Space-air-ground coordinated coastal wetland suaeda salsa restoration effect evaluation method

Abstract

The invention relates to the technical field of coastal wetland ecological restoration evaluation, and discloses an air-space-ground collaborative coastal wetland suaeda salsa restoration effect evaluation method. In order to solve the problems that ground actual measurement samples are deficient due to poor reachability of a coastal intertidal zone, conventional satellite image mixed pixel interference is serious due to short and small suaeda salsa plants and broken distribution, and a precise quantitative evaluation means is lacked in restoration effect, a ground measurement-multispectral unmanned aerial vehicle-hyperspectral satellite collaborative evaluation framework is constructed. Firstly, a high-resolution image of an unmanned aerial vehicle is used as a bridge, and the problems of scale mismatching and sample expansion between centimeter-level quadrat and ten-meter-level satellite pixels are solved; secondly, performing spatial-spectral fusion on the satellite-borne hyperspectral image, introducing a suaeda salsa hyperspectral index (SSHI) to determine suaeda salsa distribution, further extracting spectrum and texture features of the image, and screening sensitive feature combinations by adopting correlation-VIF to suppress high-dimensional data redundancy; and finally, constructing a regression model to realize biomass time sequence mapping, and quantitatively evaluating the cumulative effect of the ecological restoration project. According to the method, the estimation precision of the biomass of the single species in the complex environment is effectively improved, and scientific support is provided for coastal wetland restoration effect evaluation.

Description

Technical Field

[0001] This invention relates to the field of coastal wetland ecological restoration assessment technology, specifically to a method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach. Background Technology

[0002] Coastal wetland ecosystems play a crucial role in global carbon cycling and biodiversity conservation. As a pioneer species in coastal wetlands, *Suaeda salsa* possesses important functions in salt tolerance, shoreline protection, and carbon sequestration, and the resulting "red beach" landscape has extremely high ecological and economic value. In recent years, *Suaeda salsa* has suffered severe degradation due to the invasion of *Spartina alterniflora* and human activities. With the implementation of a series of national ecological restoration projects, accurately assessing the restoration effectiveness and quantifying the restoration gains are particularly important.

[0003] Aboveground biomass (AGB) is a key indicator for measuring vegetation health and ecosystem function. Compared to traditional field quadrat surveys, satellite remote sensing technology can achieve large-scale spatial mapping of AGB and has been widely used in monitoring various ecosystems. However, most satellite remote sensing estimation methods heavily rely on a wide and evenly distributed range of ground samples to ensure inversion accuracy.

[0004] Currently, the assessment and restoration of Suaeda salsa in coastal salt marshes faces extremely severe challenges. First, due to the complex tidal hydrological environment and poor accessibility of the intertidal zone, obtaining sufficient and representative ground-based samples is extremely difficult, resulting in limited accuracy of satellite inversion models due to insufficient training samples. Second, low-to-medium resolution satellite imagery is prone to pixel mixing effects, limiting species-level biomass retrieval. This problem is particularly prominent in coastal salt marshes because Suaeda salsa typically has small plants (only 20-50 cm tall), sparse canopy cover, and strong spatial heterogeneity, making it difficult for traditional multispectral satellites to capture its subtle features and achieve accurate estimation. In particular, relying solely on spectral features often only captures differences in spectral reflectance, neglecting the spatial texture details crucial for Suaeda salsa biomass estimation. Because Suaeda salsa exhibits a fragmented and non-uniform spatial distribution, a single spectral dimension is insufficient to fully explore the nonlinear mapping relationship between its canopy structure and biomass, making it difficult to guarantee the inversion accuracy and robustness of the model when dealing with complex spatially heterogeneous regions.

[0005] To address the limitation of insufficient ground samples, recent studies have shown that high-resolution UAV imagery can be used to estimate species-level biomass (AGB) and supplement ground-based data. While this method has been successfully applied in grassland and shrub ecosystems, research on species-level AGB estimation using UAVs in severely fragmented and hydrologically dynamic coastal wetlands remains in the exploratory stage. Furthermore, although evidence suggests that hyperspectral imagery outperforms multispectral data in coastal wetland mapping and parameter inversion, its potential for estimating Suaeda salsa biomass has not yet been systematically assessed and explored.

[0006] While existing research has incorporated machine learning algorithms such as random forests and support vector machines to construct biomass inversion models, these data-driven methods are typically highly dependent on the quantity and quality of training samples. Given the severe scarcity of measured samples in coastal wetlands, directly applying these algorithms often fails to achieve robust generalization capabilities. Furthermore, existing studies often directly input all bands or vegetation indices into the model, lacking effective feature selection mechanisms for high-dimensional data. This not only increases computational redundancy but also obscures the true contribution of key biophysical parameters to biomass, resulting in poor physical interpretability.

[0007] In summary, existing monitoring methods largely rely solely on UAV or satellite data, lacking a framework that organically combines the two. From an evaluation perspective, current methods primarily focus on monitoring "area increase / decrease" based on satellite imagery, failing to deeply characterize the recovery level of growth quality within the community. Particularly for the restoration of Suaeda salsa in saline-alkali land, aboveground biomass, as a core indicator for measuring ecological health and carbon sequestration potential, is crucial for the accurate estimation of restoration effectiveness and performance evaluation. However, at the technical implementation level, limited by the small and fragmented physical characteristics of Suaeda salsa, there is currently a lack of systematic technical solutions for utilizing UAVs as a scale conversion bridge to expand surface samples and combining the spectral advantages of satellite-borne hyperspectral imagery to alleviate the constraints of mixed pixel effects on inversion accuracy. Therefore, there is an urgent need to develop a space-air-ground collaborative biomass inversion and quantitative evaluation method to achieve accurate quantitative assessment of the restoration effectiveness of a single species in the complex environment of coastal wetlands, thereby improving the accuracy and reliability of monitoring. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] The purpose of this invention is to provide a method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach, primarily aimed at solving the following technical problems existing in the prior art:

[0010] (1) Due to the harsh environment and poor accessibility of the coastal intertidal zone, the number of measured biomass samples on the ground is insufficient and the spatial distribution is uneven, which in turn limits the generalization ability and accuracy of the large-scale remote sensing inversion model.

[0011] (2) Because the plants of Suaeda salsa are short, the canopy is sparse and the spatial heterogeneity is strong, traditional multispectral satellites are severely affected by mixed pixel effects and background noise, resulting in poor biomass estimation accuracy.

[0012] (3) There is a lack of quantitative evaluation methods for the quality and effectiveness of restoration of Suaeda salsa in saline land, making it difficult to scientifically characterize the phased gains of restoration projects in improving the quality of community growth.

[0013] (II) Technical Solution

[0014] To achieve the above objectives, the present invention provides the following technical solution:

[0015] A method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach includes the following steps:

[0016] 1. A method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach, characterized by the following steps:

[0017] S1. Construct a multi-source dataset from air, space, and ground: During the vigorous growth period of Suaeda salsa in coastal salt marsh wetlands, simultaneously acquire ground biomass data measured by ground plots in the study area, high-resolution multispectral images from UAVs, and satellite remote sensing images containing hyperspectral and multispectral data.

[0018] S2. UAV-scale feature aggregation and sample expansion: Based on high-resolution UAV surface reflectance images, the spectral index (SSSI) of Suaeda salsa is calculated and pure pixels are extracted according to the threshold. Then, the spectral features are obtained using vegetation indices. The multispectral features of the high-resolution UAV images are resampled to the satellite pixel scale using spatial aggregation methods. A biomass estimation model is constructed using the aggregated features and measured ground biomass data. This model is then applied to the aggregated UAV image features to generate ground biomass "true value" labels at the satellite scale, thereby obtaining a satellite-scale sample set containing rich biomass gradients. This sample set is then divided into a training set and a validation set.

[0019] S3. Hyperspectral satellite image spatial-spectral fusion processing: Radiometric calibration and atmospheric correction are performed on the acquired spaceborne hyperspectral images, and spatial-spectral fusion is performed with high-resolution multispectral satellite images acquired at the same time to generate a fused image with both high spatial resolution and high spectral resolution.

[0020] S4. Multidimensional feature extraction and feature optimization: Based on the fused hyperspectral image, the spectral index (SSHI) of Suaeda salsa is calculated and the distribution area of ​​Suaeda salsa in saline land is determined according to the threshold. Spectral and texture features are extracted for this area to construct an initial feature set. A screening strategy including correlation analysis and variance inflation factor (VIF) test is adopted to select the feature combination that is highly sensitive to the biomass of Suaeda salsa in saline land and has low redundancy.

[0021] S5. Construction and mapping of satellite-scale biomass inversion model: Using the selected sensitive feature combination as the independent variable and the large satellite-scale biomass training sample set expanded in step S2 as the dependent variable, a biomass inversion model of Suaeda salsa aboveground in coastal salt marsh wetland is constructed. The model is then applied to the hyperspectral fusion image of the entire study area to realize the estimation and mapping of Suaeda salsa biomass on a large scale.

[0022] S6. Accuracy Assessment: The accuracy of the biomass estimation results is verified using a UAV-based validation dataset and ground-based measured data.

[0023] S7. Quantitative assessment of ecological restoration effectiveness: Using the model constructed in step S5, multi-year hyperspectral fusion images of the study area are processed to obtain multi-temporal spatial distribution maps of Suaeda salsa aboveground biomass, in order to assess the restoration effectiveness of Suaeda salsa.

[0024] 2. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach according to claim 1, characterized in that, in step S2, the UAV-scale feature aggregation and sample expansion specifically include:

[0025] S2.1 Calculate the spectral index (SSSI) of Suaeda salsa in UAV images, set a threshold to extract pure Suaeda salsa pixels, and exclude interference from water bodies, bare beaches and Spartina alterniflora;

[0026] S2.2 Construct a multispectral vegetation index feature set for the Suaeda salsa pixels within the mask, calculate various multispectral vegetation indices related to biomass, and use the mean method to aggregate the high spatial resolution vegetation index raster space into a grid consistent with the spatial resolution of satellite pixels.

[0027] S2.3. Use Pearson correlation coefficient to screen and obtain key features that are significantly correlated with ground-measured biomass;

[0028] S2.4. Use the Random Forest Regression (RFR) algorithm to construct a mapping model between ground-measured biomass and aggregated key features. Apply this model to all aggregated UAV image grids to generate sufficient satellite-scale biomass samples as "true value" labels for subsequent satellite inversion.

[0029] 3. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach according to claim 1, characterized in that, in step S3, the spatial-spectral fusion processing specifically involves: using the CN Spectral Sharpening fusion algorithm to fuse the low spatial resolution hyperspectral data with the high spatial resolution multispectral data to generate high spatial resolution hyperspectral data; the fused image needs to be evaluated by root mean square error (RMSE), peak signal-to-noise ratio (PSNR), relative global dimension synthesis error (ERGAS), and spectral angle mapping (SAM) to ensure spectral fidelity.

[0030] 4. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach according to claim 1, characterized in that, in step S4, the multidimensional feature extraction includes:

[0031] S4.1 Extraction of Suaeda salsa targets: Calculate the Suaeda salsa hyperspectral index (SSHI) of the fused image, binarize the image according to the preset threshold, and extract the spatial distribution mask of Suaeda salsa.

[0032] S4.2 Spectral Feature Extraction: Within the spatial distribution mask, calculate the conventional multispectral vegetation index, red edge index, hyperspectral vegetation index, and spectral transformation features based on reciprocal, logarithm, square root, and their derivatives;

[0033] S4.3 Texture Feature Extraction: Select key bands that are sensitive to biomass and use the gray-level co-occurrence matrix to extract the mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation texture statistics.

[0034] 5. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach according to claim 1, characterized in that, in step S4, the specific steps for feature optimization are as follows: in the first stage, the Pearson correlation coefficient (r) between each feature and biomass is calculated, and spectral and texture features with |r| greater than a preset threshold are retained; in the second stage, the VIF of the selected features is calculated, and features with VIF>10 are removed to eliminate multicollinearity, and finally the optimal combination of sensitive features is determined.

[0035] 6. The method according to claim 5, wherein the optimal sensitive feature combination includes at least one or more of the following features: conventional vegetation index, hyperspectral index sensitive to Suaeda salsa, red edge index, and spectral features such as bands processed by mathematical transformations such as reciprocal, logarithm, and differential, as well as texture features based on the gray-level co-occurrence matrix to extract key bands.

[0036] 7. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a space-air-ground integrated approach according to claim 1, characterized in that, in step S5, the biomass inversion model construction specifically involves: using RFR, with the optimal sensitive feature combination determined in step S4 as the input variable, and the expanded sample set generated in step S2 as the target variable for training; in the model parameter settings, the number of decision trees and the maximum tree depth are optimized through grid search, and the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) are used as evaluation indicators to establish the final estimation model.

[0037] 8. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach according to claim 1, characterized in that, in step S6, the accuracy assessment specifically includes: verifying the satellite estimation results using UAV verification samples and ground quadrat data, and calculating R², RMSE and MAE.

[0038] 9. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach according to claim 1, characterized in that, in step S7, the quantitative evaluation of the ecological restoration effect specifically includes: obtaining temporal dynamic data of aboveground biomass by statistically estimating the aboveground biomass of Suaeda salsa in the study area over several years, thereby quantitatively evaluating the restoration effect of the coastal wetland ecological restoration project.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) This invention constructs a sample expansion mechanism of "pure pixel filtering-spatial aggregation". Utilizing the high resolution advantage of UAV imagery, spectral threshold screening is performed to eliminate background information interference, and only the spectral features of pure Suaeda salsa pixels are aggregated to the satellite scale. This step effectively overcomes the mixed pixel effect caused by background noise during the traditional scale conversion process, thereby constructing a satellite-scale training sample set with high spectral purity and wide spatial coverage, and realizing high-quality expansion of the ground-measured "true value".

[0041] (2) This invention fully leverages the high spectral resolution of spaceborne hyperspectral images to obtain specific spectral features of Suaeda salsa that are difficult to characterize using conventional multispectral images. By constructing a "correlation-VIF" dual screening mechanism, it effectively suppresses the "curse of dimensionality" in hyperspectral data while eliminating collinearity interference, thus achieving efficient compression of the feature space. This mechanism ensures that the feature subset possesses both strong sensitivity and statistical independence, significantly enhancing the robustness and generalization ability of the inversion model, and providing scientific variable constraints for the accurate quantification of the restoration effectiveness of a single species.

[0042] (3) In view of the characteristics of Suaeda salsa plants such as short stature, sparse canopy and strong spatial heterogeneity, this invention firstly utilizes spatial-spectral fusion technology to effectively improve the spatial resolution of hyperspectral satellite images. Secondly, texture features are introduced during feature extraction to supplement the lack of information on vegetation canopy structure in a single spectral dimension, thereby significantly improving the accuracy of biomass estimation. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the overall technical process of the method of the present invention.

[0045] Figure 2 This is a schematic diagram of the spatial aggregation of UAV images according to the present invention;

[0046] Figure 3 This invention provides a high-precision map of the aboveground biomass of Suaeda salsa based on UAV imagery.

[0047] Figure 4 This is a graph showing the accuracy of aboveground biomass estimation for Suaeda salsa in Embodiment 2 of the present invention, where only spectral features are input.

[0048] Figure 5 This is a map showing the accuracy of the aboveground biomass estimation of Suaeda salsa based on the input spectral and texture features in Example 3 of the present invention.

[0049] Figure 6 This is a spatial distribution map of the aboveground biomass of Suaeda salsa in the study area of ​​Example 3 of the present invention. Detailed Implementation

[0050] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention may be implemented in other embodiments without these specific details.

[0051] Example 1

[0052] Reference Figure 1 A method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach includes the following steps:

[0053] S1. Constructing a multi-source dataset from air, space, and ground: During the peak growth period of Suaeda salsa in coastal salt marshes, ground surveys and remote sensing data acquisition were conducted simultaneously. The specific steps are as follows:

[0054] S1.1 Ground Measured Data: Several large quadrats were established within the study area, with dimensions consistent with the pixel size of subsequent satellite imagery. Within each quadrat, vegetation was harvested, measured, and dried and weighed in the laboratory to obtain the AGB per unit area as the ground "true value".

[0055] S1.2 UAV Data: High-resolution images of the sample plot area are acquired using a UAV equipped with a multispectral camera. The image bands should include at least the visible and near-infrared bands to meet the needs of vegetation index calculation. The images are then radiometrically corrected.

[0056] S1.3 Satellite Data: Acquire satellite-borne hyperspectral and high-resolution multispectral images that are close to the time of the ground survey. The images must cover the entire study area and have sufficient cloud cover to meet the processing requirements. Then, perform atmospheric and radiometric corrections on the images to obtain surface reflectance data.

[0057] S2. UAV-scale feature aggregation and sample expansion: Leveraging the high spatial resolution of UAV data, an intermediate-scale biomass estimation model is constructed to address the problem of insufficient ground-based samples. The specific steps are as follows:

[0058] S2.1, Vegetation Extraction: Calculate the vegetation index (SSSI(1)) of the UAV image, extract pure Suaeda salsa pixels by threshold segmentation, remove interference from water bodies, bare beaches and other vegetation, and obtain Suaeda salsa distribution pixels;

[0059] (1)

[0060] S2.2 Feature Aggregation: Multiple multispectral vegetation indices were calculated within the *Suaeda salsa* region. The mean aggregation method was used to resample the centimeter-high resolution vegetation index raster from the UAV to a grid with the same spatial resolution as satellite pixels.

[0061] S2.3 Sample Expansion: Based on ground-based measured data and aggregated UAV features, key sensitive features were selected using the Pearson correlation coefficient. Subsequently, a UAV-scale biomass estimation model was constructed using RFR. This model was applied to the UAV coverage area to generate satellite-scale biomass raster data. 70% of the pixels were randomly selected as the expanded satellite-scale training sample set, and the remaining 30% served as the validation set.

[0062] S3. Hyperspectral Satellite Image Spatial-Spectral Fusion Processing: To more accurately obtain Suaeda salsa biomass information and minimize interference from complex surface backgrounds, spatial-spectral fusion processing is performed on the image data. The specific steps are as follows:

[0063] S3.1 Preprocessing: Radiometric calibration and atmospheric correction are performed on the acquired spaceborne hyperspectral images to obtain the true surface reflectance;

[0064] S3.2 Registration: Select high spatial resolution multispectral images of the same or different sources, and perform strict geometric correction and registration to ensure the consistency of spatial position;

[0065] S3.3 Fusion: The CN Spectral Sharpening fusion algorithm is adopted to organically fuse the rich spectral information of hyperspectral images with the clear spatial details of multispectral images, generating fused data with both high spatial resolution and high spectral resolution, thereby significantly improving the ability to capture small vegetation patches and the accuracy of biomass inversion.

[0066] S4. Multidimensional Feature Extraction and Feature Optimization: Based on the fused hyperspectral image, a multidimensional feature set is constructed and features are selected. The specific steps are as follows:

[0067] S4.1 Extraction of Suaeda salsa targets: Calculate the Suaeda salsa hyperspectral index (SSHI (2)) of the fused image, and perform binarization processing on the image according to the preset threshold to obtain the spatial distribution of Suaeda salsa.

[0068] (2)

[0069] in,

[0070] S4.2 Feature Extraction: Within the spatial distribution mask, spectral and texture features are extracted. Spectral features include conventional wideband vegetation indices, such as narrowband hyperspectral vegetation indices sensitive to anthocyanins, red-edge indices, and spectral features after mathematical transformations such as reciprocals, logarithms, and derivatives; texture features are extracted based on the gray-level co-occurrence matrix, extracting texture information of key bands.

[0071] S4.3 Feature Optimization: Implement a screening strategy. The first stage uses Pearson correlation coefficient for initial screening; the second stage uses VIF to handle multicollinearity, removing highly correlated redundant features while retaining high-contribution features, thereby obtaining the most concise and efficient feature subset.

[0072] S5. Construction of satellite-scale biomass inversion model: Using the optimal combination of sensitive features determined in step S4 as input variables and the satellite-scale sample set expanded in step S2 as target variables, the RFR model is used to generate a spatial distribution map of biomass in the whole region.

[0073] S6. Accuracy Assessment: Using UAV verification samples and independent ground-based measured data, calculate R², RMSE, and MAE to verify the accuracy of the inversion results at both the pixel and region scales.

[0074] S7. Quantitative assessment of ecological restoration effectiveness: Using the model constructed in step S5, multi-year hyperspectral fusion images of the study area are processed to obtain multi-temporal spatial distribution maps of Suaeda salsa aboveground biomass, in order to assess the restoration effectiveness of Suaeda salsa.

[0075] Example 2

[0076] This embodiment selects the Yellow River Delta National Nature Reserve as the study area and uses spectral characteristics to construct a model to estimate the aboveground biomass of *Suaeda salsa* in this area. (Refer to...) Figures 2-4 This embodiment aims to demonstrate the basic inversion effect with only spectral features as assistance, serving as a control group for method verification.

[0077] S1. Data acquisition and processing, the specific steps are as follows:

[0078] S1.1 Ground Data Acquisition: A field survey was conducted in September 2024 (peak biomass period of Suaeda salsa). Forty-two 10 m × 10 m quadrats were established within the study area, and the center coordinates of each quadrat were recorded using an RTK GNSS receiver. Within each quadrat, 0.5 m × 0.5 m quadrats were established for sampling: for homogeneous patches, three quadrats were randomly selected; for heterogeneous patches, five quadrats were established along the diagonal, resulting in a total of 192 quadrat data points. Aboveground vegetation within the quadrats was harvested, washed, and dried to constant weight in a laboratory oven at 85°C. The average biomass per unit area (AGB) was calculated, and the average biomass of all quadrats within the plot was used as the ground "true value" for that 10 m × 10 m pixel.

[0079] S1.2. UAV Data Acquisition: A DJI Phantom 4 multispectral UAV equipped with a six-band (blue, green, red, red-edge, near-infrared, and visible light) CMOS sensor was used. The flight time was from 10:00 to 14:00 in clear weather in September 2024, synchronized with ground measurements. The flight altitude was set at 30-50 m, and the acquired images had a ground resolution of 2-4 cm, acquiring a total of 27 images covering all ground survey plots. DJI Terra software was used for image stitching and radiometric calibration to generate multispectral surface reflectance images of the study area.

[0080] S1.3 Satellite Data Acquisition: Select the ZY1-02D ​​satellite imagery with an imaging time of October 3, 2024. This imagery contains 30 m resolution hyperspectral camera (AHSI) data and 10 m resolution visible near-infrared camera (VNIC) data.

[0081] S2. Sample expansion and spatial spectrum fusion: The specific steps are as follows:

[0082] S2.1. UAV Sample Enlargement: The Spectral Sediment Index (SSSI) of Suaeda salsa was calculated using the red, green, blue, and near-infrared bands of UAV imagery. Referring to existing research, a threshold of SSSI > 0.035 was set to extract pure Suaeda salsa patches. Fourteen multispectral vegetation indices (RVI, DVI, NDVI, GNDVI, RDVI, EVI, TVI, PVI, SAVI, MSAVI, OSAVI, VARI, ARVI, and SSSI) were calculated for the selected Suaeda salsa pixels. These centimeter-resolution vegetation indices were spatially aggregated into a 10-meter resolution grid consistent with satellite pixels using the mean method. Subsequently, based on the aggregated features and ground-measured data, Pearson correlation coefficients were used for initial screening, and a random forest model was constructed to calculate the SHAP value. Finally, the key features selected were DVI, GNDVI, MSAVI, RVI, SAVI, RDVI, NDVI, TVI, OSAVI, and EVI. An intermediate-scale biomass estimation model was constructed using RFR (R²=0.912 for training set, R²=0.902 for validation set), and this model was applied to all aggregated data to directly obtain a biomass augmentation sample set of more than 3,500 satellite-scale samples.

[0083] S2.2 Satellite Spatial-Spectral Fusion: Missing bands (1358-1425 nm, 1812-1947 nm, 2501 nm) in the ZY1-02D ​​AHSI image were removed, followed by radiometric calibration and atmospheric correction using the FLAASH module. The 30 m resolution AHSI image and the 10 m resolution VNIC image were fused using the CNSpectral Sharpening algorithm to generate a 10 m resolution hyperspectral image. Quality evaluation indicators show that the spectral fidelity of the images is good before and after fusion (RMSE=0.0005, PSNR=34.626, ERGAS=6.533, SAM=1.101).

[0084] S3. Spectral feature extraction and optimization, the specific steps are as follows:

[0085] S3.1 Feature Extraction: Based on the fused image, four types of spectral features were extracted: 14 conventional multispectral vegetation indices, 5 erythrocyte-sensitive hyperspectral indices, 7 red edge indices, and 3600 band features were generated by performing 24 mathematical transformation operations (such as reciprocal, logarithm, first derivative, etc.) on 150 bands.

[0086] S3.2 Feature Optimization: A "correlation analysis-VIF" screening strategy was adopted. First, features with a correlation coefficient |r| > 0.65 were retained; finally, features with VIF > 10 were removed. The final 10 spectral sensitivity features determined are: RVI, SSHI, SIPI, RDVI, ARI2, ARI1, , GNDVI and .

[0087] S4. Model building and accuracy verification, the specific steps are as follows:

[0088] S4.1 Model Parameter Settings: The RFR algorithm is adopted, using 10 selected spectral features as input variables and an expanded sample set as the target variable for training. The model parameters are set as follows: number of decision trees is 1000, and maximum depth is 10.

[0089] S4.2 Result Analysis: The model constructed using only spectral features has a coefficient of determination R² of 0.737, a root mean square error RMSE of 22.443 g / m², and a mean absolute error MAE of 16.148 g / m² on the validation set.

[0090] Example 3

[0091] Reference Figures 2-6 Based on Example 2, this embodiment adds texture features in the feature variable selection stage to improve estimation accuracy, and uses the selected optimal classification model to evaluate the restoration effect of Suaeda salsa in the Yellow River Delta.

[0092] S1. Data acquisition and processing: consistent with step S1 in Example 2.

[0093] S2, Sample expansion and spatial spectrum fusion: consistent with step S2 of Example 2.

[0094] S3. Multidimensional feature extraction and optimization, the specific steps are as follows:

[0095] S3.1 Feature Extraction: Based on the spectral features in Example 2, additional texture features are extracted. Thirty-five key spectral bands are selected (covering the band ranges of 391–598 nm, 658–709 nm, 881–898 nm, 967–984 nm, 1001–1044 nm, and 1972–2002 nm). Using the gray-level co-occurrence matrix, a 3×3 moving window is set with a step size of 1. Eight texture statistics are calculated: mean (MEA), variance (VAR), homogeneity (HOM), contrast (CON), dissimilarity (DIS), entropy (ENT), second moment (SEM), and correlation (COR), generating a total of 280 texture features.

[0096] S3.2 Feature Optimization: The same screening strategy was adopted. For texture features, features with |r| < 0.1 were first eliminated, and VIF < 10 was ensured. The final optimal combination of sensitive features (15 in total) included: 10 spectral features, the same as in Example 2; 5 texture features: , , , and .

[0097] S4. Model building and accuracy improvement verification, the specific steps are as follows:

[0098] S4.1 Model Construction: The same RFR model parameters as in Example 2 are used, and the input variables are increased to the above 15 feature combinations.

[0099] S4.2 Accuracy Improvement Analysis: Compared to Example 2 which only used spectral features (R²=0.737), the introduction of texture features significantly improved the model accuracy. Final validation was performed using independent ground-based measured data (42 quadrats). The predicted values ​​and measured values ​​showed a high degree of agreement, with a coefficient of determination (R²) of 0.891, a root mean square error (RMSE) of 29.75 g / m², and a mean absolute error (MAE) of 25.00 g / m². The results indicate that texture features effectively supplemented the spatial structure information of the Suaeda salsa canopy, demonstrating the crucial role of spatial structure information in the biomass inversion of Suaeda salsa and improving the inversion accuracy.

[0100] S5. Spatiotemporal Dynamics and Restoration Effectiveness Analysis: Statistical analysis of inversion data during the monitoring period revealed that the total biomass of *Suaeda salsa* in the study area significantly increased from 1440.06 tons in 2020 to 3349.48 tons in 2024. The method of this invention accurately quantifies this biomass accumulation process, demonstrating that the "ground-multispectral UAV-hyperspectral satellite" collaborative estimation method can provide reliable data support for ecological restoration projects.

Claims

1. A method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach, characterized in that, Includes the following steps: S1. Construct a multi-source dataset from air, space, and ground: During the vigorous growth period of Suaeda salsa in coastal salt marsh wetlands, simultaneously acquire ground biomass data measured by ground plots in the study area, high-resolution multispectral images from UAVs, and satellite remote sensing images containing hyperspectral and multispectral data. S2. UAV-scale feature aggregation and sample expansion: Based on high-resolution UAV surface reflectance images, the spectral index (SSSI) of Suaeda salsa is calculated and pure pixels are extracted according to the threshold. Then, the spectral features are obtained using vegetation indices. The multispectral features of the high-resolution UAV images are resampled to the satellite pixel scale using spatial aggregation methods. A biomass estimation model is constructed using the aggregated features and measured ground biomass data. This model is then applied to the aggregated UAV image features to generate ground biomass "true value" labels at the satellite scale, thereby obtaining a satellite-scale sample set containing rich biomass gradients. This sample set is then divided into a training set and a validation set. S3. Hyperspectral satellite image spatial-spectral fusion processing: Radiometric calibration and atmospheric correction are performed on the acquired spaceborne hyperspectral images, and spatial-spectral fusion is performed with high-resolution multispectral satellite images acquired at the same time to generate a fused image with both high spatial resolution and high spectral resolution. S4. Multidimensional feature extraction and feature optimization: Based on the fused hyperspectral image, the spectral index (SSHI) of Suaeda salsa is calculated and the distribution area of ​​Suaeda salsa in saline land is determined according to the threshold. Spectral and texture features are extracted for this area to construct an initial feature set. Correlation analysis and variance expansion factor (VIF) are used to screen feature combinations that are highly sensitive to the biomass of Suaeda salsa in saline land and have low redundancy. S5. Construction and mapping of satellite-scale biomass inversion model: Using the selected sensitive feature combination as the independent variable and the large satellite-scale biomass training sample set expanded in step S2 as the dependent variable, a biomass inversion model of Suaeda salsa aboveground in coastal salt marsh wetland is constructed. The model is then applied to the hyperspectral fusion image of the entire study area to realize the estimation and mapping of Suaeda salsa biomass on a large scale. S6. Accuracy Assessment: The accuracy of the biomass estimation results is verified using a UAV-based validation dataset and ground-based measured data. S7. Quantitative assessment of ecological restoration effectiveness: Using the model constructed in step S5, multi-year hyperspectral fusion images of the study area are processed to obtain multi-temporal spatial distribution maps of Suaeda salsa aboveground biomass, in order to assess the restoration effectiveness of Suaeda salsa.

2. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach, as described in claim 1, is characterized in that... In step S2, the UAV-scale feature aggregation and sample augmentation specifically include the following steps: S2.1 Calculate the SSSI of the UAV imagery, set a threshold to extract pure Suaeda salsa pixels, and exclude interference from water bodies, bare beaches and Spartina alterniflora; S2.2 Construct a multispectral vegetation index feature set for the Suaeda salsa pixels within the mask, calculate various multispectral vegetation indices related to biomass, and use the mean method to aggregate the high spatial resolution vegetation index raster space into a grid consistent with the spatial resolution of satellite pixels. S2.

3. Use Pearson correlation coefficient to screen and select key features that are significantly correlated with ground-measured biomass; S2.

4. Use the Random Forest Regression (RFR) algorithm to construct a mapping model between ground-measured biomass and aggregated key features. Apply this model to all aggregated UAV image grids to generate sufficient satellite-scale biomass samples as "true value" labels for subsequent satellite inversion.

3. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach, as described in claim 1, is characterized in that... In step S3, the spatial-spectral fusion process specifically involves: using the CN Spectral Sharpening fusion algorithm to fuse the low spatial resolution hyperspectral data with the high spatial resolution multispectral data to generate high spatial resolution hyperspectral data; the fused image needs to be evaluated by root mean square error (RMSE), peak signal-to-noise ratio (PSNR), relative global dimension synthesis error (ERGAS), and spectral angle mapping (SAM) to ensure spectral fidelity.

4. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach as described in claim 1, characterized in that... In step S4, the multidimensional feature extraction includes: S4.1 Extraction of Suaeda salsa targets: Calculate the SSHI of the fused image, binarize the image according to the preset threshold, and extract the spatial distribution mask of Suaeda salsa. S4.2 Spectral Feature Extraction: Within the spatial distribution mask, calculate the conventional multispectral vegetation index, red edge index, hyperspectral vegetation index, and spectral transformation features based on reciprocal, logarithm, square root, and their derivatives; S4.3 Texture Feature Extraction: Select key bands that are sensitive to biomass and use the gray-level co-occurrence matrix to extract the mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation texture statistics.

5. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach as described in claim 1, characterized in that, In step S4, the specific steps of the selection and optimization are as follows: S5.1, First stage: Calculate the Pearson correlation coefficient (r) between each feature and biomass, and retain spectral and texture features whose |r| is greater than the preset threshold; S5.2 Second stage: Calculate the VIF of the selected features, remove features with VIF>10 to eliminate multicollinearity, and finally determine the optimal combination of sensitive features.

6. The method according to claim 5, characterized in that, The optimal combination of sensitive features includes at least one or more of the following features: conventional vegetation indices, hyperspectral indices sensitive to Suaeda salsa, red edge indices, and spectral features of bands processed by mathematical transformations such as reciprocals, logarithms, and differentials, as well as texture features extracted from key bands based on gray-level co-occurrence matrices.

7. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach as described in claim 1, characterized in that, In step S5, the biomass inversion model is constructed as follows: RFR is used, with the optimal sensitive feature combination determined in step S4 as the input variable and the expanded sample set generated in step S2 as the target variable for training. In the model parameter settings, the number of decision trees and the maximum tree depth are optimized through grid search, and the coefficient of determination (R²), RMSE and mean absolute error (MAE) are used as evaluation indicators to establish the final estimation model.

8. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach as described in claim 1, characterized in that, In step S6, the accuracy assessment specifically includes: verifying the satellite estimation results using UAV verification samples and ground quadrat data, and calculating R², RMSE, and MAE.

9. The method for evaluating the restoration effect of Suaeda salsa in coastal wetlands using a combined air-space-ground approach, as described in claim 1, is characterized in that... In step S7, the quantitative assessment of the ecological restoration effectiveness specifically includes: obtaining temporal dynamic data of aboveground biomass by statistically estimating the aboveground biomass of Suaeda salsa in the study area over several years, thereby quantitatively assessing the restoration effectiveness of the coastal wetland ecological restoration project.

Images