Forest biomass estimation method combining machine learning and spatial statistical analysis

Abstract

The invention discloses a forest biomass estimation method combining machine learning and spatial statistical analysis, and belongs to the technical field of forest management. The method mainly comprises the steps of establishing a sample plot horizontal biomass reference map, establishing a sample plot horizontal biomass estimation map based on multi-source data by utilizing machine learning, determining reasons influencing uncertainty by utilizing a spatial statistical method, and re-estimating the biomass by introducing the reasons influencing the biomass estimation uncertainty. According to the technology, the estimation precision of the regional scale forest biomass is improved, a more accurate basis is provided for forest management decision, and sustainable forestry development is promoted.

Description

Technical Field

[0001] The present invention belongs to the field of forest management technology, and relates to a technology for accurately estimating forest biomass, and more specifically, to a technology for improving the accuracy of subtropical forest biomass estimation at a regional scale. A method for improving the accuracy of subtropical forest biomass mapping is invented based on machine learning and spatial statistical analysis. Background Art

[0002] An integrated approach combining ground surveys, satellite imagery, radar, and machine learning has been shown to be one of the effective ways to improve regional-scale aboveground biomass estimates. By integrating data and technology, the uncertainty of aboveground biomass estimates can be greatly reduced. The diversity and uncertainty of multiple factors and their different impact intensities and directions during subtropical forest vegetation succession lead to heterogeneity in the spatial distribution of aboveground biomass. Large-scale human activities further enhance the heterogeneity of forest aboveground biomass. The spatiotemporal heterogeneity of these factors and the challenges of incorporating characteristic parameters such as stand structure, soil, and topography into aboveground biomass estimates result in similar regional aboveground biomass estimates, but significant local differences within regions. Therefore, there is still a large degree of uncertainty in forest aboveground biomass mapping at a regional scale. The average uncertainty of aboveground biomass mapping at the pixel scale for all continents around the world is about ±30%, while the uncertainty in subtropical regions is about ±33%. At a regional scale, forest heterogeneity and differences in remote sensing image quality lead to increased uncertainty in aboveground biomass estimates, ranging from ±6% to ±53%. Therefore, identifying and improving estimation methods is crucial to obtain more accurate and precise spatial distribution maps of forest aboveground biomass.

[0003] Currently, three data sources are mainly used for forest aboveground biomass mapping, namely ground observations, satellite images, and radar. Each data source has its advantages and disadvantages. Ground observations have high measurement accuracy, can verify the accuracy of aboveground biomass estimation and mapping, and provide auxiliary factors that affect mapping accuracy, such as terrain, soil, and stand structure attributes. However, ground observations require a lot of time, money, and manpower, and only limited points can be measured within the sample plot. In addition, it is difficult to sample large areas due to insufficient sampling density. Only by implementing a well-designed sampling scheme can more representative and comprehensive data collection be achieved, thereby providing reasonable confidence intervals for estimation results and significantly improving accuracy. Satellite images have the advantages of multi-scale, cross-regional, and multi-temporal characteristics in describing surface heterogeneity. They provide relatively easy and inexpensive data collection, as well as reliable indices such as the Normalized Difference Vegetation Index and other characteristic variables. Quantitative relationships can be derived from these indices and then inverted to obtain large-scale aboveground biomass maps. However, satellite images have some problems, such as poor signal penetration into the forest canopy, inability to directly extract forest vertical structure information, and light saturation. Radar has the technical advantage of extracting 3D structural information of vegetation and overcoming the light saturation problem in satellite imagery. However, compared to satellite imagery, radar usually requires higher acquisition cost and processing intensity if it is processed from point clouds rather than from final products such as CHM.

[0004] Only by measuring forest aboveground biomass in the field and establishing a quantitative spatial relationship between ground observations, satellite images, and radar can the accuracy of aboveground biomass mapping be fundamentally improved. Combining machine learning and spatial statistical analysis techniques can give full play to the advantages of the above three data sources, making the algorithm more flexible and unrestricted in variable types. Due to its non-parametric nature, there are no strict requirements on the relationship and distribution between the predictor and response variables, and higher prediction accuracy can be achieved through this method.

[0005] Spatial statistical analysis is a statistical technique that uses attribute information and spatial location to analyze data. Compared with traditional statistical methods, spatial statistical analysis incorporates spatial factors such as geographic location, proximity, and spatial interactivity into the research model, breaking the assumption of independence between samples in traditional statistical methods. Spatial statistical analysis not only considers the correlation between values, but also better understands the causes of spatial heterogeneity, geographic location characteristics, spatial distribution characteristics, and spatial patterns within the research object. Therefore, compared with traditional methods, spatial statistical analysis has obvious advantages. However, although there are many prediction models based on spatial statistical methods in the literature, there are relatively few studies on the estimation of forestry-related variables. Therefore, we hypothesize that by using multi-source data in a coordinated manner, spatial statistical analysis can be used to explore the influence of multiple auxiliary factors and further improve the accuracy of mapping.

[0006] In summary, it is necessary to develop a forest biomass estimation method that combines machine learning and spatial statistical analysis. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a forest biomass estimation method combining machine learning and spatial statistical analysis.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: A forest biomass estimation method combining machine learning and spatial statistical analysis, characterized in that it comprises the following steps: (1) Preprocessing of multi-source data, including ground measurement data, SRTM DEM, Landsat OLI, and airborne lidar data; (2) Use machine learning methods to estimate aboveground biomass from multiple source data and evaluate them to obtain a preliminary predicted spatial distribution map of aboveground biomass; (3) Use spatial statistical analysis methods to identify auxiliary factors that affect the accuracy of aboveground biomass estimation; (4) Use spatial statistical analysis methods to integrate the best data source combination and auxiliary factors to further improve the accuracy of aboveground biomass estimation.

[0009] In step (1), the data preprocessing of ground observation data includes calculating the biomass of the sample plots and obtaining the spatial distribution data of auxiliary factors; the data preprocessing of SRTM DEM includes extracting relevant variables; the data preprocessing of Landsat OLI includes radiation calibration, atmospheric correction, terrain correction, cloud filtering and extracting relevant variables; the data preprocessing of airborne lidar includes noise point removal, point cloud filtering, point cloud data rasterization, elevation normalization and extracting relevant variables.

[0010] In step (2), the biomass of the sample plots extracted from the ground observation data is divided into a training set and a validation set, random forest modeling is performed based on the training set, and the estimation accuracy is verified using the validation set. Then, a preliminary predicted spatial distribution map of the aboveground biomass is obtained by inversion.

[0011] In step (3), the residuals and auxiliary factors of biomass are used as input data, and the geographic detector in the spatial statistical method is used to obtain the auxiliary factors that significantly affect the biomass estimation.

[0012] In step (4), the residual value of the sample plot is used as the main variable, and the main auxiliary factors affecting the aboveground biomass estimation are used as covariates to perform empirical Bayesian regression prediction interpolation, thereby obtaining the regional scale aboveground biomass residual spatial distribution map. The aboveground biomass spatial distribution map of the aboveground forest predicted by the random forest model is added to the aboveground biomass residual spatial distribution map predicted by the empirical Bayesian regression prediction interpolation to obtain the optimized regional scale aboveground biomass spatial distribution map. The optimized regional scale aboveground biomass spatial distribution map is verified using the validation set to obtain the optimized estimation accuracy.

[0013] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses machine learning algorithms to integrate multi-source remote sensing data such as ground observations, satellite images, and radar to estimate biomass, effectively leveraging the advantages of multi-source data; (2) The present invention uses the residuals for empirical Bayesian regression prediction interpolation to correct the preliminary predicted aboveground biomass. Compared with the random forest model that only considers multi-source data variables, the aboveground biomass prediction accuracy is higher; (3) The present invention uses auxiliary factors that affect biomass estimation as covariates for empirical Bayesian regression prediction interpolation, fully considering the relationship between residuals and the spatial distribution of aboveground biomass in large areas with complex terrain; (4) The present invention proposes the idea of ​​combining the advantages of machine learning algorithms and spatial statistical analysis to carry out spatial mapping of aboveground biomass, taking into account the spatial autocorrelation between data, and incorporating the interpolation results of the residuals of the empirical Bayesian regression prediction into the aboveground biomass estimation and mapping strategy, thereby reducing the overfitting phenomenon that occurs when using a single random forest model, especially in areas with high aboveground biomass. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the method of the present invention; Figure 2 The accuracy of the preliminary aboveground biomass estimate; Figure 3 This is the preliminary predicted spatial distribution map of aboveground biomass; Figure 4 is the relationship between the auxiliary factors and the accuracy of aboveground biomass estimation; Figure 5 The accuracy of aboveground biomass estimation after residual optimization; Figure 6 Spatial distribution map of aboveground biomass after residual optimization. Implementation

[0015] The present invention is further described in detail below with reference to examples. Embodiment 1:

[0016] A forest biomass estimation method combining machine learning and spatial statistical analysis includes the following steps: (1) Overview of the pilot area The test area is located in Longyan City, Fujian Province, China (115°50′56′′–117°44′15′′E, 24°22′31′′–26°2′35′′N), with an area of ​​19,028 square kilometers. The area has a subtropical marine monsoon climate, with an average annual temperature of 20.5°C and an average annual precipitation of 1479 mm. Longyan City has a high terrain in the east and a low terrain in the west, sloping from northeast to southwest, with an average altitude of 460 meters. Mountains and hills occupy 94.83% of the total area of ​​the city. About 86% of the test area is covered by forests, with the main tree species being Pinus massoniana, Cunninghamia lanceolata and Bamboo.

[0017] (2) Data preprocessing The preprocessing of the national inventory data included the following steps: applying different allometric equations to calculate the aboveground biomass of individual trees; calculating the aboveground biomass of all individual trees; dividing the sum of the aboveground biomass of all individual trees in the plot by the area of ​​the plot to obtain the measured aboveground biomass at the plot level. In addition, regional-scale auxiliary factors were obtained from the national inventory data, including 9 forest stand structure variables (dominant tree species, forest species, origin, age group, shrub cover, average shrub height, herb cover, average herb height, and canopy density), 4 terrain variables (elevation, slope, aspect, and slope position), and 3 soil attribute variables (soil name, soil layer thickness, and humus layer thickness). The elevation variable was extracted from the SRTM DEM data. The LandsatOLI data was radiometrically calibrated, atmospherically corrected, terrain corrected, and cloud filtered, and then the Green_sum average, EVI, and SWIR2_dissimilarity variables were extracted. The airborne lidar data was subjected to noise point removal, point cloud filtering, point cloud data rasterization, and elevation normalization, and then the H was extracted. mean variable.

[0018] (3) Random Forest Modeling The variables extracted from different data sources were used as independent variables for random forest modeling. The training data were divided to establish a random forest model and predict the preliminary aboveground biomass of each plot. The estimation accuracy of the preliminary aboveground biomass was calculated based on the evaluation indicators, including the coefficient of determination (R 2 ), root mean square error (RMSE), mean square error (MSE). Figure 2 is the result of precision assessment. The random forest model trained with the training set was used to invert the preliminary regional aboveground biomass spatial distribution map, such as Figure 3 shown.

[0019] (4) Identify the main influencing factors The residual is obtained by subtracting the predicted aboveground biomass from the measured aboveground biomass of the sample plot. The residual and auxiliary factors are used as input data to identify the main auxiliary factors affecting the aboveground biomass estimation using the geographic detector. The main auxiliary factors are: dominant tree species, herbaceous cover, and canopy density, such as Figure 4 shown.

[0020] (5) Empirical Bayesian regression prediction interpolation of residuals The residual value of the sample plot is used as the main variable, and the dominant tree species, herbaceous coverage, and canopy density are used as covariates to perform empirical Bayesian regression prediction interpolation. Then, the spatial distribution map of aboveground biomass residuals at the regional scale is obtained. The spatial distribution map of aboveground biomass of aboveground forest predicted by the random forest model is added to the spatial distribution map of aboveground biomass residuals predicted by the empirical Bayesian regression prediction interpolation, and the optimized spatial distribution map of aboveground biomass of aboveground forest at the regional scale is obtained, as shown in Figure 2. Figure 5 The validation set was used to validate the optimized regional-scale aboveground forest biomass spatial distribution map, and the optimized estimation accuracy was obtained, as shown in Figure 6 As shown in Figure 2. From the perspective of verification accuracy, all indicators after combining spatial statistics to correct residuals are better. 2 From 0.734 to 0.766, RMSE decreased by 6.186% and MSE decreased by 11.990%. The predicted values ​​before and after residual optimization were compared with the measured values ​​to generate a scatter plot. The scatter points after residual optimization were closer to the 1:1 standard trend line than those before residual optimization, that is, the predicted values ​​were closer to the measured values. This proves that the introduction of spatial statistics to correct the residuals can reduce the overfitting phenomenon of random models when predicting in areas with dense forests and high aboveground biomass values ​​to a certain extent.

Claims

1. A forest biomass estimation method combining machine learning and spatial statistical analysis, characterized in that: The steps include: (1) Preprocessing of multi-source data, including ground measurement data, SRTM DEM, Landsat OLI, and airborne lidar data; (2) Use machine learning methods to estimate aboveground biomass from multiple source data and evaluate them to obtain a preliminary predicted spatial distribution map of aboveground biomass; (3) Use spatial statistical analysis methods to identify auxiliary factors that affect the accuracy of aboveground biomass estimation; (4) Use spatial statistical analysis methods to integrate the best data source combination and auxiliary factors to further improve the accuracy of aboveground biomass estimation.

2. In step (1), the data preprocessing of ground observation data includes calculating the biomass of the sample plots and obtaining the spatial distribution data of auxiliary factors; the data preprocessing of SRTM DEM includes extracting relevant variables; the data preprocessing of Landsat OLI includes radiation calibration, atmospheric correction, terrain correction, cloud filtering and extracting relevant variables; the data preprocessing of airborne lidar includes noise point removal, point cloud filtering, point cloud data rasterization, elevation normalization and extracting relevant variables.

3. In step (2), the biomass of the sample plots extracted from the ground observation data is divided into a training set and a validation set. Random forest modeling is performed based on the training set, and the estimation accuracy is verified using the validation set. Then, a preliminary predicted spatial distribution map of the aboveground biomass is obtained by inversion.

4. In step (3), the residuals and auxiliary factors of biomass are used as input data, and the geographic detector in the spatial statistical method is used to obtain the auxiliary factors that significantly affect the biomass estimation.

5. In step (4), the residual value of the sample plot is used as the main variable, and the main auxiliary factors affecting the aboveground biomass estimation are used as covariates to perform empirical Bayesian regression prediction interpolation, thereby obtaining the regional scale aboveground biomass residual spatial distribution map. The aboveground biomass spatial distribution map of the aboveground forest predicted by the random forest model is added to the aboveground biomass residual spatial distribution map predicted by the empirical Bayesian regression prediction interpolation to obtain the optimized regional scale aboveground biomass spatial distribution map. The optimized regional scale aboveground biomass spatial distribution map is verified using the validation set to obtain the optimized estimation accuracy.

Images