wetland health index of the source area of ​​the Yangtze River in 2022.

[0108] Example 1: The technical solution of the present invention is used to monitor hydrological changes in the Zoige Wetland The Zoige Wetland is a typical representative of China's alpine wetlands, and its hydrological changes are crucial to the regional ecosystem. This embodiment aims to monitor the dynamic hydrological changes of the Zoige Wetland through multi-temporal SAR and optical images.

[0109] step: (1) Data collection: SAR data: Sentinel-1 images (VV and VH polarization) will be acquired annually from 2015 to 2023, with sampling times in the wet season (July) and the dry season (October).

[0110] Optical data: Sentinel-2 images are acquired during the same period (three views in April, July and October each year).

[0111] Auxiliary data: DEM data is obtained for the extraction of low-lying areas in wetlands; meteorological data is obtained for the analysis of the correlation between precipitation and wetland hydrological changes.

[0112] (2) Data preprocessing: 1) Preprocessing of SAR data: Use Lee filter to denoise and reduce speckle noise interference; Ensure consistency with optical imagery through georeferencing; Extract the surface scattering type based on Freeman-Durden decomposition.

[0113] 2) Processing of optical data: Use the Fmask algorithm to remove clouds and cloud shadows; Apply the atmospheric correction algorithm (6S) to obtain the surface reflectance; The MNDWI (Modified Water Body Index) is calculated to enhance water body boundary identification.

[0114] (3) Wetland feature extraction: 1) SAR data: extract high scattering areas of wetland water bodies through VV polarization signals, and analyze wetland soil moisture by combining VH polarization textures; 2) Optical data: MNDWI is used to distinguish the boundary between wetland water bodies and vegetation, and NDVI is used to analyze the relationship between vegetation and water bodies.

[0115] (4) Multimodal fusion and monitoring model: 1) Using the spatiotemporal convolutional network (ST-CNN): Input: SAR scattering characteristics and optical water index; Convolutional structure: Convolutional extraction of SAR and optical features is performed separately, and the attention mechanism is combined to enhance the representation of wetland area features; Temporal dimension analysis: capturing wetland dynamics through time series ST-CNN.

[0116] 2) Results: A wetland area change map from 2015 to 2023 was generated to distinguish the dynamic range of seasonal wetlands and permanent wetlands.

[0117] (5) Output and analysis: 1) Surface water distribution area change curve: The surface water area of ​​wetlands was calculated year by year, and a time series change curve was drawn. It was found that the surface water fluctuated between 2015 and 2023, with two peaks in 2018 and 2021. The continuous water body has an increasing trend in the past eight years, with a growth rate of about 1.5km2 / a, and the seasonal water body has a slight downward trend as a whole.

[0118] 2) Spatial distribution map: Generate a water distribution map of typical areas of Zoige Wetland, showing that seasonal water bodies are mainly distributed in the surrounding areas of lakes and rivers, as well as low-lying areas in the wetland. The seasonal water distribution area changes most significantly in the low-lying areas of the wetland.

[0119] 3) Report generation: By analyzing the correlation between water body changes and precipitation in combination with meteorological data, it was found that the reduction of water bodies is closely related to wetland precipitation and surface evapotranspiration.

[0120] It is recommended to implement dynamic water monitoring in areas with significant water changes; and output wetland dynamic change reports for reference by wetland management departments.

[0121] Example 2: Health evaluation of the Yangtze River source wetland using the technical solution of the present invention The Yangtze River source wetlands are an important ecological barrier on the Qinghai-Tibet Plateau, and their health directly affects the stability of the downstream river system. This example focuses on assessing the health of wetlands and providing protection and restoration recommendations.

[0122] step: (1) Data collection: SAR data: Acquire Sentinel-1 multi-temporal images from 2018 to 2022 and monthly images during the growing season (April to October) each year.

[0123] Optical data: Acquire Sentinel-2 multispectral images from the same period, focusing on analyzing wetland vegetation coverage.

[0124] Auxiliary data: Obtain meteorological data (precipitation, temperature) and DEM data.

[0125] (2) Data preprocessing: The same processing method as in Example 1 is used to complete SAR and optical image denoising, correction and registration to ensure the consistency of multi-modal images.

[0126] (3) Extraction of wetland health indicators: Water Health Index (WHI): The water distribution is extracted through MNDWI, and the water health score is calculated based on the trend of water area changes; Vegetation Health Index (VHI): quantifies vegetation growth and cover status based on NDVI and LST (land surface temperature); Soil moisture index (SMI): It uses the texture characteristics of SAR images to estimate soil moisture and reflect the hydrological conditions of wetlands.

[0127] (4) Comprehensive evaluation model: The comprehensive health score was calculated using the analytic hierarchy process (AHP): The weights were set to 0.4 for WHI, 0.4 for VHI, and 0.2 for SMI; The comprehensive score is divided into four levels: excellent (>0.8), good (0.6-0.8), fair (0.4-0.6), and poor (<0.4).

[0128] (5) Evaluation result output: Zoning map: Generate a healthy zoning map of the Yangtze River source wetland; Trend analysis: Combining time series analysis with wetland health change trends, it was found that the vegetation health index has been declining year by year, which is significantly correlated with rising temperature and decreasing precipitation.

[0129] (6) Repair suggestions: Implementing restoration projects for degraded wetlands and restricting grazing activities in “poor” rated areas; Propose zoning management suggestions, implement strict protection for “excellent” areas, and carry out ecological water replenishment for “medium” areas.

[0130] Conclusion: Provide wetland protection and restoration priority planning, assist in the formulation of ecological protection policies; output health assessment reports to support the design of wetland ecological compensation programs.

[0131] The present invention provides a method and system for dynamic monitoring and health assessment of alpine wetlands based on multi-source remote sensing, which has the following beneficial effects: (1) The technical solution of the present invention overcomes the limitation of optical images being obscured by clouds by fusing SAR and optical data, and analyzes the dynamic changes of wetland soil moisture and water bodies through SAR data. This combination of multi-source data improves the ability to jointly analyze wetland hydrology, vegetation and soil moisture.

[0132] (2) The present invention integrates dynamic monitoring and health assessment. Through the spatiotemporal convolutional neural network (ST-CNN) combined with the attention mechanism, it can simultaneously process the spatiotemporal characteristics of SAR and optical data, and realize long-term monitoring of the dynamic changes of wetlands. Especially under the dual influence of climate change and human activities, the system can provide timely ecological health assessment.

[0133] (3) The multi-task deep learning network of the present invention designs a multi-task network based on spatiotemporal convolution and attention mechanism, which improves the accuracy of wetland classification and change detection while meeting the output requirements of multi-task results.

[0134] (4) The present invention optimizes the characteristics of alpine wetlands. In view of the unique seasonal changes and surface complexity of alpine wetlands, the SAR polarization decomposition model and the calculation method of the optical wetland index are optimized to make the model more suitable for actual application scenarios.

[0135] (5) This invention provides a scientific basis for the management of wetland reserves and improves the efficiency of ecological restoration; supports the formulation of water resource management and ecological protection policies in alpine regions; and provides data support for the study of wetland carbon sequestration functions and the impact of climate change.

[0136] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The dynamic monitoring and health assessment method of alpine wetlands based on multi-source remote sensing includes the following steps: Step 1: Data collection and preprocessing; Acquire SAR image data, optical image data, auxiliary data and other data, and perform preprocessing respectively; Step 2: Multimodal feature extraction; Step 3: Multimodal data fusion; By adopting a spatiotemporal convolutional neural network, inputting SAR and optical multi-temporal data, and combining it with an attention mechanism, the sensitivity to wetland characteristics is improved; Step 4: Dynamic monitoring; Through the fused feature data, based on random forest or support vector machine classification algorithm, wetlands are divided into permanent wetlands, seasonal wetlands, peatlands and degraded wetlands; through time series data, the dynamic change trend of wetlands, wetland area, vegetation coverage and water body expansion and contraction are analyzed; Step 5: Wetland health assessment.

2. The method for dynamic monitoring and health assessment of alpine wetlands based on multi-source remote sensing according to claim 1 is characterized in that: The data acquisition in step 1 is SAR image data, using Sentinel-1 VV / VH polarization data or ALOS PALSAR data, supporting multi-phase and multi-polarization, focusing on obtaining the scattering characteristics of wetland water and soil moisture; optical images, using Sentinel-2 10m resolution multispectral images or Landsat series 30m resolution, to obtain wetland vegetation coverage, surface water distribution and surface temperature; Auxiliary data: digital elevation model, meteorological precipitation, temperature data, etc.

3. The method for dynamic monitoring and health assessment of alpine wetlands based on multi-source remote sensing according to claim 1 is characterized in that: The SAR feature extraction method in step 2, the scattering feature, based on VV and VH polarization, extracts the high scattering characteristics of wetland water, the medium scattering characteristics of vegetation and the low scattering characteristics of bare land; Texture Features,extract wetland soil moisture and roughness based on gray level co-occurrence matrix; Dynamic characteristics,Analysis of seasonal changes of wetland soil moisture using time series SAR data.

4. The method for dynamic monitoring and health assessment of alpine wetlands based on multi-source remote sensing according to claim 1 is characterized in that: In step 3, a spatiotemporal convolutional neural network is used to input SAR and optical multi-temporal data, and an attention mechanism is combined to improve the sensitivity to wetland characteristics; the network is divided into two, network branch one: processing SAR scattering and texture features; network branch two: processing optical vegetation and water body features, and outputting wetland classification, dynamic change trend diagram and key feature distribution diagram.

5. Dynamic monitoring and health assessment system of alpine wetlands based on SAR and optical remote sensing, including the following systems: Data acquisition and preprocessing module: used for the acquisition and preprocessing of SAR image data, optical data, and auxiliary data; Multimodal feature extraction module: used for SAR feature and optical feature extraction; Multimodal data fusion module: used for the fusion of SAR feature and optical feature data, using spatiotemporal convolutional neural network and sub-network processing technology to achieve the sensitivity of wetland features; Wetland dynamic monitoring module: using the fused feature data, based on random forest or support vector machine classification algorithm, wetlands are divided into permanent wetlands, seasonal wetlands, peatlands and degraded wetlands; through time series data, the dynamic change trend of wetlands, such as wetland area, vegetation coverage and water body expansion and contraction, is analyzed; Output: annual wetland change map, wetland dynamic curve (such as area change curve); Wetland health assessment module: Construct a health assessment model based on weight distribution, AHP hierarchical analysis method, and integrate various indicators into a comprehensive score of wetland ecological health for wetland health assessment; Output and decision support module: Generate wetland health grade zoning map of excellent, good, fair and poor; Provide wetland protection and restoration recommendations, such as the demarcation of protected areas and the prioritization of restoration of degraded areas.

6. The method for dynamic monitoring and health assessment of alpine wetlands based on multi-source remote sensing according to claim 5 is characterized in that: The multi-module data fusion module learning architecture is as follows: 1) Using a spatiotemporal convolutional neural network (ST-CNN) with SAR and optical multi-temporal data as input and an attention mechanism to improve sensitivity to wetland characteristics; 2) Network branch 1: Processing SAR scattering and texture features; 3) Network branch 2: processing optical vegetation and water characteristics; Result output: Output wetland classification, dynamic change trend chart and key feature distribution map.

Images