Large-scale forest green carbon inversion method and system based on multi-source remote sensing data

Abstract

The invention relates to the technical field of remote sensing space information, in particular to a large-scale forest green carbon inversion method and system based on multi-source remote sensing data, and the method comprises the steps: resolving remote sensing parameters, and obtaining the crown breadth and height of a single-plant crown of a sample plot; taking the crown breadth and the height of the single tree crown of the sample plot as independent variables, and establishing different-speed growth models of different tree species; performing inversion based on original echo signal waveform data of the satellite-borne laser radar to obtain forest structure parameters; according to the forest structure parameters, calculating laser spot scale biomass by using a different-speed growth model; and based on the laser spot scale biomass and the multi-source remote sensing image features, using a U-Net deep learning model to obtain large-scale biomass data with continuous space, and obtaining a large-scale forest above-ground biomass distribution map according to the large-scale biomass data. According to the method, high-precision inversion from single-plant crown parameters to large-scale forest above-ground biomass is realized, and the efficiency of forest biomass estimation is improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing spatial information technology, and particularly to a large-scale forest green carbon inversion method and system based on multi-source remote sensing data. Background Art

[0002] As a core component of the terrestrial ecosystem carbon sink, the accurate estimation of forest above-ground biomass is of great significance for global carbon cycle research and ecosystem management. Traditional above-ground biomass estimation mainly relies on forest plot inventory methods. Although this method can provide relatively reliable biomass estimation results, it depends on a large number of field surveys. Not only is the data acquisition cycle long and the workload large, but it is also easily affected by factors such as forest spatial heterogeneity and differences in ground plot observation times. As a result, in the estimation of large-area forest biomass, it is difficult to unify the plot survey standards and the representativeness of plot types is insufficient, making the large-scale biomass estimation have great uncertainties.

[0003] Optical remote sensing technology is widely used in forest biomass estimation because of its easy data acquisition and simple processing. However, optical remote sensing data mainly reflects the reflection spectral signals at the top of the canopy and is difficult to penetrate the canopy interior to obtain forest structure information, which limits its estimation accuracy in high-biomass forests. Although the correlation between multi-spectral data and biomass can be improved by adding different combinations of spectral indices and spatial texture features, the established optical biomass models often lack universality and are difficult to be promoted in different study areas. Synthetic aperture radar (SAR) technology, on the other hand, provides the ability to penetrate clouds and canopies and can obtain forest structure information all day and all weather. However, there are differences in the emission wavelengths and penetrations of different radar data, resulting in uneven biomass estimation accuracies of different radar data. When the biomass level is high, the radar backscattering tends to saturate, and the saturation point varies slightly with changes in forest structure, but the general trend is that the saturation point decreases with the decrease of wavelength, further limiting its estimation effect in high-biomass forests.

[0004] In addition, spaceborne lidar technology provides a new idea for solving the saturation problem in biomass estimation by optical and radar remote sensing data by recording echo signals at different canopy heights. However, when using spaceborne lidar to invert biomass, constructing a high-precision laser speckle allometric equation becomes a key step. Traditional allometric equations based on diameter at breast height and tree height are difficult to apply at the satellite remote sensing scale because satellite remote sensing is difficult to effectively obtain diameter at breast height information, which limits the application of traditional allometric equations at the plot scale.

[0005] In summary, in the current estimation of large-scale forest aboveground biomass, both traditional methods and remote sensing technologies face many challenges. Therefore, exploring a large-scale forest aboveground green carbon inversion method based on allometric equations and spaceborne remote sensing data to solve the problems existing in the prior art has become an important topic in the research on carbon sinks in forest ecosystems. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a large-scale forest green carbon inversion method and system based on multi-source remote sensing data.

[0007] In a first aspect, the present invention provides a large-scale forest green carbon inversion method based on multi-source remote sensing data, and the method includes the following steps:

[0008] Obtain remote sensing parameters, and perform calculations on the remote sensing parameters to obtain the crown diameter of a single tree in a plot and the crown height of a single tree in the plot; the remote sensing parameters include high-resolution optical remote sensing images of the plot and lidar point cloud data of the plot;

[0009] Taking the crown diameter of a single tree in the plot and the crown height of a single tree in the plot as independent variables, establish allometric growth models for different tree species;

[0010] Invert based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; the forest structure parameters include percentile height and canopy coverage;

[0011] According to the forest structure parameters, use the allometric growth models of different tree species to calculate the biomass at the laser spot scale, and extract multi-source remote sensing image features from optical and radar remote sensing images;

[0012] Based on the biomass at the laser spot scale and the multi-source remote sensing image features, use the U-Net deep learning model to predict spatially continuous large-scale biomass data;

[0013] Obtain a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data.

[0014] In a further embodiment, the step of performing calculations on the remote sensing parameters to obtain the crown diameter of a single tree in the plot and the crown height of a single tree in the plot includes:

[0015] Use a semantic segmentation algorithm to segment the crowns of high-resolution optical remote sensing images of the plot to obtain crown vector boundaries, and calculate the crown diameter of a single tree in the plot according to the crown vector boundaries;

[0016] Perform ground filtering on the lidar point cloud data of the plot, extract the lidar point cloud covered by the crowns, and calculate the height difference of the lidar point cloud covered by the crowns to obtain the crown height of a single tree in the plot.

[0017] In a further embodiment, the allometric growth model adopts a power exponent model.

[0018] In a further embodiment, the steps of inverting the forest structure parameters based on the original echo signal waveform data of spaceborne lidar include:

[0019] Using the Gaussian decomposition algorithm to decompose the original echo signal waveform data of spaceborne lidar into multiple echo peaks, and obtaining echo peak parameters; the echo peak parameters at least include ground echo peak parameters and canopy echo peak parameters;

[0020] Performing terrain slope correction on the forest canopy height based on the echo peak parameters to obtain the forest canopy height after terrain slope correction;

[0021] According to the corrected forest canopy height, performing cumulative energy sorting on the echo energy from the ground reference height to the maximum forest canopy height, calculating the height value corresponding to the position where the cumulative energy is 95% of the total energy, and obtaining the percentile height;

[0022] Inverting the canopy echo energy after Gaussian decomposition and the ground echo energy to obtain the canopy coverage of the laser spot scale.

[0023] In a further embodiment, the canopy coverage is obtained by calculating the proportion of the sum of all canopy echo energies in the total echo energy, where the total echo energy is defined as the sum of the sum of all canopy echo energies and the ground echo energy.

[0024] In a further embodiment, the terrain slope correction formula for the forest canopy height is:

[0025] RH s = gpCR - SB - 3×(σ gp - σ t )

[0026] In the formula, RH s is the forest canopy height after terrain slope correction; gpCR is the ground echo center position; SB is the starting position of the forest canopy echo; σ gp is the standard deviation of the ground echo after Gaussian decomposition; σ t is the standard deviation of the transmitted wave after Gaussian decomposition.

[0027] In a further embodiment, the steps of extracting multi-source remote sensing image features from optical and radar remote sensing images include:

[0028] Obtain radar remote sensing images, and extract radar image polarization features and radar image texture features from the radar remote sensing images; the radar image polarization features include the radar image HH polarization band, the radar image VV polarization band, and the radar image polarization band ratio;

[0029] Obtain optical remote sensing images, perform time series image synthesis on the optical remote sensing images, and obtain optical image time series quantile features.

[0030] In a further embodiment, the step of using the U-Net deep learning model to predict spatially continuous large-scale biomass data based on the laser spot scale biomass and the multi-source remote sensing image features includes:

[0031] Use the recursive feature elimination method to evaluate the feature importance of the multi-source remote sensing image features, and select the multi-source remote sensing image features with feature importance greater than a preset importance threshold as the feature variables of the U-Net deep learning model;

[0032] Perform enhanced sampling on the laser spot scale biomass to obtain enhanced samples of lidar spot biomass;

[0033] Based on the enhanced samples of lidar spot biomass and the feature variables, use the pre-trained U-Net deep learning model to predict spatially continuous large-scale biomass data.

[0034] In a further embodiment, after the step of using the pre-trained U-Net deep learning model to predict spatially continuous large-scale biomass data based on the enhanced samples of lidar spot biomass and the feature variables, it further includes:

[0035] Construct an iterative deep learning network model, and perform accuracy evaluation on the large-scale biomass data to screen out high-value underestimated samples and low-value overestimated samples in biomass inversion; the iterative deep learning network model is a U-Net architecture combining upsampling and downsampling with skip connections;

[0036] Use the high-value underestimated samples and low-value overestimated samples as biomass deviation samples, and use the iterative deep learning network model to iterate the biomass deviation samples to generate synthetic samples;

[0037] Re-input the synthetic samples into the U-Net deep learning model for updated prediction to obtain deviation-corrected large-scale biomass data.

[0038] In a second aspect, the present invention provides a large-scale forest green carbon inversion system based on multi-source remote sensing data, and the system includes:

[0039] A data acquisition module, which is used to acquire remote sensing parameters, calculate the acquired remote sensing parameters, and obtain the crown width of a single tree crown in a sample plot and the height of a single tree crown in the sample plot; the remote sensing parameters include high-resolution optical remote sensing images of the sample plot and lidar point cloud data of the sample plot;

[0040] A model establishment module, which is used to establish allometric growth models of different tree species with the crown width of a single tree crown in the sample plot and the height of a single tree crown in the sample plot as independent variables;

[0041] A data inversion module, which is used to perform inversion based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; the forest structure parameters include percentile height and canopy coverage;

[0042] A feature extraction module, which is used to calculate the biomass of laser spot scale by using allometric growth models of different tree species according to the forest structure parameters, and extract multi-source remote sensing image features from optical and radar remote sensing images;

[0043] A model prediction module, which is used to predict spatially continuous large-scale biomass data by using a U-Net deep learning model based on the biomass of laser spot scale and the multi-source remote sensing image features;

[0044] A distribution map acquisition module, which is used to obtain a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data.

[0045] The present invention provides a large-scale forest green carbon inversion method and system based on multi-source remote sensing data. The method obtains remote sensing parameters, calculates the remote sensing parameters, and obtains the crown width of a single tree crown in a sample plot and the height of a single tree crown in the sample plot; establishes allometric growth models of different tree species with the crown width of a single tree crown in the sample plot and the height of a single tree crown in the sample plot as independent variables; performs inversion based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; calculates the biomass of laser spot scale by using allometric growth models of different tree species according to the forest structure parameters, and extracts multi-source remote sensing image features from optical and radar remote sensing images; predicts spatially continuous large-scale biomass data by using a U-Net deep learning model based on the biomass of laser spot scale and the multi-source remote sensing image features; obtains a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data. Compared with the prior art, by fusing multi-source remote sensing data, combining allometric growth models and U-Net deep learning models, the method realizes high-precision inversion from single-tree crown parameters to large-scale forest aboveground biomass, and effectively improves the accuracy and spatial continuity of forest biomass estimation. Description of the Drawings

[0046] Figure 1Schematic diagram of the large-scale forest green carbon inversion method based on multi-source remote sensing data provided by an embodiment of the present invention;

[0047] Figure 2 Block diagram of the large-scale forest green carbon inversion process provided by an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of the iterative process of the iterative deep learning network model provided by an embodiment of the present invention;

[0049] Figure 4 Block diagram of the large-scale forest green carbon inversion system based on multi-source remote sensing data provided by an embodiment of the present invention. Detailed implementation manners

[0050] The following specifically clarifies the implementation manners of the present invention in conjunction with the accompanying drawings. The given embodiments are only for illustrative purposes and should not be construed as limiting the present invention. The accompanying drawings are only for reference and illustration and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0051] Refer to Figure 1 An embodiment of the present invention provides a large-scale forest green carbon inversion method based on multi-source remote sensing data. As Figure 1 shown, the method includes the following steps:

[0052] S1. Obtain remote sensing parameters and perform calculations on the remote sensing parameters to obtain the crown width of a single tree in a plot and the crown height of a single tree in a plot; the remote sensing parameters include high-resolution optical remote sensing images of the plot and lidar point cloud data of the plot.

[0053] In some implementation manners, the step of performing calculations on the remote sensing parameters to obtain the crown width of a single tree in a plot and the crown height of a single tree in a plot includes:

[0054] Use a semantic segmentation algorithm to segment the crowns of the high-resolution optical remote sensing images of the plot to obtain the crown vector boundaries, and calculate the crown width of a single tree in the plot according to the crown vector boundaries;

[0055] Perform ground filtering on the lidar point cloud data of the plot, extract the lidar point cloud covered by the crowns, and calculate the height difference of the lidar point cloud covered by the crowns to obtain the crown height of a single tree in the plot.

[0056] Specifically, as Figure 2As shown in the figure, in this embodiment, high-resolution optical remote sensing images corresponding to the sample plots are obtained. The high-resolution optical remote sensing images of the sample plots include sub-meter satellite high-resolution images, such as GF-7 panchromatic band images or centimeter-level visible light three-band images obtained by drones, etc., ensuring that the acquisition time of the images corresponds to the growing season of forest trees in order to accurately reflect the growth status of the trees. In this embodiment, these high-resolution optical remote sensing images of the sample plots are used to segment the tree crowns in the high-resolution optical remote sensing images of the sample plots by using the watershed algorithm or semantic segmentation algorithms such as YOLOV8. After segmentation, the vector boundaries of the tree crowns are obtained. Among them, the watershed algorithm regards the image as a terrain surface and uses the flow characteristics of water to segment objects; the YOLOV8 algorithm is a deep learning-based object detection algorithm that can quickly and accurately identify and segment tree crowns. In this embodiment, through the segmented vector boundaries of the tree crowns, the crown width of each tree crown can be accurately calculated.

[0057] Meanwhile, in this embodiment, LiDAR point cloud data corresponding to the sample plots are obtained. These LiDAR point cloud data of the sample plots can be obtained by using an unmanned aerial vehicle (UAV) LiDAR system, a backpack LiDAR device, or a handheld LiDAR device, and it is ensured that the acquisition time of the LiDAR point cloud data of the sample plots corresponds to the growing season of forest trees. Ground filtering is performed on the obtained LiDAR point cloud data of the sample plots. For example, ground filtering methods such as morphological filtering algorithms or machine learning-based classification algorithms are used to distinguish ground points and non-ground points, thereby generating a canopy height model. Based on the canopy height model, in this embodiment, the LiDAR points in the tree crown coverage area are extracted, and by calculating the height difference of the tree crown coverage point cloud, that is, the vertical distance between the highest point in the point cloud and the ground point, the height of a single tree crown in the sample plot is determined.

[0058] S2. Taking the crown width of a single tree crown in the sample plot and the height of a single tree crown in the sample plot as independent variables, establish allometric growth models for different tree species.

[0059] Specifically, in this embodiment, according to the extracted crown width of a single tree crown in the sample plot and the height of a single tree crown in the sample plot, allometric growth models are established for different tree species. The allometric growth model adopts the form of a power exponential model, that is:

[0060] F(C, H) = a * C b * H c

[0061] In the formula, F(C, H) is the allometric growth model function; C is the crown width of a single tree crown in the sample plot; H is the height of a single tree crown in the sample plot; a, b, and c are all model fitting parameters.

[0062] In this embodiment, by fitting the crown width and tree height data of different tree species, an allometric growth equation corresponding to each type of tree species can be obtained. If there are multiple tree species in the plot, each type of tree species will have a corresponding allometric growth equation, thereby achieving accurate calculation and modeling of the crown width and tree height of different tree species in the plot.

[0063] S3. Invert the original echo signal waveform data based on spaceborne lidar to obtain forest structure parameters; the forest structure parameters include percentile height and canopy cover.

[0064] In some embodiments, the step of inverting the original echo signal waveform data based on spaceborne lidar to obtain forest structure parameters includes:

[0065] Use the Gaussian decomposition algorithm to decompose the original echo signal waveform data of spaceborne lidar into multiple echo peaks, and obtain echo peak parameters; the echo peak parameters at least include ground echo peak parameters and canopy echo peak parameters;

[0066] Perform terrain slope correction on the forest canopy height based on the echo peak parameters to obtain the forest canopy height after terrain slope correction;

[0067] According to the corrected forest canopy height, perform cumulative energy sorting on the echo energy from the ground reference height to the maximum forest canopy height, and calculate the height value corresponding to the position where the cumulative energy is 95% of the total energy to obtain the percentile height;

[0068] Invert the canopy echo energy after Gaussian decomposition and the ground echo energy to obtain the canopy cover of the laser spot scale.

[0069] Specifically, in this embodiment, the original echo signal waveform data of the spaceborne lidar is obtained as the input. The original echo signal waveform data of the spaceborne lidar contains the lidar echo information of the forest area, and the Gaussian decomposition algorithm is used to decompose the original echo signal waveform data, dividing the original echo signal waveform data into several echo peaks. Each echo peak corresponds to different targets such as the canopy and the ground respectively. The echo peaks at least include the canopy echo peak and the ground echo peak. During the decomposition process, this embodiment needs to record key parameters such as the start and end points, the peak positions, widths, and energy ratios of each intermediate echo, that is, the echo peak parameters at least include the echo peak position, the echo peak width, and the echo energy information. Through these key parameters, different canopy levels can be effectively distinguished, providing accurate data support for subsequent terrain correction and parameter calculation. Then, this embodiment adopts a terrain correction method based on the lidar physical model to correct the terrain slope of the forest canopy height. Specifically, this embodiment calculates the forest canopy height after slope correction according to the center position of the ground echo, the start position of the forest canopy echo, and the standard deviations of the ground and the transmitted wave after decomposition. In this embodiment, the terrain slope correction formula for the forest canopy height is:

[0070] RH s = gpCR - SB - 3×(σ gp - σ t )

[0071] In the formula, RH s is the forest canopy height after terrain slope correction; gpCR is the center position of the ground echo; SB is the starting position of the forest canopy echo; σ gp is the standard deviation of the ground echo after Gaussian decomposition; σ t is the standard deviation of the transmitted wave after Gaussian decomposition.

[0072] Through the terrain slope correction formula of the forest canopy height, this embodiment can accurately correct the terrain, eliminate the influence of terrain undulation on lidar data, and thus obtain accurate forest canopy height information. Then, this embodiment sorts the cumulative energy of the echo energy after terrain correction from the ground to the maximum height. By finding the height value corresponding to when the cumulative energy reaches 95%, the percentile height RH95 is obtained, thereby reflecting the height distribution characteristics of the forest canopy.

[0073] Finally, in this embodiment, the canopy coverage is calculated using the laser echo model. The canopy coverage reflects the cumulative canopy width of multiple tree crowns at the scale of the light spot. In this embodiment, the canopy coverage is obtained by calculating the proportion of the total echo energy of all canopy echoes in the total echo energy. The total echo energy is defined as the sum of the total echo energy of all canopy echoes and the ground echo energy. In this embodiment, the forest canopy coverage is determined by calculating the ratio of the non-ground canopy echo energy to the sum of the ground echo energy and the non-ground canopy echo energy. The higher this ratio is, the higher the forest coverage is, that is, the larger the area covered by the tree crowns and the denser the forest. In this embodiment, the calculation formula for the canopy coverage is:

[0074]

[0075] In the formula, RC is the forest canopy coverage, that is, the degree of canopy coverage; ∑ i＝NLM (A i ·S i ) represents the summation of the energies of all non-ground (NLM, Non-Ground Layer) canopy echoes; i represents the i-th canopy echo. In this embodiment, i is an index variable used to distinguish different canopy echoes; NLM is the non-ground canopy, that is, all canopy echoes that do not belong to the ground echo; A i is the amplitude of the i-th canopy echo, which reflects the intensity of this canopy echo; S i is the area or coverage range of the i-th canopy echo; A LM is the amplitude of the ground echo, which reflects the intensity of the ground echo; S LM is the area or coverage range of the ground echo; A LM ·S LM represents the energy of the ground echo; A i ·S i represents the energy of the i-th canopy (non-ground: NLM) echo, that is, the product of the amplitude and area of this canopy echo, reflecting the total energy of this canopy echo.

[0076] S4. According to the forest structure parameters, calculate the biomass at the laser spot scale using the allometric growth models of different tree species, and extract multi-source remote sensing image features from optical and radar remote sensing images.

[0077] In some embodiments, the step of extracting multi-source remote sensing image features from optical and radar remote sensing images includes:

[0078] Obtain radar remote sensing images, and extract radar image polarization features and radar image texture features from the radar remote sensing images; the radar image polarization features include the radar image HH polarization band, the radar image VV polarization band, and the radar image polarization band ratio; the radar image texture features include the variance, contrast, sum of variance, inertia, prominence clustering, and dissimilarity of the radar image texture.

[0079] Obtain optical remote sensing images, perform time series image synthesis on the optical remote sensing images to obtain optical image time series quantile features; the optical image time series quantile features include the maximum value, minimum value, quartile value, mean value, and quantile value features.

[0080] Specifically, in this embodiment, by using the percentile height RH95 and canopy cover obtained by inversion, combined with the allometric growth model established in advance for different tree species, and calculating by substituting the percentile height RH95 and canopy cover into the allometric growth model, the biomass at the laser spot scale can be accurately estimated. Among them, when the percentile height RH95 and canopy cover are input into the allometric growth model, C represents the canopy cover, H represents the percentile height RH95, and a, b, and c are all model fitting parameters. These parameters are determined through a large amount of sample data and statistical analysis methods, and can reflect the growth laws of different tree species and the relationship between biomass and structural parameters. Then, in this embodiment, multi-source remote sensing image feature extraction is carried out, and the specific steps are as follows:

[0081] First, obtain Sentinel-1 radar remote sensing image data and Sentinel-2 optical remote sensing image data. The radar remote sensing image data includes HH and VV polarization bands, and these data contain radar echo information under different polarization modes; the optical remote sensing image data includes multiple bands such as visible light, near-infrared, red edge, and short-wave infrared. Then, in this embodiment, remote sensing processing software or programming means are used to process the Sentinel-1 radar remote sensing image data and Sentinel-2 optical remote sensing image data. For the Sentinel-1 radar remote sensing image data, in this embodiment, the ratio of the HH and VV polarization bands (HH / VV) is extracted, and various texture features of the Sentinel-1 radar image are calculated, specifically including texture features such as variance, contrast, sum variance, inertia, cluster prominence, and dissimilarity. Among them, variance reflects the dispersion degree of the image gray values, and in this embodiment, the variance of the gray values within the neighborhood of each pixel can be calculated; contrast is used to measure the size of the difference between the image gray levels, and in this embodiment, the contrast between different gray levels is calculated to reflect the clarity of the image; sum variance is the variance calculated based on the gray-level co-occurrence matrix, reflecting the overall texture features of the image; inertia represents the uniformity of the gray distribution, and the larger the inertia value, the more uneven the gray distribution; cluster prominence reflects the prominence of the clusters in the image, and the larger the cluster prominence value, the more obvious the clusters; dissimilarity is used to measure the dissimilarity degree between different gray levels, and the larger the value, the higher the dissimilarity. These features can reflect the roughness and structural information of the land cover.

[0082] At the same time, for the Sentinel-2 optical images, in this embodiment, cloud-free Sentinel-2 optical remote sensing image data within a year is selected, and time-series image synthesis is performed on the cloud-free Sentinel-2 optical remote sensing image data within a year through time-series synthesis technology to generate a synthetic image, and the maximum value, minimum value, quartile value, mean value, and 10% and 90% percentile value features of the synthetic image are calculated using remote sensing processing software or programming methods. Among them, the maximum value of the synthetic image reflects the maximum value of the pixel gray values in the image, which represents the maximum reflectance of the ground object in a certain band; the minimum value of the synthetic image reflects the minimum value of the pixel gray values in the image, which represents the minimum reflectance of the ground object in a certain band; these features can capture the growth status and change information of the surface vegetation at different time points, providing rich data sources for subsequent analysis.

[0083] S5. Based on the laser spot scale biomass and the multi-source remote sensing image features, use the U-Net deep learning model to predict spatially continuous large-scale biomass data.

[0084] In some embodiments, the step of using the U-Net deep learning model to predict spatially continuous large-scale biomass data based on the laser spot scale biomass and the multi-source remote sensing image features includes:

[0085] Use the recursive feature elimination method to evaluate the feature importance of the multi-source remote sensing image features, and select the multi-source remote sensing image features with feature importance greater than a preset importance threshold as the feature variables of the U-Net deep learning model;

[0086] Perform enhanced sampling on the laser spot scale biomass to obtain an enhanced sample of lidar spot biomass;

[0087] Based on the enhanced sample of lidar spot biomass and the feature variables, use the pre-trained U-Net deep learning model to predict spatially continuous large-scale biomass data.

[0088] In this embodiment, the recursive feature elimination method is used to evaluate the feature importance of the multi-source remote sensing image features. Specifically, in this embodiment, the recursive feature elimination (RFE) method is used to evaluate the importance of the HH and VV polarization bands, HH / VV ratio, and various texture features of the input Sentinel-1 radar image. Among them, the texture features include variance, contrast, sum variance, inertia, prominence clustering, and dissimilarity, etc. At the same time, the recursive feature elimination method is used to evaluate the importance of the maximum, minimum, quartile, mean, and 10%, 90% percentile features synthesized from the input Sentinel-2 optical remote sensing image and its time series. The recursive feature elimination method is used to evaluate the importance of different variables, and the features with relatively high importance are selected as the feature variables for input to the deep learning model for prediction.

[0089] Then, in this embodiment, the SMOTE (Synthetic Minority Oversampling Technique) technology is adopted to perform enhanced sampling on the lidar light spot biomass samples to obtain lidar light spot biomass enhanced samples, so as to expand the proportion of minority samples such as high values and low values, solve the problem of sample imbalance, improve the generalization ability and prediction accuracy of the model. In this embodiment, a U-Net deep learning model is used to perform model training and prediction by combining the selected feature variables and the lidar light spot biomass enhanced samples, predict spatially continuous biomass, realize large-scale biomass mapping, and evaluate the model accuracy according to the output uncertainty. In some embodiments, after the step of predicting spatially continuous large-scale biomass data by using the pre-trained U-Net deep learning model based on the lidar light spot biomass enhanced samples and the feature variables, the following steps are further included:

[0090] Construct an iterative deep learning network model, evaluate the accuracy of the large-scale biomass data, and screen out high-value underestimated samples and low-value overestimated samples in biomass inversion; the iterative deep learning network model is a U-Net architecture that combines upsampling, downsampling and skip connections;

[0091] Use the high-value underestimated samples and the low-value overestimated samples as biomass deviation samples, and use the iterative deep learning network model to iterate the biomass deviation samples to generate synthetic samples;

[0092] Re-enter the synthetic samples into the U-Net deep learning model for updated prediction to obtain large-scale biomass data with bias correction.

[0093] In the process of lidar spot biomass estimation, there is often a problem of aboveground biomass estimation deviation caused by insufficient sample quantity, specifically manifested as underestimation of high-value biomass and overestimation of low-value biomass. This is mainly because in the actual biomass distribution, the proportion of higher and lower values is small, and overall it approximately satisfies the normal distribution law. In the seamless expansion of lidar light spots and multi-source remote sensing data, there will be a problem of sample imbalance, resulting in the inversion model being unable to accurately predict some outlier samples. To solve this problem, this embodiment proposes an iterative deep learning network model to achieve seamless expansion of biomass to solve the problem of sample imbalance, as Figure 3As shown, in this embodiment, high-value and low-value samples are selected from the prediction results of the U-Net deep learning model and re-added to the training set to solve the problem of sample imbalance, thereby improving the phenomenon of underestimation of high values and overestimation of low values in biomass inversion. Specifically, for the problems of underestimation of high values and overestimation of low values, this embodiment further constructs an iterative deep learning network model. The iterative deep learning network model is a U-Net architecture that combines upsampling, downsampling, and skip connections. The iterative deep learning network model inputs the prediction results of the first time into the iterative process, screens out some high-value and low-value samples for expanding the training set, generates biomass deviation samples, and through continuous iterative training, this embodiment gradually optimizes the model to improve the accuracy of biomass prediction. During the training process, this embodiment uses the mean square error loss function to constrain the overall performance of the biomass inversion model to ensure that the difference between the predicted value and the true value is minimized. At the same time, this embodiment introduces a correlation coefficient loss function to constrain the distribution of the predicted biomass, making the distribution of the predicted biomass closer to the distribution of the true biomass. The correlation coefficient loss function is based on the strategy of calculating the correlation coefficient and measures the ratio of the covariance to the variance between the predicted biomass and the true biomass. Combining these two loss functions, this embodiment forms a comprehensive loss function by weighted summation with a balance factor to constrain the overall iterative network. In this embodiment, the mathematical expression of the comprehensive loss function is:

[0094] L3 = L1×σ+(1 - L2)

[0095] Wherein,

[0096]

[0097] In the formula, L3 is the comprehensive loss function; L1 is the mean square error loss function, which is used to measure the difference between the biomass predicted by the model and the true biomass; L2 is the loss function based on the correlation coefficient, which is used to constrain the distribution of the predicted biomass to make it close to the distribution of the true biomass; σ is the balance factor, and its value is taken as the reciprocal of the mean value of the true biomass, that is y AGB,j is the true biomass of the jth biomass sample; y AGB is the true biomass; f(x) AGB,j is the predicted biomass of the jth biomass sample, which is the biomass predicted value calculated by combining the U-Net deep learning model with multi-source remote sensing image features and lidar light spot biomass; j is the biomass sample index; n is the total number of biomass samples; Cov(y AGB,j , f(x) AGB,j ) is the covariance between the true biomass and the predicted biomass; Var(y AGB,j ) is the variance of the true biomass; Var(f(x) AGB,j ) is the variance of the predicted biomass.

[0098] Finally, in this embodiment, the newly trained synthetic samples are re-introduced into the U-Net deep learning model, and the biomass and important features of the lidar light spots are recombined for prediction. Finally, spatially continuous large-scale biomass data are obtained. These data are presented in the form of a map, forming a high-precision large-scale biomass mapping result, which can display the distribution of forest biomass. This result can accurately reflect the spatial distribution characteristics of biomass, effectively solve the problems of high-value underestimation and low-value overestimation in biomass inversion, improve the accuracy and reliability of biomass mapping, and provide strong support for fields such as ecological monitoring and resource assessment.

[0099] S6. Obtain a spatially continuous large-scale forest aboveground biomass distribution map based on the spatially continuous large-scale biomass data.

[0100] Traditional large-scale biomass inversion techniques highly rely on plot data measured manually in the field. By establishing a statistical regression model between the artificial plots and remote sensing data to deduce the biomass distribution. However, this statistical regression model is limited by the spatio-temporal distribution characteristics of the samples, resulting in its limitations in the application scope. To solve this problem, this embodiment constructs a light spot-scale allometric growth model based on satellite observation data. This model integrates lidar, optical, and radar data to improve the accuracy of forest biomass inversion. Compared with traditional methods, the green carbon inversion method proposed in this embodiment abandons the traditional allometric growth equation that relies on diameter at breast height and tree height, and instead uses the crown width and tree height information extracted from high-resolution satellite images to establish an allometric growth equation adapted to the characteristics of satellite parameters for the plots, providing a more accurate basis for biomass estimation. It not only simplifies the data collection process, but also improves the applicability and accuracy of the model. At the same time, to solve the common problem of unbalanced distribution of small-type samples in the biomass seamless extrapolation model, this embodiment uses an iterative deep learning network and combines sample enhancement techniques. The overall model is constrained by the mean square error loss function, the matching degree between the predicted biomass and the true biomass distribution frequency is adjusted by the correlation coefficient, and an iterative network cycle selection strategy is adopted to re-introduce high-value and low-value samples for training, effectively balancing the sample distribution and improving the generalization ability of the model.

[0101] An embodiment of the present invention provides a large-scale forest green carbon inversion method based on multi-source remote sensing data. The method includes obtaining remote sensing parameters and resolving the remote sensing parameters to obtain the crown width of a single tree in a plot and the tree height of a single tree in the plot; using the crown width of a single tree in the plot and the tree height of a single tree in the plot as independent variables to establish allometric growth models for different tree species; performing inversion based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; according to the forest structure parameters, using the allometric growth models of different tree species to calculate the biomass at the laser spot scale, and extracting multi-source remote sensing image features from optical and radar remote sensing images; based on the biomass at the laser spot scale and the multi-source remote sensing image features, using a U-Net deep learning model to predict spatial continuous large-scale biomass data; and obtaining a spatial continuous large-scale forest aboveground biomass distribution map according to the spatial continuous large-scale biomass data. Compared with the prior art, by fusing multi-source remote sensing data, combining allometric growth models and a U-Net deep learning model, this method realizes high-precision inversion from single-tree crown parameters to large-scale forest aboveground biomass, effectively improving the accuracy and spatial continuity of forest biomass estimation.

[0102] It should be noted that the magnitudes of the serial numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0103] In one embodiment, as Figure 4 shown, an embodiment of the present invention provides a large-scale forest green carbon inversion system based on multi-source remote sensing data. The system includes:

[0104] A data acquisition module 101, configured to obtain remote sensing parameters and resolve the remote sensing parameters to obtain the crown width of a single tree in a plot and the tree height of a single tree in the plot; the remote sensing parameters include the high-resolution optical remote sensing image of the plot and the lidar point cloud data of the plot;

[0105] A model establishment module 102, configured to use the crown width of a single tree in the plot and the tree height of a single tree in the plot as independent variables to establish allometric growth models for different tree species;

[0106] A data inversion module 103, configured to perform inversion based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; the forest structure parameters include the percentile height and the canopy coverage;

[0107] A feature extraction module 104, configured to calculate the biomass at the laser spot scale according to the forest structure parameters by using the allometric growth models of different tree species, and extract multi-source remote sensing image features from optical and radar remote sensing images;

[0108] The model prediction module 105 is configured to predict spatially continuous large-scale biomass data by using a U-Net deep learning model based on the laser spot scale biomass and the multi-source remote sensing image features;

[0109] The distribution map acquisition module 106 is configured to obtain a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data.

[0110] For the specific limitations of a large-scale forest green carbon inversion system based on multi-source remote sensing data, reference may be made to the above limitations on a large-scale forest green carbon inversion method based on multi-source remote sensing data, which will not be elaborated here. Those of ordinary skill in the art can realize that, in combination with the various modules and steps described in the embodiments disclosed in the present application, they can be implemented by hardware, software, or a combination of both. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0111] An embodiment of the present invention provides a large-scale forest green carbon inversion system based on multi-source remote sensing data. The system acquires remote sensing parameters through a data acquisition module, and performs calculations on the remote sensing parameters to obtain the crown width of a single tree in a sample plot and the tree height of a single tree in a sample plot. The model establishment module establishes allometric growth models for different tree species with the crown width of a single tree in a sample plot and the tree height of a single tree in a sample plot as independent variables. The data inversion module performs inversion based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters. The feature extraction module calculates the laser spot scale biomass by using the allometric growth models of different tree species according to the forest structure parameters, and extracts multi-source remote sensing image features from optical and radar remote sensing images. The model prediction module predicts spatially continuous large-scale biomass data by using a U-Net deep learning model based on the laser spot scale biomass and the multi-source remote sensing image features. The distribution map acquisition module obtains a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data. Compared with the prior art, by fusing multi-source remote sensing data and combining allometric growth models and U-Net deep learning models, the system realizes high-precision inversion from single-tree crown parameters to large-scale forest aboveground biomass, effectively improving the accuracy and spatial continuity of forest biomass estimation.

[0112] The above-described embodiments merely represent several preferred embodiments of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claims described above.

Claims

1. A large-scale forest green carbon inversion method based on multi-source remote sensing data, characterized in that, Including the following steps: Obtain remote sensing parameters, and resolve the remote sensing parameters to obtain the crown width of a single tree in the plot and the height of the crown of a single tree in the plot; the remote sensing parameters include high-resolution optical remote sensing images of the plot and lidar point cloud data of the plot; Taking the crown width of a single tree in the plot and the height of the crown of a single tree in the plot as independent variables, establish allometric growth models for different tree species; Invert based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; the forest structure parameters include percentile height and canopy coverage; According to the forest structure parameters, use the allometric growth models of different tree species to calculate the biomass at the laser spot scale, and extract multi-source remote sensing image features from optical and radar remote sensing images; Based on the biomass at the laser spot scale and the multi-source remote sensing image features, use the U-Net deep learning model to predict spatially continuous large-scale biomass data; Obtain a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data.

2. The large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 1, characterized in that, The step of resolving the remote sensing parameters to obtain the crown width of a single tree in the plot and the height of the crown of a single tree in the plot includes: Use the semantic segmentation algorithm to segment the crowns of the high-resolution optical remote sensing images of the plot to obtain the crown vector boundaries, and resolve the crown width of a single tree in the plot according to the crown vector boundaries; Perform ground filtering on the lidar point cloud data of the plot, extract the lidar points covered by the crowns, and calculate the height difference of the lidar points covered by the crowns to obtain the height of the crown of a single tree in the plot.

3. A large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 1, characterized in that: The allometric growth model uses a power exponential model.

4. A large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 1, characterized in that, The step of inverting based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters includes: Use the Gaussian decomposition algorithm to decompose the original echo signal waveform data of spaceborne lidar into multiple echo peaks, and obtain echo peak parameters; the echo peak parameters at least include ground echo peak parameters and canopy echo peak parameters; Perform terrain slope correction on the forest canopy height based on the echo peak parameters to obtain the terrain slope-corrected forest canopy height; According to the corrected forest canopy height, perform cumulative energy sorting on the echo energy from the ground reference height to the maximum forest canopy height, and calculate the height value corresponding to the position where the cumulative energy is 95% of the total energy to obtain the percentile height; Invert the canopy echo energy after Gaussian decomposition and the ground echo energy to obtain the canopy coverage at the laser spot scale.

5. The large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 4, characterized in that: The canopy coverage is obtained by calculating the proportion of the sum of all canopy echo energies in the total echo energy, where the total echo energy is defined as the sum of the sum of all canopy echo energies and the ground echo energy.

6. The large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 4, wherein, The terrain slope correction formula for the forest canopy height is: RH s = gpCR - SB - 3×(σ gp - σ t ) where, RH s is the forest canopy height after terrain slope correction; gpCR is the ground echo center position; SB is the starting position of the forest canopy echo; σ gp is the standard deviation of the ground echo after Gaussian decomposition; σ t is the standard deviation of the transmitted wave after Gaussian decomposition.

7. The large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 1, characterized in that The step of extracting multi-source remote sensing image features from optical and radar remote sensing images includes: Obtain radar remote sensing images, and extract radar image polarization features and radar image texture features from the radar remote sensing images; the radar image polarization features include the HH polarization band of the radar image, the VV polarization band of the radar image, and the radar image polarization band ratio; Obtain an optical remote sensing image, perform time-series image synthesis on the optical remote sensing image, and obtain the quantile characteristics of the optical image time series.

8. A large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 1, characterized in that, The step of predicting the spatially continuous large-scale biomass data by using the U-Net deep learning model based on the laser spot scale biomass and the multi-source remote sensing image features includes: Use the recursive feature elimination method to evaluate the feature importance of the multi-source remote sensing image features, and select the multi-source remote sensing image features with feature importance greater than the preset importance threshold as the feature variables of the U-Net deep learning model; Perform enhanced sampling on the laser spot scale biomass to obtain an enhanced sample of the lidar spot biomass; Based on the enhanced sample of the lidar spot biomass and the feature variables, use the pre-trained U-Net deep learning model to predict the spatially continuous large-scale biomass data.

9. A large-scale forest green carbon inversion method based on multi-source remote sensing data according to claim 8, characterized in that, After the step of predicting the spatially continuous large-scale biomass data by using the pre-trained U-Net deep learning model based on the enhanced sample of the lidar spot biomass and the feature variables, it further includes: Construct an iterative deep learning network model, and perform accuracy evaluation on the large-scale biomass data to screen out the high-value underestimated samples and low-value overestimated samples in biomass inversion; the iterative deep learning network model is a U-Net architecture that combines upsampling and downsampling with skip connections; Use the high-value underestimated samples and low-value overestimated samples as biomass deviation samples, and use the iterative deep learning network model to iterate on the biomass deviation samples to generate synthetic samples; Re-input the synthetic samples into the U-Net deep learning model for updated prediction to obtain the large-scale biomass data with bias correction.

10. A large-scale forest green carbon inversion system based on multi-source remote sensing data, characterized in that, The system includes: A data acquisition module, configured to acquire remote sensing parameters and perform calculations on the remote sensing parameters to obtain the crown width of a single tree in a sample plot and the height of a single tree in a sample plot; the remote sensing parameters include the high-resolution optical remote sensing image of the sample plot and the lidar point cloud data of the sample plot; A model establishment module, configured to establish allometric growth models of different tree species with the crown width of a single tree in the sample plot and the height of a single tree in the sample plot as independent variables; A data inversion module, configured to perform inversion based on the original echo signal waveform data of spaceborne lidar to obtain forest structure parameters; the forest structure parameters include percentile height and canopy coverage; A feature extraction module, configured to calculate the laser spot scale biomass according to the forest structure parameters by using allometric growth models of different tree species, and extract multi-source remote sensing image features from optical and radar remote sensing images; A model prediction module, configured to predict the spatially continuous large-scale biomass data by using the U-Net deep learning model based on the laser spot scale biomass and the multi-source remote sensing image features; A distribution map acquisition module, configured to obtain a spatially continuous large-scale forest aboveground biomass distribution map according to the spatially continuous large-scale biomass data.

Images