Forest aboveground biomass estimation method based on multi-source remote sensing data

Abstract

The invention provides a forest above-ground biomass estimation method based on multi-source remote sensing data, and the method comprises the steps: obtaining multi-source remote sensing data which comprises a GF-7 stereo satellite image, synthetic aperture radar data, a sentinel-2 multispectral image and an ALOS digital elevation model; performing feature extraction based on the multi-source remote sensing data to obtain a feature data set; constructing a biomass estimation model based on the feature data set; and estimating the biomass in the forest through the biomass estimation model. According to the method, the vertical structure information and the spectral characteristics are combined, so that the AGB estimation performance can be remarkably improved. Compared with similar research results, the method provided by the invention can be applied to a larger research area, has fewer direct variables, and has relatively high accuracy in the aspect of AGB estimation.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing data application, and in particular relates to a forest aboveground biomass estimation method based on multi-source remote sensing data. Background Art

[0002] The global forest area is about 4 billion hectares, accounting for about 31% of the earth's land area. It is the most extensive ecosystem in the global terrestrial environment. Forest ecosystems play an irreplaceable role in maintaining regional ecological balance, global carbon balance and mitigating the impact of global climate change. The carbon pool and function of forest vegetation have become the focus of international attention. Among terrestrial ecosystems, global forest aboveground biomass (AGB) accounts for a large part of global carbon sequestration. AGB plays a very important role in global carbon emission monitoring and forest carbon stock estimation. The Global Climate Observing System (GCOS) led by the United Nations Framework Convention on Climate Change (UNFCCC) records AGB as an essential terrestrial carbon variable (ECV). In addition, information on forest AGB and other biophysical structural parameters also contributes to various activities related to carbon accounting and conservation purposes. Accurately estimating biomass distribution information at local, regional and global scales can detect land changes caused by deforestation and degradation. Therefore, estimating AGB is also the basis for carbon cycle research and global climate change research.

[0003] Although biomass estimates using traditional techniques based on field measurements have high accuracy, conducting such field surveys requires considerable time and resources, which limits the number of plots that can be reasonably surveyed, thus affecting the estimation of biomass over large areas. Summary of the invention

[0004] In order to solve the above-mentioned problem, the present invention provides a method for estimating forest aboveground biomass based on multi-source remote sensing data.

[0005] The technical solution of the present invention is:

[0006] A method for estimating forest aboveground biomass based on multi-source remote sensing data, the method comprising:

[0007] Acquiring multi-source remote sensing data, including GF-7 stereo satellite images, synthetic aperture radar data, Sentinel-2 multispectral images, and ALOS digital elevation models;

[0008] Extract features based on multi-source remote sensing data to obtain feature data sets;

[0009] constructing a biomass estimation model based on the characteristic data set;

[0010] The biomass estimation model is used to estimate the biomass in the forest.

[0011] Furthermore, the feature data set includes tree height, radar polarization parameters, spectral reflectivity and vegetation index.

[0012] Furthermore, the tree height is obtained from GF-7 satellite images and digital elevation models, specifically:

[0013] Tree height = DSM GF-7 –DEM ALOS

[0014] Among them, DSM GF-7 is the surface elevation from GF-7, DEM ALOS is a digital elevation model from ALOS.

[0015] Furthermore, the radar polarization parameter is VV polarization, which means that the electromagnetic waves emitted by the radar are polarized in the vertical direction, and the received reflected waves are also polarized in the vertical direction, which is obtained from the synthetic aperture radar data.

[0016] Furthermore, the spectral reflectance and vegetation index are obtained from the Sentinel-2 multispectral image.

[0017] Furthermore, the spectral reflectance is B5, and the vegetation index is NDI45 and NDRE1.

[0018] Furthermore, constructing a biomass estimation model based on the characteristic data set is specifically as follows:

[0019] Train the RF regression model using the feature dataset;

[0020] Then the accuracy of the RF regression model is evaluated; and the trained RF regression model is determined to be a biomass estimation model.

[0021] Furthermore, the accuracy evaluation includes root mean square error RMSE, relative root mean square error rRMSE, mean absolute error MAE, mean absolute percentage error MAPE, discrimination coefficient R 2 , Akaike information criterion AIC and variance inflation factor VIF.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] The present invention combines vertical structural information (tree height) and spectral features (spectral bands and vegetation index) to significantly improve the performance of AGB estimation. Compared with similar research results, the method of the present invention can be applied to a larger study area, has fewer direct variables, and has a relatively high accuracy in AGB estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to illustrate the embodiments of the invention. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.

[0025] Figure 1 A schematic diagram of the process of the present invention is shown;

[0026] Figure 2 A scatter plot of the observed data and estimated values ​​of AGB (including tree height) of the present invention is shown;

[0027] Figure 3 A scatter plot of the observed data and estimated values ​​of AGB (excluding tree height) of the present invention is shown;

[0028] Figure 4 The spatial distribution diagram of the AGB (including tree height) estimation results of the present invention is shown;

[0029] Figure 5 The spatial distribution diagram of the AGB (excluding tree height) estimation results of the present invention is shown. DETAILED DESCRIPTION

[0030] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] Example 1

[0032] The study site is located in Genhe City in northeastern China. It covers an area of ​​about 20,010 square kilometers (120.196°~122.90°E, 50.414°N~52.517°N). It is part of the Greater Khingan Range, which extends roughly from northeast to southwest. Genhe is mainly located on the western slope of the northern section of the Greater Khingan Range, and its landform is low mountains and hills. The terrain is low in the southwest and high in the northeast, with an altitude of 700-1,300 meters and an average altitude of 1,000 meters. The terrain is low in the southwest and high in the northeast, with an altitude of 700-1,300 meters and an average altitude of 1,000 meters. The north is dominated by hilly terraces, the middle is low mountains and hills with gentle slopes, and the south is low mountains and hills with steep slopes. The frost-free period is about 80 to 90 days per year. The average annual precipitation at the study site is 650 to 825 mm. Snowfall and rainfall occur from November to March and June to August, respectively. With a forest area of ​​more than 1.84 million hectares and a coverage rate of 91.92%, Genhe is one of the most important forest areas and timber production bases in China. The forest is mainly dominated by larch, with an area of ​​1.14 million hectares, accounting for 82.77% of the total forest area in the city. The total area of ​​birch (Betula platyphylla, Sukaczev) and poplar (Populus davidians L.) is 290,000 hectares, accounting for 9.15% of the total forest area in the city. As the "coldest city in China", Genhe has extremely rich cold-temperate natural forests, so it is a suitable area for studying forest biomass.

[0033] Satellite data and preprocessing

[0034] GF-7 Stereo Image and Preprocessing

[0035] The GF-7 satellite is China's first civilian sub-meter resolution stereo mapping satellite launched in 2019. It is equipped with a dual-line camera (DLC) that can acquire images of 0.65 meters backward and 0.8 meters forward, integrating mapping and resource survey functions. In addition, it is equipped with a backward multispectral camera (MUX) that can acquire 2.6-meter four-band multispectral images. At present, the application of GF-7 satellites in forestry is still limited. The present invention collects 37 GF-7DLC images and attempts to generate continuous forest height maps using the stereo information provided by the GF-7DLC images and ALOSDEM, aiming to improve the accuracy of forest aboveground biomass estimation.

[0036] The DSM (digital surface model representing the top of the forest canopy) was extracted from the GF-7 stereo satellite images using the OrthoEngine module of Geomatica 2018 software. The main purpose of the OrthoEngine module is to generate a mathematical model that relates the columns and rows of matched pixels to the ground coordinates (X, Y) and elevation (Z) by using rational polynomial coefficients (RPC), ground control points (GCP) and tie points (TP). The quality of the DSM is determined by the number and quality of points extracted from the forward and backward views. The main steps to extract the DSM data are as follows. First, the relative orientation is determined by the coplanarity equation, the elements of which are precisely determined using the rational polynomial coefficients provided in the RPC file distributed with the GF-7 data file to determine the relative position of the stereo images. The module uses the semi-global matching (SGM) algorithm to match the pixels. Secondly, in this invention, the image is divided into four grid areas and the horizontal (XY) GCPs are automatically extracted based on the Google reference image. (h) GCPs are also extracted from ALOSDEM to minimize field work. A total of 325 GCPs were collected to ensure uniform distribution of GCPs, and points with an error threshold of more than 1 pixel were removed. These GCPs play a key role in achieving absolute orientation of the image in the ground survey coordinate system, involving translation, rotation, and scaling adjustments. The panchromatic images of GF-7 were used as a guide to place these images throughout the study area to ensure accurate geolocation. Third, the connection points between adjacent images were automatically collected, and then the points with large errors were manually removed. A total of 3250 connection points were generated by the present invention, and the overall accuracy of the connection points was 0.841 pixels. Fourth, the DSM was extracted and its accuracy was evaluated using 100 points in the Google reference image. The results showed that the average error of the DSM extracted from the forward and backward views was 0.384 meters.

[0037] Sentinel-1 imagery and preprocessing

[0038] The Sentinel-1 satellite consists of two satellites (A satellite and B satellite) and is a synthetic aperture radar (SAR) sensor carrying C band. There are currently four imaging modes, such as interferometric wide scan wave (IW), (WV), ultra-wide scan (EW) and strip map (SM). The ground range detection (GRD) of the IW-1 level product has a very high spatial resolution (10 meters) and radiometric resolution. It contains multi-visual intensity data with associated backscatter coefficients and can be widely used for land AGB inversion. The present invention obtains 10m resolution Sentinel-1 images covering the study area in the same period from the Google Earth Engine (GEE) cloud platform. The images have two polarization types: vertical transmission vertical reception (VV) and vertical transmission horizontal reception (VH). In GEE, the Sentinel-1 images are preprocessed into an analyzable data format, and the boundaries and thermal noise are removed again, as well as radiation and terrain calibration. Radar data are not greatly affected by clouds, so a considerable number of complete images can be obtained every month. However, radar data may be affected by rainfall or high winds, so clear cloud-free images close to the date of field data collection were selected for analysis.

[0039] Sentinel-2 imagery and preprocessing

[0040] The Sentinel-2 satellite is a European wide-scan, high-resolution, multispectral imaging instrument consisting of two satellites (Sentinel-2A and Sentinel-2B) that supports research in vegetation, soil, and water monitoring. The Sentinel-2 satellite carries a multispectral imager with 13 spectral bands. Considering that the study area of ​​the present invention is covered by a large area of ​​coniferous forest and the purpose of the present invention is to obtain the true tree height, the present invention uses GF-7 stereo images and subtracts ALOS DEM data (Formula 1) to extract tree heights, and obtains a spatially continuous tree height map of the study area.

[0041] Tree height (m) = DSM GF-7 –DEM ALOS (1)

[0042] Among them, DSMGF-7 is the surface elevation from GF-7, and DEM is from ALOS. The extracted tree height can provide useful vertical structure information, thereby reducing the data saturation problem and effectively improving the forest AGB estimation.

[0043] Select relevant predictors

[0044] The selection of appropriate variables is crucial for building a biomass estimation model because some variables are weakly correlated with AGB or these variables may be interdependent. The selected variables should be significantly correlated with AGB but not interdependent. In order to obtain effective predictors, correlation analysis was first used to select candidate variables. Pearson correlation coefficient (P) and scatter plots were used to test the correlation between AGB, and then the variables were determined to be accepted for further analysis based on their significance (P<0.05). In addition to the p-value, the variation inflation factor (VIF) generated for each predictor variable was used to minimize multicollinearity in the model. VIF measures the increase in variance of the estimated regression coefficients due to collinearity, indicating how much the variance is increased compared to independent variables that do not have a linear relationship in the model. A VIF of 1 indicates no collinearity, and some studies use VIF<10 to avoid severe multicollinearity between the selected predictors. In general, VIF>10 indicates high collinearity with other predictors in the model, indicating that the regression estimates are unstable. In the field of remote sensing, VIF is used to check multicollinearity in models with independent predictors. To test the collinearity among the selected variables, a variance inflation factor (VIF) of 10 was used as the threshold.

[0045] Establishing AGB estimation model using appropriate variables

[0046] The random forest regression model was developed on the R programming platform. First, feature extraction was performed on the data set, and a total of 35 features were extracted, but not all features are critical for AGB estimation. How to select the predictors with the least number, the strongest predictive power, and help explain the final model is a challenge. In order to explore the applicability of different data sets and their combinations in estimating AGB, the present invention conducted 12 experiments. In order to overcome the difficulty of selecting the least predictive variables to provide the best predictive power, the present invention used the importance ranking method provided by RF to rank the predictive variables. Subsequently, the present invention adopted the backward feature elimination method, starting from all possible predictive variables, and gradually eliminating the least promising variables. Model optimization and comparison used absolute and relative fit, which were determined by the error statistics of the model, including root mean square error (RMSE), relative root mean square error (rRMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), discrimination coefficient (R2), Akaike information criterion (AIC) and variance inflation factor (VIF) and other criteria. The consistency index was used to rank the effectiveness of different RF models. Then the determination coefficient R was selected. 2 The variable subset with the highest, rRMSE, VIF, and lowest AIC values ​​was used to predict AGB.

[0047] In order to understand whether incorporating forest vertical structure information can reduce the saturation problem in AGB estimation, two RF AGB estimation models were established after selecting the best characteristic factor combination. One model incorporates tree height into the optimal combination of variables, and the second model only uses the optimal combination of variables. The parameters of the two models are uniformly set to ntree 500 and mtry 3.

[0048] Model Validation

[0049] The field measured data set (n=83) was randomly divided into 58 / 25 parts for training and validation respectively. The training data set was used to optimize the random forest regression and train the prediction model, and 30% of the validation samples were used to evaluate the goodness of fit of each model, the accuracy and reliability of the prediction model. Previous studies have widely adopted validation techniques such as leave-one-out cross-validation and k-fold cross-validation, using reference data to evaluate model performance. The present invention uses repeated k-fold cross-validation and bootstrapping to evaluate the sensitivity of the model to the selection of training and validation data sets, where the data are randomly assigned and spatially independent. The present invention uses a k-fold cross-validation procedure to test whether there is overfitting by splitting the data 5 times. The present invention also uses 1000 bootstrappings to evaluate the uncertainty of the model.

[0050] In the constructed RF regression model, the AGB estimation accuracy estimated by the combination of VV, B5, NDI45 and NDRE1 features is the highest. In order to explore the application potential of forest vertical structure parameters in AGB models. The present invention adds the tree height variable to the optimal RF model, and then compares and analyzes the two RF models: one RF model includes tree height, VV, B5, NDI45 and NDRE1, marked as "a"; the other RF model does not include tree height, but only includes VV, B5, NDI45 and NDRE1 features, marked as "b". Among the two AGB estimation models, the AGB model marked as "a" (including tree height) has the highest estimation accuracy (R 2 =0.811, root mean square error = 18.159 Mg / ha, rRMSE = 21.257%). The AGB model marked as "b" has a slightly lower estimation accuracy (R 2 =0.776, RMSE = 19.759 Mg / ha, rRMSE = 23.131%). The results show that the forest AGB estimation model including tree height features achieves higher accuracy than the model established using only spectral features, which indicates that tree height information can improve the accuracy of AGB estimation.

[0051] To further analyze the performance of these two modes, Figure 2 The scatter plot shows the AGB results of the RF model including the tree height variable estimation. Figure 3The scatter plots of AGB results without tree height variable estimation in the RF model are shown. Although overestimation in low AGB areas and underestimation in high AGB areas can be found in both scatter plots. However, model "a", which is the RF model including tree height variables, has significantly lower residual values ​​than model "b", especially in low AGB areas, and is more stable across the AGB region. This confirms that incorporating tree height information can greatly improve the accuracy of coniferous forest AGB estimation and reduce data saturation issues. Therefore, the RF regression model including tree height, VV, B5, NDI45, and NDRE1 is the best for estimating forest parameters.

[0052] In terms of evaluation parameters, the radar polarization parameter VV has a better correlation with AGB, and VV polarization is more effective in estimating AGB than VH polarization. This result is consistent with the results of Pham et al., who found that VV polarization performed better in estimating AGB and was more sensitive to the increase of AGB than VH. Although in the present invention, the combination of VV and VH polarization improves R 2 However, when evaluating the importance of variables, after variable screening, the VH variable was excluded, so only VV was selected to participate in the construction of the AGB estimation model.

[0053] The spectral characteristics of plants mainly reflect the physiological and biochemical characteristics of the plants themselves and the forest canopy, including chlorophyll content, water content and the growth status of forest vegetation. In the study of the present invention, only the B5 band variable was selected for the model among the spectral band variables. Through the previous analysis, on the one hand, the B5 band has a strong ability to reflect the chlorophyll content of vegetation and the complexity of forest stand structure; on the other hand, the B5 band has a strong inversion ability. For the calculation and inversion of forest AGB, the B5 parameter is the main indicator. On the other hand, compared with other bands, the B5 band is less affected by soil reflection interference, which helps to better highlight vegetation information and reduce the impact of soil interference on AGB estimation. This can achieve a more accurate estimation of AGB to a large extent.

[0054] Compared with the original spectral information, the vegetation index can better represent the pure vegetation characteristics by synthesizing related spectral signals. Since this synthesis process can effectively reduce spectral noise and enhance pixel contrast, it helps to enhance the prediction ability of AGB. Although NDVI and EVI are still the two most widely used vegetation indices, NDRE1 and NDI45 have surpassed them in estimating coniferous forest AGB. Among the vegetation indices used in the present invention, NDRE1 plays the greatest role in AGB estimation. This result is consistent with the results of Otsu et al., that is, NDRE1 performs better than traditional NDVI. NDRE1 is constructed by the red edge band and the red band, where the red band is sensitive to vegetation coverage and plant growth status, and the red edge band can effectively monitor vegetation health information. In addition, the present invention also found that the NDI45 spectral index is slightly better than NDVI. Through normalized calculation, the light noise is reduced, and the differences between NDRE1, NDI45 and NDVI are further highlighted.

[0055] The tree height variable represents the vertical structural characteristics, and the information it provides is different from the spectral characteristics, which can alleviate the biomass saturation problem in AGB estimation. In the present invention, tree height is considered to have a higher estimation ability for forest AGB. Since the present invention compares models with and without tree height information, incorporating tree height variables into the model can improve the accuracy of AGB estimation. This confirms that the vertical structural characteristics of the forest are closely related to AGB and are not affected by the data saturation problem.

[0056] Topographic factors affect vegetation growth and forest structure through factors such as drainage and slope, which in turn affect tree species composition, forest complexity and vegetation growth rate. These characteristics can provide information on changes in local canopy structure and help obtain more accurate aboveground biomass (AGB) distribution maps. At the same time, biophysical variables can effectively describe vegetation functions and are crucial for biomass estimation. Although many studies have confirmed their importance, the study of the present invention excluded topographic factors and biophysical variables from the Sentinel-2 study because their importance ranking was not high enough. This may be because the topography of the study area of ​​the present invention is not sufficient to affect AGB estimation, and biophysical parameters are obtained through optical image inversion, so factors such as the inversion algorithm and the selection of specific parameters may affect the accuracy of biophysical parameter estimation.

[0057] The present invention establishes an RF regression model to predict the AGB of cold temperate coniferous forests. Among all the estimation models, the combination of tree height, VV, B5, NDI45 and NDRE1 was determined to construct the RF estimation model with the highest accuracy (model marked as "a"), with R 2The value is 0.811 and the RMSE value is 18.159. In addition, in order to evaluate the importance of the vertical structure of tree height in forest biomass estimation, an RF model named "b" was constructed using variables VV, B5, NDI45 and NDRE1. Compared with using a single variable or a single sensor, the combination of tree height, synthetic aperture radar signal and S2 spectral band can provide higher accuracy in estimating medium and high biomass. Combining tree height can estimate biomass greater than 150Mg / ha and effectively reduce the saturation effect in high-density forest areas. This also shows that model "a" is most consistent with ground observation data of coniferous forests in cold regions.

[0058] The present invention uses GF-7 stereo images, synthetic aperture radar (SAR), Sentinel-2 multispectral images and ALOSDEM data to estimate the tree height and AGB potential of the cold temperate coniferous forest in Genhe Forest Farm in Northeast China. The main finding is that by using multiple data types in combination, the AGB of the cold temperate coniferous forest can be estimated with high accuracy. The estimation results of AGB show that the biomass value in the north and south is the highest, ranging from 80 to 165 Mg / ha, and the biomass value in the middle of the study area is lower, less than 80 Mg / ha. The present invention can draw the following conclusions:

[0059] First, combining continuous tree height data with synthetic aperture radar and multispectral imagery can improve the accuracy of local and regional biomass estimates to a higher level than estimates using a single sensor alone. The RF model constructed by combining tree height, VV, B5, NDI45, and NDRE14 variables can significantly improve AGB estimates in densely forested areas and reduce saturation effects. In addition, the resolution of sentinel images (10 meters) alleviates the problem of mixed pixels that often occur in medium spatial resolution data (30 meters, such as Landsat 8), thereby improving the accuracy of biomass estimates.

[0060] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for estimating forest aboveground biomass based on multi-source remote sensing data, characterized in that: The method comprises: Acquiring multi-source remote sensing data, including GF-7 stereo satellite images, synthetic aperture radar data, Sentinel-2 multispectral images, and ALOS digital elevation models; Extract features based on multi-source remote sensing data to obtain feature data sets; constructing a biomass estimation model based on the characteristic data set; The biomass estimation model is used to estimate the biomass in the forest.

2. According to the method for estimating forest aboveground biomass based on multi-source remote sensing data according to claim 1, the feature data set includes tree height, radar polarization parameters, spectral reflectance and vegetation index.

3. According to the method for estimating forest aboveground biomass based on multi-source remote sensing data as claimed in claim 2, the tree height is obtained by GF-7 satellite images and digital elevation models, specifically: Tree height = DSM GF-7 –DEM ALOS in, DSM GF-7 is the surface elevation from GF-7, DEM ALOS is a digital elevation model from ALOS.

4. According to the forest aboveground biomass estimation method based on multi-source remote sensing data as described in claim 2, the radar polarization parameter is VV polarization, which means that the electromagnetic waves emitted by the radar are polarized in the vertical direction, and the received reflected waves are also polarized in the vertical direction, which is obtained from the synthetic aperture radar data.

5. According to the method for estimating forest aboveground biomass based on multi-source remote sensing data as described in claim 2, the spectral reflectance and vegetation index are obtained from the Sentinel-2 multispectral image. 6 . The method for estimating aboveground forest biomass based on multi-source remote sensing data according to claim 5 , wherein the spectral reflectance is B5, and the vegetation index is NDI45 and NDRE1.

7. According to the method for estimating forest aboveground biomass based on multi-source remote sensing data as claimed in claim 5, constructing a biomass estimation model based on the characteristic data set is specifically as follows: Train the RF regression model using the feature dataset; Then the accuracy of the RF regression model was evaluated; The trained RF regression model was determined as the biomass estimation model.

8. The method for estimating forest aboveground biomass based on multi-source remote sensing data according to claim 7, wherein the accuracy evaluation includes root mean square error RMSE, relative root mean square error rRMSE, mean absolute error MAE, mean absolute percentage error MAPE, discrimination coefficient R 2 , Akaike information criterion AIC and variance inflation factor VIF.

Images