Large-scale mangrove canopy height accurate inversion method based on multi-source remote sensing images and XGBoost regression model

Abstract

The present invention provides a method for accurate remote sensing inversion of mangrove canopy height based on multi-source remote sensing data and an XGBoost regression model. First, image collection and preprocessing are performed: UAV-LiDAR data of a typical area during low tide is obtained, and the collected UAV-LiDAR data is preprocessed; high-quality SDGSAT-1 images and time-series Sentinel-1 / 2 images are screened, and an initial feature set for mangrove canopy height inversion is constructed based on feature parameters extracted from satellite images; secondly, an optimal feature set for mangrove canopy height inversion is constructed based on an XGBoost regression model, a cross-validation recursive feature elimination method, and a SHAP interpretation framework; finally, the mangrove canopy height is accurately inverted based on the XGBoost regression model and the optimal feature set, and a large-scale spatial distribution map of mangrove canopy height is drawn. The present invention uses remote sensing technology to achieve rapid and accurate inversion of large-scale mangrove canopy height, improves the accuracy and credibility of remote sensing inversion of mangrove canopy height, and has repeatability. The results can provide important information for the restoration and management of mangrove ecosystems.

Description

Technical Field

[0001] The present invention relates to the field of accurate inversion of large-scale mangrove structural parameters, and specifically to a method for rapid and accurate inversion of mangrove canopy height based on multi-source remote sensing data and an XGBoost regression model. Background Art

[0002] As a unique transitional wetland between land and sea, mangrove ecosystems have multiple ecological values, such as resisting storm surges, mitigating sea level rise and coastal erosion, and regulating water quality and nutrient cycles in coastal areas. In addition, mangroves can fix and store carbon in the atmosphere and ocean, reducing the negative impacts of global climate change, and are therefore called "blue carbon ecosystems." Although mangroves only account for 0.5% of the global coastal area, their contribution to the storage of organic carbon in coastal areas is as high as 10% to 15% (24TgCy -1 Canopy height is a necessary parameter for characterizing mangrove biomass, directly reflecting the carbon sequestration potential of mangrove ecosystems and a key factor in determining their carbon storage and sequestration. Therefore, accurately retrieving large-scale mangrove canopy height is of great scientific significance for the precise calculation of mangrove biomass and carbon storage in China, and serves as an important basis for formulating sustainable management policies.

[0003] Mangroves are mostly located in the poorly accessible intertidal zone. Traditional plot-scale canopy height surveys are time-consuming, labor-intensive, and costly, with a limited monitoring range. They are also insufficient to reflect the large-scale spatial distribution patterns of mangrove canopy height. Remote sensing, with its wide coverage, low labor costs, and high spatiotemporal resolution, has become the preferred method for monitoring and assessing mangrove resources. Because remote sensing data cannot simultaneously possess high spatial resolution, high spectral resolution, high temporal resolution, and large-area coverage, multi-source remote sensing data synergy can take into account the temporal, spectral, coverage, and resolution characteristics of imagery. However, current research on inverting mangrove canopy height using multi-source remote sensing data suffers from low spatial resolution, suboptimal model inversion accuracy, and unclear mechanisms for inverting mangrove canopy height, making it difficult to accurately calculate mangrove biomass and carbon stocks.

[0004] In summary, how to fully explore the key remote sensing characteristics of mangrove canopy height based on the complex nonlinear characteristics of large-scale mangroves and relying on high spatial resolution remote sensing data, introduce high-performance interpretable models to construct a remote sensing inversion model suitable for large-scale mangrove canopy height, and improve the inversion accuracy of mangrove canopy height, is an urgent problem to be solved in the research on accurate inversion of mangrove canopy height. Summary of the Invention

[0005] In order to solve the problems of low spatial resolution, unsatisfactory model inversion accuracy, and unclear key features of mangrove canopy height inversion in existing remote sensing technologies when inverting large-scale mangrove canopy height, the present invention provides a method for quickly and accurately inverting large-scale mangrove canopy height based on UAV-LiDAR, SDGSAT-1 and time-series Sentinel-1 / 2 images combined with the XGBoost regression model.

[0006] The technical solution of the present invention for rapidly and accurately inverting large-scale mangrove canopy height based on UAV-LiDAR, SDGSAT-1 and time-series Sentinel-1 / 2 images combined with the XGBoost regression model is:

[0007] Step 1: Collect UAV-LiDAR point cloud data of mangroves in typical areas during low tide, pre-process the collected point cloud data, and use the processed canopy height data as training data and verification data for the mangrove canopy height inversion model;

[0008] Step 2: Screen the SDGSAT-1 images and time-series Sentinel-1 / 2 images at low tide and without clouds in the mangrove area, extract image features and construct the initial feature set for the mangrove canopy height inversion model;

[0009] Step 3: Construct the optimal feature set for mangrove canopy height inversion based on the initial feature set, XGBoost regression model, RFECV method and SHAP interpretation framework;

[0010] Step 4: Based on the training data obtained in step 1, the optimal feature set for mangrove canopy height inversion obtained in step 3, and the XGBoost regression model, a mangrove canopy height inversion model is constructed, and the model parameters are optimized;

[0011] Step 5: Migrate the mangrove canopy height inversion model to other areas and draw a large-scale mangrove canopy height spatial distribution map; based on the verification data obtained in step 1, use the ten-fold crossover method to verify the inversion results and evaluate the model accuracy.

[0012] A further improvement of the present invention is that step one is specifically as follows:

[0013] 1. The point cloud data preprocessing steps include point cloud denoising: removing errors caused by high-order and low-order gross errors; point cloud filtering: classifying the denoised point cloud data into ground points and ground features; and generating digital model products: interpolating the classified ground points using the inverse distance weighted (IDW) algorithm to generate a digital elevation model (DEM) and a digital surface model (DSM). Based on the research spatial scale, the DEM and DSM are exported as raster data with a spatial resolution of 10m. The true value of the mangrove canopy height is calculated based on the DEM and DSM using the following formula:

[0014] H ture =DSM-DEM;

[0015] Among them, H ture is the true value of canopy height;

[0016] 2 Based on the true value of mangrove canopy height, sampling points were selected using uniform distribution random sampling method, 80% of which were used for model training and 20% for model validation.

[0017] A further improvement of the present invention is that step 2 is specifically as follows:

[0018] 1. Extract spectral reflectance of 7 bands as initial features based on SDGSAT-1 images;

[0019] 2. Extract the time series VV polarization and VH polarization based on the time series Sentinel-1 image, and calculate and extract the time series NDI, RVI and RI radar index as the initial features. The calculation formula is as follows;

[0020]

[0021] ③ Based on the time series Sentinel-2 images, the spectral reflectance of 10 bands, including visible light, near infrared and shortwave infrared, is extracted. The time series NDVI, GCI, MVI, LSWI, mNDWI and EVI indices are calculated and extracted as initial features. The calculation formula is as follows:

[0022]

[0023] A further improvement of the present invention is that step three is specifically as follows:

[0024] 1. Input the initial feature set into the XGBoost regression model and the SHAP interpretation framework to obtain the Shapley value of each feature, and delete the features with a Shapley value of 0;

[0025] 2. Combined with RFECV, the feature input model with the lowest Shapley value is deleted in turn to evaluate the model inversion accuracy. When the model inversion accuracy reaches the maximum value and tends to be stable, the feature set output by the model is the optimal feature set;

[0026] For a feature in a feature set, the Shapley value is calculated as follows:

[0027]

[0028] Where N is the feature set, S is any feature subset that does not contain feature i, S is the number of features in set S, υ(S) is the contribution of feature set S to the model prediction, and υ(S∪{i}) is the contribution of feature set S∪{i} containing feature i to the model prediction.

[0029] The large-scale mangrove canopy height precision inversion method based on multi-source remote sensing images and XGBoost regression model provided by the present invention has at least the following technical effects:

[0030] 1. The present invention uses UAV-LiDAR data as model training data, which can match canopy height inversion at multiple spatial scales and compress the workload of 10 days of field sampling to 1 day.

[0031] 2. This invention provides a methodological approach for achieving accurate inversion of large-scale mangrove canopy height. By collaborating with currently readily available multi-source remote sensing data with high spatial resolution, the spatial resolution of large-scale mangrove canopy height inversion results is improved. In combination with the SHAP interpretation framework, the key features of mangrove canopy height inversion are clarified.

[0032] 3. This paper combines the advanced XGBoost regression model to construct a mangrove canopy height inversion model, which improves the model inversion efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following drawings required for use in the implementation will be introduced. The following drawings are used together with the embodiments of the present invention to explain the present invention, but are not intended to limit the present invention. In the drawings:

[0034] Figure 1 This is the UAV-LiDAR point cloud data of the Zhenhai Bay Mangrove Reserve in Jiangmen City, Guangdong Province. The left side is the tiled image and the right side is the cross-sectional image.

[0035] Figure 2 The actual mangrove canopy heights generated based on UAV-LiDAR point cloud data. The left side shows the Gaoqiao Mangrove Reserve in Zhanjiang City, Guangdong Province, and the right side shows the Zhenhaiwan Mangrove Reserve in Jiangmen City, Guangdong Province.

[0036] Figure 3 These are the 20 key features of mangrove canopy height retrieval;

[0037] Figure 4 This is the spatial distribution map of mangrove canopy height in Guangdong Province;

[0038] Figure 5(a) shows the model self-validation results of the accuracy evaluation results of the mangrove canopy height inversion based on the XGBoost regression model;

[0039] Figure 5(b) shows the independent validation data verification results of the accuracy evaluation results of the mangrove canopy height inversion based on the XGBoost regression model. DETAILED DESCRIPTION

[0040] To make the purpose, technical solutions and advantages of the present invention more clear, the following will be combined with the accompanying drawings of the embodiments of the present invention to clearly and completely describe the technical solutions of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0041] This implementation method leverages high-spatial-resolution multi-source remote sensing data, fully exploits key remote sensing features for mangrove canopy height inversion, and introduces a high-performance interpretable model to achieve accurate inversion of large-scale mangrove canopy height.

[0042] Step 1: Collect UAV-LiDAR point cloud data of mangroves in the Gaoqiao Mangrove Nature Reserve in Zhanjiang City, Guangdong Province and the Zhenhaiwan Mangrove Nature Reserve in Jiangmen City, Guangdong Province during the low tide period in 2023. Figure 1 The collected point cloud data is preprocessed. The preprocessing steps include point cloud denoising: removing errors caused by high-order and low-order errors; point cloud filtering: dividing the denoised point cloud data into ground points and ground objects; generating digital model products: based on the classified ground points, interpolation is performed based on the Inverse Distance Weighted (IDW) algorithm to generate digital elevation models (DEM) and digital surface models (DSM). According to the research spatial scale, DEM and DSM are exported to raster data with a spatial resolution of 10m. Figure 2 .

[0043] The true value of mangrove canopy height is calculated based on DEM and DSM. The calculation formula is as follows:

[0044] H ture =DSM-DEM;

[0045] Among them, H ture is the true value of canopy height;

[0046] Based on the true value of mangrove canopy height, sampling points were selected using a uniformly distributed random sampling method, 80% of which were used for model training and 20% for model validation.

[0047] Step 2: Select the SDGSAT-1 image of Guangdong Province at low tide and without clouds in 2023, and extract the spectral reflectance of seven bands as the initial features. The band introduction is shown in Table 1.

[0048] The 2023 time-series Sentinel-1 images of Guangdong Province were screened, and the time-series VV polarization and VH polarization were extracted based on the time-series Sentinel-1 images. The time-series NDI, RVI, and RI radar indices were calculated and extracted as initial features. The calculation formula is as follows;

[0049]

[0050] The 2023 time-series Sentinel-2 images of Guangdong Province were screened. First, low-tide images were selected to obtain the image with the largest average NDVI among all images in each month of 2023. The low-tide images were combined with the s2cloudless declouding algorithm to detect cloudiness. Cloud-obstructed images were declouded, and other cloud-free images of the month were selected to fill in the gaps. Image features were extracted to construct the initial feature set for the mangrove canopy height inversion model.

[0051] Based on the time-series Sentinel-2 imagery, the spectral reflectance of 10 bands, including visible light, near infrared, and shortwave infrared, is extracted as initial features. The band description is shown in Table 1. The time-series NDVI, GCI, MVI, LSWI, mNDWI, and EVI indices are also calculated and extracted as initial features. The calculation formula is as follows:

[0052]

[0053] Table 1 Introduction to Sentinel-1, Sentinel-2 and SDGSAT-1 satellite image bands

[0054]

[0055] Step 3: Construct the optimal feature set for mangrove canopy height inversion based on the initial feature set, XGBoost algorithm, RFECV, and SHAP interpretation framework. Specifically, the initial feature set is input into the XGBoost algorithm and SHAP interpretation framework to obtain the Shapley value of each feature, and features with a Shapley value of 0 are deleted. For the features in the feature set, the Shapley value is calculated as follows:

[0056]

[0057] Where N is the feature set, S is any feature subset that does not contain feature i, S is the number of features in set S, v(S) is the contribution of feature set S to the model prediction, and v(S∪{i}) is the contribution of feature set S∪{i} containing feature i to the model prediction.

[0058] Combined with the RFECV algorithm, the feature input model with the lowest Shapley value is deleted in turn to evaluate the model inversion accuracy. When the model inversion accuracy reaches the maximum value and tends to be stable, the feature set output by the model is the optimal feature set; combined with the SHAP interpretation framework, the top 20 parameters in the optimal feature set are output. Figure 3 .

[0059] Step 4: Based on the training data obtained in step 1, the optimal feature set for mangrove canopy height inversion obtained in step 3, and the XGBoost algorithm, a mangrove canopy height inversion model is constructed, and the model parameters are optimized at the same time; the model parameters are divided into general parameters, Booster parameters, and objective function parameters. The general parameters and target parameters can be kept at the default parameters to ensure the optimization of the model; when adjusting the Booster parameters, step-by-step control adjustment is adopted, and one of the parameters is optimized in turn according to the parameter weight while the others remain unchanged. The order of parameter adjustment and the final result are: max_depth: 6; min_child_weight: 0; gamma: 1.1; subsample: 0.8; colsample_bytree: 0.8; eta: 0.1.

[0060] Step 5: Migrate the mangrove canopy height inversion model to the entire study area and draw the spatial distribution map of mangrove canopy height in Guangdong Province. Figure 4 ; Based on the verification data obtained in step 1, the inverted mangrove canopy height data of Guangdong Province was verified to evaluate the model accuracy, referring to Figure 5(a) and Figure 5(b), where Figure 5(a) is the model self-verification result and Figure 5(b) is the independent verification data verification result.

Claims

1. A large-scale mangrove canopy height inversion method based on multi-source remote sensing images and XGBoost regression model, which includes the following steps: Step 1: Collect UAV-LiDAR point cloud data of mangroves in typical areas during low tide, pre-process the collected point cloud data, and use the processed canopy height data as training data and verification data for the mangrove canopy height inversion model; Step 2: Screen the SDGSAT-1 images and time-series Sentinel-1 / 2 images at low tide and without clouds in the mangrove area, extract image features and construct the initial feature set for the mangrove canopy height inversion model; Step 3: Construct the optimal feature set for mangrove canopy height inversion based on the initial feature set, XGBoost regression model, RFECV and SHAP interpretation framework; Step 4: Based on the training data obtained in step 1, the optimal feature set for mangrove canopy height inversion obtained in step 3, and the XGBoost regression model, a mangrove canopy height inversion model is constructed, and the model parameters are optimized; Step 5: Migrate the mangrove canopy height inversion model to other areas and draw a large-scale mangrove canopy height spatial distribution map; Based on the verification data obtained in step 1, use the ten-fold cross-validation method to verify the inversion results and evaluate the model accuracy; The step 1 is specifically as follows: ① Point cloud data preprocessing steps include point cloud denoising: removing errors caused by high-order and low-order gross errors; point cloud filtering: classifying the denoised point cloud data into ground points and ground features; generating digital model products: interpolating the classified ground points using the inverse distance weighted algorithm to generate digital elevation models and digital surface models; exporting the DEM and DSM to raster data with a spatial resolution of 10m based on the research spatial scale; and calculating the true value of the mangrove canopy height based on the DEM and DSM. The calculation formula is as follows: H ture =DSM-DEM; in, H ture is the true value of canopy height; ② Based on the actual value of mangrove canopy height, uniformly distributed random sampling was used to select sampling points, 80% of which were used for model training and 20% for model validation; The step 2 is specifically as follows: ① Extract spectral reflectance of seven bands as initial features based on SDGSAT-1 images; ② Extract the time-series VV polarization and VH polarization based on the time-series Sentinel-1 image, and calculate and extract the time-series NDI, RVI, and RI radar index as the initial features. The calculation formula is as follows; ③ Based on the time series Sentinel-2 images, the spectral reflectance of 10 bands, including visible light, near infrared and shortwave infrared, is extracted. The time series NDVI, GCI, MVI, LSWI, mNDWI and EVI indices are calculated and extracted as initial features. The calculation formula is as follows: The step three is specifically as follows: ① Input the initial feature set into the XGBoost algorithm and the SHAP interpretation framework to obtain the Shapley value of each feature, and delete the features with a Shapley value of 0; ② Combined with the RFECV algorithm, the feature input model with the lowest Shapley value is deleted in turn to evaluate the model inversion accuracy. When the model inversion accuracy reaches the maximum value and tends to be stable, the feature set input when the model accuracy is the maximum after tending to be stable is the optimal feature set.

2. The large-scale mangrove canopy height accurate inversion method based on multi-source remote sensing images and XGBoost regression model according to claim 1 is characterized in that: In step one, using UAV-LiDAR data as model training data can match canopy height inversion at multiple spatial scales and compress the workload of 10 days of field sampling to 1 day.

3. The large-scale mangrove canopy height accurate inversion method based on multi-source remote sensing images and XGBoost regression model according to claim 1 is characterized in that: In step 2, the low tide time series Sentinel-2 images are selected by screening the images with the largest average NDVI among all images of each month; the cloud-free images are detected based on the low tide images combined with the s2cloudless declouding algorithm to detect whether there are clouds, and the images obscured by clouds are declouded, and other cloud-free images of the month are selected to fill the gaps.

4. The large-scale mangrove canopy height accurate inversion method based on multi-source remote sensing images and XGBoost regression model according to claim 1 is characterized in that: In step 3, for the features in the feature set, the Shapley value is calculated as follows: Where N is the feature set, S is any feature subset that does not contain feature i, |S| is the number of features in set S, υ(S) is the contribution of feature set S to the model prediction, and v(S∪{i}) is the contribution of feature set S∪{i} containing feature i to the model prediction.

5. The large-scale mangrove canopy height accurate inversion method based on multi-source remote sensing images and XGBoost regression model according to claim 1 is characterized in that: In step 4, the parameters and order of model optimization are: max_depth, min_child_weight, gamma, subsample, colsample_bytree, and eta.

Images