Forest carbon reserve estimation method and system fusing satellite-borne LiDAR features and three-dimensional enhanced spectral index

Abstract

The invention relates to a forest carbon reserve estimation method and system fusing satellite-borne LiDAR features and a three-dimensional enhanced spectral index. The method comprises the steps that discrete footprint point data and Sentinel-2 optical image data of the full-waveform satellite-borne laser radar are acquired and preprocessed; based on the preprocessed data, a random forest regression algorithm is adopted as a spatial interpolator to carry out linear mapping, and a continuous satellite-borne laser radar feature surface is obtained; the forest canopy height is obtained based on the continuous satellite-borne laser radar feature surface, the vegetation coverage is obtained through calculation based on the preprocessed Sentinel-2 optical image data, a pixel multiplication fusion method is adopted for the forest canopy height and the vegetation coverage, and a three-dimensional enhanced spectral index is obtained; and on the basis of the three-dimensional enhanced spectral index, a forest carbon reserve prediction model is constructed by adopting a relative importance and statistical significance level calculation method, and forest carbon reserve prediction is realized. According to the invention, high-precision forest carbon reserve estimation is realized.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing information processing and application technology, specifically relating to a method and system for estimating forest carbon storage by fusing spaceborne LiDAR features and three-dimensional enhanced spectral indices. Background Technology

[0002] As the largest carbon sink in terrestrial ecosystems, forests are crucial for accurately estimating their carbon storage, which is essential for studying the global carbon cycle and addressing climate change. Traditional plot surveys can provide localized, high-precision forest carbon storage data, but they are costly and time-consuming, failing to meet the demands of modern forestry for efficient and dynamic monitoring. The development of remote sensing technology has provided a new approach for large-scale forest carbon storage monitoring.

[0003] Spaceborne LiDAR technology can directly acquire vertical structural information such as tree height, canopy height profiles, and vertical porosity, gradually becoming a powerful tool for three-dimensional forest structure detection and providing important data support for overcoming the saturation problem of optical remote sensing. Combining LiDAR and optical remote sensing data can significantly improve the accuracy of forest carbon storage prediction, especially in high-carbon storage areas where traditional optical models fail. However, spaceborne LiDAR data is usually sparsely sampled, and how to effectively combine it with full-coverage optical imagery remains a problem that urgently needs further exploration. Currently, common combination strategies include multi-source feature cascaded input and hierarchical modeling, but the performance of different methods varies depending on factors such as region and vegetation type, and a universally applicable fusion paradigm has not yet been formed. Therefore, developing a forest carbon storage estimation method that integrates spaceborne LiDAR features and three-dimensional enhanced spectral indices has great application value. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and system for estimating forest carbon storage by integrating spaceborne LiDAR features and three-dimensional enhanced spectral indices, aiming to achieve high-precision forest carbon storage estimation.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for estimating forest carbon storage that integrates spaceborne LiDAR features and three-dimensional enhanced spectral indices, the method comprising: Step 1: Acquire discrete footprint data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and perform preprocessing; Step 2: Based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, the random forest regression algorithm is used as a spatial interpolator to perform linear mapping to obtain continuous spaceborne lidar feature surfaces. Step 3: Obtain forest canopy height based on continuous spaceborne lidar feature surfaces, calculate vegetation cover based on preprocessed Sentinel-2 optical image data, and use pixel multiplication fusion method to obtain three-dimensional enhanced spectral index for forest canopy height and vegetation cover. Step 4: Based on the three-dimensional enhanced spectral index, a forest carbon storage prediction model is constructed using the relative importance and statistical significance level calculation method to achieve forest carbon storage prediction.

[0006] Preferably, the method for obtaining discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data in step 1 and performing preprocessing includes: Data quality checks, effective signal extraction, echo signal smoothing, waveform data decomposition, and energy and geometric feature extraction are performed on the discrete footprint point data of the full-waveform spaceborne lidar. Radiometric calibration and atmospheric correction were performed on Sentinel-2 optical image data, and spectral and texture features were extracted. Spatial matching was performed between the discrete footprint data of the full-waveform spaceborne lidar and the Sentinel-2 optical image data to obtain the geographical location, energy characteristics, and geometric characteristics of the discrete footprint data of the full-waveform spaceborne lidar, as well as the spectral and texture characteristics of the Sentinel-2 optical image data, in the same coordinate system.

[0007] Preferably, step 2, based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, uses a random forest regression algorithm as a spatial interpolator to perform linear mapping to obtain continuous spaceborne lidar feature surfaces. The random forest regression algorithm is used as a spatial interpolator. The geographical location corresponding to the discrete footprint point data of the full-waveform spaceborne lidar and the spectral and texture features of the Sentinel-2 optical image data are used as input variables. The energy features or geometric features corresponding to the discrete footprint point data of the full-waveform spaceborne lidar are used as target variables. The nonlinear mapping relationship between geographical location, spectral and texture features and energy features or geometric features is established, and the sparse spaceborne lidar footprint point data is transformed into continuous raster data, and the continuous spaceborne lidar feature surface is output.

[0008] Preferably, the method for obtaining forest canopy height based on continuous spaceborne lidar feature surfaces in step 3 includes: ; Where CH represents the forest canopy height. P len The length of the wave crest. Peak first and Peaklast These represent the positions of the first and last valid peaks of the echo signal, respectively.

[0009] Preferably, the method for calculating vegetation cover based on preprocessed Sentinel-2 optical image data in step 3 includes: Normalized Difference Vegetation Index (NDVI) was calculated using bands from Sentinel-2 optical image data. ; Wherein, NDVI is the normalized vegetation index, and NIR and Red represent the spectral reflectance in the near-infrared band and the red band, respectively. Pixels are divided into two based on the statistical distribution characteristics of the normalized vegetation index, and vegetation cover is calculated using a piecewise function: ; Wherein, FVC represents vegetation cover. and These represent the 5th and 95th percentile values ​​of NDVI, respectively.

[0010] Preferably, the method for obtaining the three-dimensional enhanced spectral index in step 3 by using a pixel multiplication fusion method to calculate forest canopy height and vegetation cover includes: ; in, CHVI s For three-dimensional enhanced spectral indices, A pixel Represents pixel area. FVC For vegetation coverage, CH The height of the forest canopy. VI s Spectral and texture features of Sentinel-2 optical image data.

[0011] Preferably, step 4, based on the three-dimensional enhanced spectral index, employs methods for calculating relative importance and statistical significance levels to construct a forest carbon storage prediction model. The method for predicting forest carbon storage includes: The random forest algorithm is used to measure the relative importance of spaceborne lidar features, 3D enhanced spectral index, and forest carbon storage. By randomly shuffling the values ​​of a predicted feature variable in the out-of-bag samples, the percentage mean square error of the model's predictions after shuffling is calculated to obtain the relative importance. ; Here, %IncMSE represents relative importance. Ntrees The total number of trees in the random forest. For the first t The original mean square error of the tree on the out-of-bag samples After the substitution variable, the first t The mean squared error of each tree on samples outside the same bag; Statistical inference of relative importance based on permutation tests is performed by repeatedly and randomly permuting the values ​​of a predictive feature variable in out-of-bag samples, calculating the relative importance under each permutation, and generating a distribution of "relative importance" for each variable under preset conditions. The actual observed relative importance of the variable is compared with the distribution obtained from the permutation, thereby generating a statistical significance level for the relative importance. ; in, pj For variables j Statistical significance level p value, nrep To replace the number of repetitions, The relative importance is calculated based on the original data. For the first r The relative importance of the permutation data calculation I (·) is an indicator function that returns 1 when the condition is true, otherwise returns 0; The feature with the highest relative importance is used as the first modeling variable, and feature variables are introduced step by step. At each step, the feature that minimizes the model error is selected and added to the model. ; in, f k+1 For the first k +1 features to be introduced f As a candidate feature, F For feature set, S k For the currently selected k A feature subset, F \ S k These are the remaining features that have not yet been selected. J This is the error evaluation function; Based on predicted and observed values ​​of forest carbon storage, and with the goal of minimizing model error, the root mean square error is used to evaluate the changes in model error during the introduction of variables. The evaluation results guide the selection of model variables. The evaluation indicators are calculated as follows: ; in, and These are the model predictions. This represents the sample size.

[0012] The present invention also provides a forest carbon storage estimation system that integrates spaceborne LiDAR features and three-dimensional enhanced spectral indices. The system is used to implement the aforementioned method and includes: a data acquisition and processing module, a linear mapping module, a fusion module, and a prediction module. The data acquisition and processing module is used to acquire discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and perform preprocessing. The linear mapping module is used to perform linear mapping based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, using the random forest regression algorithm as a spatial interpolator to obtain continuous spaceborne lidar feature surfaces. The fusion module is used to obtain forest canopy height based on continuous feature surfaces of spaceborne lidar, calculate vegetation cover based on preprocessed Sentinel-2 optical image data, and use pixel multiplication fusion method to obtain three-dimensional enhanced spectral index for forest canopy height and vegetation cover. The prediction module is used to construct a forest carbon storage prediction model based on the three-dimensional enhanced spectral index and employs methods for calculating relative importance and statistical significance levels, thereby enabling the prediction of forest carbon storage.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method and system for estimating forest carbon storage by fusing spaceborne LiDAR features and three-dimensional enhanced spectral indices. First, it acquires and preprocesses discrete footprint data from a full-waveform LiDAR system and Sentinel-2 optical imagery to obtain the geographical location, energy, and geometric features corresponding to the spaceborne LiDAR footprints, as well as the spectral and textural features of Sentinel-2. Using the geographical location information of the spaceborne LiDAR footprints and the global spectral and textural features of Sentinel-2 as input variables, and the corresponding spaceborne LiDAR footprint features as target variables, a random forest regression algorithm is used as a spatial interpolator to transform the sparse spaceborne LiDAR footprint data into continuous raster data, providing a reliable data foundation for subsequent high-precision, spatially continuous forest carbon storage distribution mapping. Forest canopy height values ​​were obtained using spaceborne lidar features, and vegetation cover was calculated based on Sentinel-2 data. A pixel multiplication fusion method was used to physically fuse vegetation cover (representing horizontal density), forest canopy height (representing vertical structure), and spectral and textural features (representing species / physiological state), generating a three-dimensional enhanced spectral index. This provides richer and more physically meaningful input variables for subsequent high-precision forest carbon storage retrieval. Combining a variable importance assessment method with permutation tests, the relative importance and statistical significance of all spaceborne lidar features and the three-dimensional enhanced spectral index with forest carbon storage were calculated. Based on the relative importance ranking and model error changes, the optimal feature subset was gradually selected, ensuring that the constructed model achieved optimal prediction accuracy under given data conditions. The final forest carbon storage prediction model achieves the complementary advantages of spaceborne lidar and optical remote sensing, more accurately reflecting the relationship between forest carbon storage and input features, reducing computational resource consumption during model prediction, and thus enabling large-scale, high-precision continuous spatial distribution mapping of forest carbon storage. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are 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.

[0015] Figure 1 This is a schematic diagram of a forest carbon storage estimation method that integrates spaceborne LiDAR features and three-dimensional enhanced spectral indices according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the three-dimensional enhanced spectral index method according to an embodiment of the present invention; Figure 3The diagram shows the comparison of fitting results of forest carbon storage estimation with different combinations of variables in the embodiments of the present invention. Among them, (a)~(c) are schematic diagrams of the results of Sentinel-2 feature estimation, (d)~(f) are schematic diagrams of the results of Sentinel-2 feature and spaceborne lidar feature estimation respectively, and (g)~(i) are schematic diagrams of the results of three-dimensional enhanced spectral index estimation respectively. Figure 4 This is a schematic diagram illustrating the spatial distribution of forest carbon storage in an embodiment of the present invention. (a), (b), and (c) are spatial distribution maps obtained by constructing a random forest model to estimate forest carbon storage after variable screening using Sentinel-2 features, Sentinel-2 features + spaceborne lidar features, and three-dimensional enhanced spectral index, respectively. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This invention provides a method for estimating forest carbon storage by fusing spaceborne LiDAR features and three-dimensional enhanced spectral indices, including: Step 1: Acquire discrete footprint data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and perform preprocessing; Step 2: Based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, the random forest regression algorithm is used as a spatial interpolator to perform linear mapping to obtain continuous spaceborne lidar feature surfaces. Step 3: Obtain forest canopy height based on continuous spaceborne lidar feature surfaces, calculate vegetation cover based on preprocessed Sentinel-2 optical image data, and use pixel multiplication fusion method to obtain three-dimensional enhanced spectral index for forest canopy height and vegetation cover. Step 4: Based on the three-dimensional enhanced spectral index, a forest carbon storage prediction model is constructed using the relative importance and statistical significance level calculation method to achieve forest carbon storage prediction.

[0019] like Figure 1 As shown, the specific implementation process of the present invention is as follows: Step 1 involves acquiring discrete footprint data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, and performing preprocessing. The discrete footprint data of the full-waveform spaceborne lidar and Sentinel-2 optical image data are acquired. The discrete footprint data and Sentinel-2 optical image data are preprocessed to obtain the geographical location information, energy and geometric features corresponding to the discrete footprint data (spaceborne lidar footprints) and the spectral and texture features of the Sentinel-2 optical image data.

[0020] Optical remote sensing (such as Sentinel-2) technology, with its advantages of wide coverage, short revisit period, and rich multispectral information, has become an important tool for large-scale forest carbon storage monitoring. Spaceborne lidar (LiDAR) technology, by actively emitting laser pulses, can directly and accurately acquire three-dimensional vertical structure information of forests, effectively overcoming the shortcomings of optical remote sensing in structural parameter detection. Combining the vertical structure features acquired by spaceborne lidar with the spectral and textural information provided by optical remote sensing can significantly improve the accuracy of forest carbon storage prediction, providing a reliable technical approach for achieving accurate, large-scale forest carbon storage assessment.

[0021] Specifically as follows: Data quality checks, effective signal extraction, echo signal smoothing, waveform data decomposition, and energy and geometric feature extraction are performed on the discrete footprint point data of the full-waveform spaceborne lidar. Radiometric calibration and atmospheric correction were performed on Sentinel-2 optical image data, and spectral and texture features were extracted. Spatial matching was performed between the discrete footprint data of the full-waveform spaceborne lidar and the Sentinel-2 optical image data to obtain the geographical location, energy characteristics, and geometric characteristics of the discrete footprint data of the full-waveform spaceborne lidar, as well as the spectral and texture characteristics of the Sentinel-2 optical image data, in the same coordinate system.

[0022] Specifically, the energy and geometric features corresponding to the footprint points of the spaceborne lidar are extracted, including: The energy characteristics specifically include the echo waveform quantile height, which refers to the height value corresponding to the point where the accumulated energy reaches a specific percentage of the total energy in an echo waveform of a full-waveform lidar. ; in, H q The first echo waveform in the energy characteristics q Quantile height, t 0 and t1 represents the start and end times of the echo signal in the energy characteristics. t q The cumulative signal energy in the energy characteristics reaches the total energy. q % of the time point, p ( t ) represents the amplitude of the echo signal in the energy characteristics.

[0023] Geometric features specifically include waveform length W len crest length P len Waveform leading edge length W ll and waveform trailing edge length W tl : ; ; ; ; in, Peak first and Peak last These represent the positions of the first and last valid peaks of the echo signal, respectively. t first and t last These represent the start and end positions of the echo signal, respectively.

[0024] Furthermore, step 2, based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, uses a random forest regression algorithm as a spatial interpolator to perform linear mapping to obtain continuous spaceborne lidar feature surfaces. The method includes: By utilizing the geographical location information of spaceborne lidar footprints and the global spectral and texture features of Sentinel-2 optical image data, a random forest regression algorithm is used as a spatial interpolator to automatically establish a nonlinear mapping relationship between the geographical location of spaceborne lidar footprints, the global spectral and texture features of Sentinel-2, and the features of spaceborne lidar footprints. This allows for the prediction and generation of continuous spaceborne lidar feature surfaces.

[0025] Traditional spatial interpolation methods are typically based on linear or stationary assumptions, limiting their applicability in complex, non-stationary forest environments. This implementation uses a random forest regression algorithm as the spatial interpolator. It takes the geographic location information of spaceborne lidar footprints and global spectral and texture features extracted from Sentinel-2 optical imagery as input variables, and the energy or geometric features of the corresponding spaceborne lidar footprints as target variables. It automatically establishes a nonlinear mapping relationship between geospatial coordinates, the spectral and texture features of Sentinel-2 optical imagery, and the features of spaceborne lidar footprints, thereby generating a spatially continuous spaceborne lidar feature surface. This implementation effectively solves the problem of the spatially discrete distribution of spaceborne lidar footprints, transforming sparse point data into continuous raster data, providing a reliable data foundation for subsequent high-precision, spatially continuous forest carbon storage distribution mapping.

[0026] Specifically, spectral and texture feature information of Sentinel-2 optical image data is extracted, including: Spectral characteristics specifically include: Single-band reflectivity B i The ratio of the two bands to the vegetation index SR ij : ; ; in, Band i and Band j Representing the Sentinel-2 thirteenth generation respectively i Bands and the j Spectral reflectance of the band.

[0027] Enhanced vegetation index (EVI), normalized difference vegetation index (NDVI), soil-modified vegetation index (SAVI), atmospheric resistance vegetation index (ARVI), red-edged normalized difference vegetation index (NDVIre), red-edged chlorophyll index (IRECI), and red-edged pigment ratio index (CIRE): ; ; ; ; ; ; ; Among them, NIR, Red, Blue, RedE1, RedE2, RedE3 and SWIR1 represent the spectral reflectance of the near-infrared band, red band, blue band, red edge 1 band, red edge 2 band, red edge 3 band and short-wave infrared 1 band, respectively.

[0028] Texture features specifically include: mean (ME), variance (VA), homogeneity (HO), contrast (CT), dissimilarity (DI), entropy (EN), second moment (SM), and correlation (CO). ; ; ; ; ; ; ; ; in, P (i,j) This represents the normalized gray-level co-occurrence matrix, where each element represents the gray level at a specific direction and distance. i The pixels and gray levels are j The probability of pixels appearing simultaneously; N The maximum number of gray levels in the image; , and , The gray levels are respectively i The pixels and gray levels are j The mean and standard deviation of the pixels.

[0029] Specifically, utilizing the geographical location information of spaceborne lidar footprint points and the global spectral and texture features of Sentinel-2, a random forest regression algorithm is used as a spatial interpolator to automatically establish a nonlinear mapping relationship between this information and the features of spaceborne lidar footprint points. This allows for the prediction and generation of continuous spaceborne lidar feature surfaces, including: Using geographic location information and spectral and texture features extracted from Sentinel-2 imagery as input variables, and energy or geometric features of corresponding spaceborne lidar footprint points as target variables, a random forest regression algorithm is used as a spatial interpolator to establish a nonlinear mapping relationship between spatial coordinates, Sentinel-2 features and spaceborne lidar footprint point features. This transforms sparse spaceborne lidar footprint point data into continuous raster data and outputs continuous spaceborne lidar feature surfaces.

[0030] Furthermore, Figure 2 The flowchart of the three-dimensional enhanced spectral index method in step 3 is presented. First, the peak length is obtained as the forest canopy height of the footprint point based on the decomposition results of the satellite lidar waveform data. Then, the forest canopy height value corresponding to each pixel in the target area is obtained by using the generated output continuous satellite lidar peak length features. The normalized vegetation index of each pixel is calculated based on Sentinel-2 data, and the vegetation cover is calculated using a pixel bisection model. The obtained forest canopy height and vegetation cover are fused with the spectral and texture features extracted from the Sentinel-2 data to generate the three-dimensional enhanced spectral index.

[0031] Specifically, the method for obtaining forest canopy height based on continuous spaceborne lidar feature surfaces in step 3 includes: Wave crest length reflects the propagation process of a laser signal as it passes through the ground and vegetation, and is often used to estimate tree height or canopy height in forests. A continuous wave crest length feature surface from continuous spaceborne lidar feature surfaces obtained from spaceborne lidar data is used as a feature of forest canopy height. ; Where CH represents the forest canopy height. P len The length of the wave crest. Peak first and Peak last These represent the positions of the first and last valid peaks of the echo signal, respectively.

[0032] Specifically, the method for calculating vegetation cover based on preprocessed Sentinel-2 optical image data in step 3 includes: Normalized Difference Vegetation Index (NDVI) was calculated using bands from Sentinel-2 optical image data. ; Wherein, NDVI is the normalized vegetation index, and NIR and Red represent the spectral reflectance in the near-infrared band and the red band, respectively. Pixels are divided into two based on the statistical distribution characteristics of the normalized vegetation index, and vegetation cover is calculated using a piecewise function: ; Wherein, FVC represents vegetation cover. and These represent the 5th and 95th percentile values ​​of NDVI, respectively.

[0033] Specifically, in step 3, the method for obtaining the three-dimensional enhanced spectral index by using a pixel multiplication fusion method to calculate forest canopy height and vegetation cover includes: Using a pixel-multiplication fusion method, vegetation cover representing horizontal density, forest canopy height representing vertical structure, and spectral and textural features representing species / physiological state are physically fused to generate a three-dimensional enhanced spectral index, including: Vegetation cover quantifies the projected density of vegetation in the horizontal dimension, forest canopy height defines the vertical extension space of vegetation, and spectral and texture features carry information such as the physiological state of vegetation. Simple feature stacking can only list this information in parallel. This method uses pixel multiplication fusion to integrate vegetation cover, which represents horizontal density, forest canopy height, which represents vertical structure, and spectral and texture features, which represent species / physiological state, into a new three-dimensional enhanced spectral index, providing input variables with richer information and clearer physical meaning. ; in, CHVI s For three-dimensional enhanced spectral indices, A pixel Represents pixel area. FVC For vegetation coverage, CH The height of the forest canopy. VI s Spectral and texture features of Sentinel-2 optical image data.

[0034] This step utilizes the peak length features of the generated continuous satellite-borne lidar to obtain the forest canopy height value corresponding to each pixel within the target area. Based on Sentinel-2 data, the normalized vegetation index (NDI) for each pixel is calculated, and a pixel-based bisection model is used to calculate vegetation cover. Vegetation cover quantifies the projected density of vegetation in the horizontal dimension, forest canopy height defines the vegetation's vertical extension space, and spectral and texture features carry information such as the physiological state of the vegetation. Simple feature stacking can only list this information in parallel. Pixel multiplication fusion is used to fuse the obtained forest canopy height and vegetation cover with the original spectral and texture features of Sentinel-2, generating a three-dimensional enhanced spectral index. The generated three-dimensional enhanced spectral index simultaneously contains information about the forest's vertical structure (contributed by forest canopy height), horizontal density (contributed by vegetation cover), and species / physiological information (contributed by spectral and texture features). This comprehensive feature can more clearly distinguish forests with different stand structures and growth states, providing input variables with richer information and clearer physical meaning for subsequent high-precision forest carbon storage inversion.

[0035] Furthermore, step 4, based on the three-dimensional enhanced spectral index, employs methods for calculating relative importance and statistical significance levels to construct a forest carbon storage prediction model. The methods for predicting forest carbon storage include: Combining the variable importance assessment method of the permutation test significance level, the relative importance and statistical significance levels of all spaceborne lidar energy characteristics and geometric characteristics, three-dimensional enhanced spectral index, and forest carbon storage were calculated. The feature with the highest relative importance ranking was selected as the first modeling variable. With the goal of minimizing model error, features that contribute significantly to the model were gradually introduced as modeling features according to the changes in model error. The optimal feature subset corresponding to the minimum model error was selected to construct the forest carbon storage prediction model, thereby achieving forest carbon storage prediction. The specific process is as follows: The random forest algorithm is used to measure the relative importance of the energy and geometric features of spaceborne lidar, the 3D enhanced spectral index, and forest carbon storage. This is achieved by randomly shuffling the values ​​of a predictive feature variable in an out-of-bag sample (an out-of-bag sample is a complete subset of data containing both "features" and "true values"; in this invention, the predictive feature variables in the out-of-bag sample are: the energy and geometric features of the spaceborne lidar, and the 3D enhanced spectral index; the target variable is: the measured value of forest carbon storage). The percentage mean square error of the prediction model after shuffling is then calculated to obtain the relative importance. ; Here, %IncMSE represents relative importance. Ntrees The total number of trees in the random forest. For the first t The original mean square error of the tree on the out-of-bag samples After the substitution variable, the first t The mean square error of each tree on samples outside the same bag.

[0036] Statistical inference of relative importance based on permutation tests is performed by repeatedly and randomly permuting the values ​​of a predictive feature variable in out-of-bag samples, calculating the relative importance under each permutation, and generating a distribution of "relative importance" for each variable under preset conditions. The actual observed relative importance of the variable is compared with the distribution obtained from the permutation, thereby generating a statistical significance level for the relative importance. ; in, pj For variables j Statistical significance level p value, nrep To replace the number of repetitions, The relative importance is calculated based on the original data. For the first r The relative importance of the permutation data calculation I (·) is an indicator function that returns 1 if the condition is true, and 0 otherwise.

[0037] The feature with the highest relative importance is used as the first modeling variable, and feature variables are introduced step by step. At each step, the feature that minimizes the model error is selected and added to the model. ; in, f k+1 For the first k +1 features to be introduced f As a candidate feature, F For feature set, S k For the currently selected k A feature subset, F \ S k These are the remaining features that have not yet been selected. J This is the error evaluation function; Based on predicted and observed values ​​of forest carbon storage, and with the goal of minimizing model error, the root mean square error is used to evaluate the changes in model error during the introduction of variables. The evaluation results guide the selection of model variables. The evaluation indicators are calculated as follows: ; in, and These are the model predictions. This represents the sample size.

[0038] This step calculates the relative importance and statistical significance of all spaceborne lidar features and 3D enhanced spectral indices with measured carbon storage data. The feature with the highest relative importance is selected as the first modeling variable to construct an initial prediction model, and its prediction error is recorded. Then, features that contribute significantly to the model are gradually introduced as modeling features based on changes in model error. The feature subset corresponding to the minimum model error is selected as the optimal feature subset for constructing the final prediction model. This step-by-step selection, aimed at minimizing model error, ensures that the constructed model achieves optimal prediction accuracy under given data conditions. Through systematic feature selection, features with weak correlation to forest carbon storage or those that are redundant are automatically eliminated. The most concise feature subset reduces computational resource consumption during model prediction and improves the efficiency of large-scale mapping.

[0039] Furthermore, the forest carbon storage estimation method of the present invention also includes: Based on predicted and observed values ​​of forest carbon storage, the coefficient of determination R0 was used. 2 Model accuracy is evaluated and model parameters are optimized using relative root mean square (rRMSE).

[0040] The evaluation indicators are calculated as follows: ; ; in, and These are the model predictions. This is the average of the measured values. This represents the sample size.

[0041] Figure 3 This image shows a comparison of the fitting results for forest carbon storage estimations using different variable combinations. The remote sensing data used included Sentinel-2 optical imagery and the "Jumang" satellite-borne lidar data. Sentinel-2 features, lidar features, and three-dimensional enhanced spectral indices were obtained, respectively. The measured forest carbon storage data was obtained from a survey conducted in Xintian County, Yongzhou City, Hunan Province in 2023. The forest carbon storage ranged from 1.843 to 119.048 Mg C / hm², with a mean of 36.435 Mg C / hm². The obtained Sentinel-2 features, lidar features, and three-dimensional enhanced spectral indices were combined to construct k-nearest neighbor (kNN), XGBoost, and random forest (RF) models for forest carbon storage prediction, yielding fitting results for forest carbon storage estimations using different variable combinations. Figure 3 (a)~(c) are the results of Sentinel-2 feature estimation. Figure 3 (d)~(f) show the results of the estimation of Sentinel-2 features and spaceborne lidar features, respectively. Figure 3 (g)~(i) represent the results of the three-dimensional enhanced spectral index estimation, respectively. Compared with using only Sentinel-2 features and spaceborne lidar features, the estimation results obtained using the three-dimensional enhanced spectral index show a significant improvement in model accuracy, with the relative root mean square error generally reduced to below 36%, verifying the effectiveness of this invention in forest carbon storage estimation.

[0042] Figure 4 This is a map illustrating the spatial distribution of forest carbon storage. Figure 4 (a) Figure 4 (b) and Figure 4 (c) These are spatial distribution maps of forest carbon storage estimated using a random forest model constructed after variable selection using Sentinel-2 features, Sentinel-2 features combined with spaceborne lidar features, and the 3D enhanced spectral index. The 3D enhanced spectral index effectively combines the forest canopy height from spaceborne lidar with the spectral texture features of Sentinel-2, improving the underestimation of forest carbon storage. The resulting spatial distribution of forest carbon storage is closer to the actual range, and the spatial distribution of forest carbon storage is largely consistent with the spatial distribution of vegetation types in Xintian County, which is in good agreement with the actual situation.

[0043] In summary, this invention provides a method for estimating forest carbon storage by fusing spaceborne LiDAR features and three-dimensional enhanced spectral indices. First, it acquires and preprocesses discrete footprint data from a full-waveform LiDAR system and Sentinel-2 optical imagery to obtain the geographic location information, energy, and geometric features corresponding to the spaceborne LiDAR footprints, as well as the spectral and texture features from Sentinel-2 imagery. Using the geographic location information of the spaceborne LiDAR footprints and the global spectral and texture features extracted from the Sentinel-2 imagery as input variables, and the energy or geometric features of the corresponding spaceborne LiDAR footprints as target variables, a random forest regression algorithm is used as a spatial interpolator to transform the sparse spaceborne LiDAR footprint data into continuous raster data, providing a reliable data foundation for subsequent high-precision, spatially continuous forest carbon storage distribution mapping. The forest canopy height is obtained using the aforementioned spaceborne LiDAR features, and vegetation cover is calculated based on the Sentinel-2 data. By employing a pixel-multiplication fusion method, vegetation cover (representing horizontal density), forest canopy height (representing vertical structure), and spectral and textural features (representing species / physiological state) are physically fused to generate a three-dimensional enhanced spectral index. This provides richer and more physically meaningful input variables for subsequent high-precision forest carbon storage inversion. Combining a variable importance assessment method with permutation tests, the relative importance and statistical significance of all spaceborne lidar features and the three-dimensional enhanced spectral index with forest carbon storage are calculated. Based on the relative importance ranking and model error changes, the optimal feature subset is gradually selected, ensuring that the constructed model achieves optimal prediction accuracy under given data conditions. The final forest carbon storage prediction model achieves complementary advantages between spaceborne lidar and optical remote sensing, more accurately reflecting the relationship between forest carbon storage and input features, reducing computational resource consumption during model prediction, and thus enabling large-scale, high-precision continuous spatial distribution mapping of forest carbon storage.

[0044] Example 2 Based on the same inventive concept, the present invention also provides a forest carbon storage estimation system that integrates spaceborne LiDAR features and three-dimensional enhanced spectral indices, for implementing the method described in the foregoing embodiments. The system includes: a data acquisition and processing module, a linear mapping module, a fusion module, and a prediction module. The data acquisition and processing module is used to acquire discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and perform preprocessing. The linear mapping module is used to perform linear mapping based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, using the random forest regression algorithm as a spatial interpolator to obtain continuous spaceborne lidar feature surfaces. The fusion module is used to obtain forest canopy height based on continuous feature surfaces of spaceborne lidar, calculate vegetation cover based on preprocessed Sentinel-2 optical image data, and use pixel multiplication fusion method to obtain three-dimensional enhanced spectral index for forest canopy height and vegetation cover. The prediction module is used to construct a forest carbon storage prediction model based on the three-dimensional enhanced spectral index and employs methods for calculating relative importance and statistical significance levels, thereby enabling the prediction of forest carbon storage.

[0045] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for estimating forest carbon storage by integrating spaceborne LiDAR features and three-dimensional enhanced spectral indices, characterized in that, The method includes: Step 1: Acquire discrete footprint data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and perform preprocessing; Step 2: Based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, the random forest regression algorithm is used as a spatial interpolator to perform linear mapping to obtain continuous spaceborne lidar feature surfaces. Step 3: Obtain forest canopy height based on continuous spaceborne lidar feature surfaces, calculate vegetation cover based on preprocessed Sentinel-2 optical image data, and use pixel multiplication fusion method to obtain three-dimensional enhanced spectral index for forest canopy height and vegetation cover. Step 4: Based on the three-dimensional enhanced spectral index, a forest carbon storage prediction model is constructed using the relative importance and statistical significance level calculation method to achieve forest carbon storage prediction.

2. The method according to claim 1, characterized in that, The method for acquiring discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and performing preprocessing in step 1 includes: Data quality checks, effective signal extraction, echo signal smoothing, waveform data decomposition, and energy and geometric feature extraction are performed on the discrete footprint point data of the full-waveform spaceborne lidar. Radiometric calibration and atmospheric correction were performed on Sentinel-2 optical image data, and spectral and texture features were extracted. Spatial matching was performed between the discrete footprint data of the full-waveform spaceborne lidar and the Sentinel-2 optical image data to obtain the geographical location, energy characteristics, and geometric characteristics of the discrete footprint data of the full-waveform spaceborne lidar, as well as the spectral and texture characteristics of the Sentinel-2 optical image data, in the same coordinate system.

3. The method according to claim 2, characterized in that, Step 2, based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, uses a random forest regression algorithm as a spatial interpolator for linear mapping to obtain continuous spaceborne lidar feature surfaces. The method includes: The random forest regression algorithm is used as a spatial interpolator. The geographical location corresponding to the discrete footprint point data of the full-waveform spaceborne lidar and the spectral and texture features of the Sentinel-2 optical image data are used as input variables. The energy features or geometric features corresponding to the discrete footprint point data of the full-waveform spaceborne lidar are used as target variables. The nonlinear mapping relationship between geographical location, spectral and texture features and energy features or geometric features is established, and the sparse spaceborne lidar footprint point data is transformed into continuous raster data, and the continuous spaceborne lidar feature surface is output.

4. The method according to claim 1, characterized in that, The method for obtaining forest canopy height based on continuous spaceborne lidar feature surfaces in step 3 includes: ； Where CH represents the forest canopy height. P len The length of the wave crest. Peak first and Peak last These represent the positions of the first and last valid peaks of the echo signal, respectively.

5. The method according to claim 4, characterized in that, The method for calculating vegetation cover based on preprocessed Sentinel-2 optical image data in step 3 includes: Normalized Difference Vegetation Index (NDVI) was calculated using bands from Sentinel-2 optical image data. ； Wherein, NDVI is the normalized vegetation index, and NIR and Red represent the spectral reflectance in the near-infrared band and the red band, respectively. Pixels are divided into two based on the statistical distribution characteristics of the normalized vegetation index, and vegetation cover is calculated using a piecewise function: ； Wherein, FVC represents vegetation cover. and These represent the 5th and 95th percentile values ​​of NDVI, respectively.

6. The method according to claim 5, characterized in that, The method for obtaining the three-dimensional enhanced spectral index by using a pixel multiplication fusion method to calculate forest canopy height and vegetation cover in step 3 includes: ； in, CHVI s For three-dimensional enhanced spectral indices, A pixel Represents pixel area. FVC For vegetation coverage, CH The height of the forest canopy. VI s Spectral and texture features of Sentinel-2 optical image data.

7. The method according to claim 1, characterized in that, Step 4, based on the three-dimensional enhanced spectral index, employs methods for calculating relative importance and statistical significance levels to construct a forest carbon storage prediction model. The method for predicting forest carbon storage includes: The random forest algorithm is used to measure the relative importance of spaceborne lidar features, 3D enhanced spectral index, and forest carbon storage. By randomly shuffling the values ​​of a predicted feature variable in the out-of-bag samples, the percentage mean square error of the model's predictions after shuffling is calculated to obtain the relative importance. ； Here, %IncMSE represents relative importance. Ntrees The total number of trees in the random forest. For the first t The original mean square error of the tree on the out-of-bag samples After the substitution variable, the first t The mean squared error of each tree on samples outside the same bag; Statistical inference of relative importance based on permutation tests is performed by repeatedly and randomly permuting the values ​​of a predictive feature variable in out-of-bag samples, calculating the relative importance under each permutation, and generating a distribution of "relative importance" for each variable under preset conditions. The actual observed relative importance of the variable is compared with the distribution obtained from the permutation, thereby generating a statistical significance level for the relative importance. ； in, pj For variables j Statistical significance level p value, nrep To replace the number of repetitions, The relative importance is calculated based on the original data. For the first r The relative importance of the permutation data calculation I (·) is an indicator function that returns 1 when the condition is true, otherwise returns 0; The feature with the highest relative importance is used as the first modeling variable, and feature variables are introduced step by step. At each step, the feature that minimizes the model error is selected and added to the model. ； in, f k+1 For the first k +1 features to be introduced f As a candidate feature, F For feature set, S k For the currently selected k A feature subset, F \ S k The remaining features that have not yet been selected. J This is the error evaluation function; Based on predicted and observed values ​​of forest carbon storage, and with the goal of minimizing model error, the root mean square error is used to evaluate the changes in model error during the introduction of variables. The evaluation results guide the selection of model variables. The evaluation indicators are calculated as follows: ； in, and These are the model predictions. This is the average of the measured values. This represents the sample size.

8. A forest carbon storage estimation system integrating spaceborne LiDAR features and three-dimensional enhanced spectral indices, said system being used to implement the method described in any one of claims 1-7, characterized in that, The system includes: a data acquisition and processing module, a linear mapping module, a fusion module, and a prediction module; The data acquisition and processing module is used to acquire discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data and perform preprocessing. The linear mapping module is used to perform linear mapping based on the preprocessed discrete footprint point data of the full-waveform spaceborne lidar and Sentinel-2 optical image data, using the random forest regression algorithm as a spatial interpolator to obtain continuous spaceborne lidar feature surfaces. The fusion module is used to obtain forest canopy height based on continuous feature surfaces of spaceborne lidar, calculate vegetation cover based on preprocessed Sentinel-2 optical image data, and use pixel multiplication fusion method to obtain three-dimensional enhanced spectral index for forest canopy height and vegetation cover. The prediction module is used to construct a forest carbon storage prediction model based on the three-dimensional enhanced spectral index and employs methods for calculating relative importance and statistical significance levels, thereby enabling the prediction of forest carbon storage.

Images