A forest canopy height inversion method and device, electronic equipment and storage medium

Abstract

The application provides a forest canopy height inversion method and device, electronic equipment and storage medium, the method comprises the following steps: firstly, integrating spaceborne lidar, unmanned aerial vehicle lidar and multi-source auxiliary data, after targeted quality screening and pretreatment, the standardized basic data is obtained; then, taking the unmanned aerial vehicle lidar data as the true value reference, the system deviation of the spaceborne data is calculated and differential calibration is carried out, and the calibrated data with unified scale is obtained. Then, based on the multi-source auxiliary data, a fusion feature set is constructed, the modeling input features are obtained by optimization and screening; based on the calibrated spaceborne data, an image and table double-branch fusion model is constructed, a regression model is trained after input feature extraction and fusion, and a canopy height inversion model is generated. Finally, the model is applied to the target area, the spatial continuous high-resolution product is generated through sliding window reasoning, and the accuracy verification and uncertainty quantification are completed combined with the previous data. The application can realize the accurate and continuous inversion of the canopy height of the cold temperate zone forest.

Description

Technical Field

[0001] This application relates to the field of forest ecological monitoring, and in particular to a method, device, electronic equipment and storage medium for forest canopy height inversion. Background Technology

[0002] Forest canopy height is a core parameter characterizing the structure and function of forest ecosystems, and its accurate acquisition is of great significance for forest carbon storage assessment and biodiversity conservation. Cold-temperate forests, as key carbon reservoirs and ecological barriers, face significantly greater challenges in canopy height retrieval compared to other forest types due to their complex topography, strong spatial heterogeneity, and the existence of insufficient coverage and significant biases in high-latitude satellite remote sensing data.

[0003] Existing inversion methods mostly rely on single spaceborne LiDAR, optical data, or traditional machine learning models, which have significant shortcomings: spaceborne data is prone to systematic underestimation and sensor calibration is inconsistent; the fusion degree of multi-source data is low; traditional models have weak generalization ability; and inversion products lack spatial continuity and cannot reveal ecological driving patterns. Although UAV LiDAR has high accuracy, its coverage is limited; existing deep learning solutions have poor adaptability and weak interpretability. There is an urgent need for an integrated inversion solution that is adapted to the characteristics of temperate and cold-climate forests and balances accuracy and practicality. Summary of the Invention

[0004] In view of this, the embodiments of this application provide a method, apparatus, electronic device and storage medium for forest canopy height inversion, which can solve the problems of large deviation of satellite data, insufficient fusion of multi-source data, weak model generalization, low product resolution and lack of ecological mechanism analysis in the inversion of canopy height of cold temperate forests. It can realize accurate and continuous inversion of canopy height of cold temperate forests, and at the same time reveal its ecological driving law, providing technical support for forest resource management and ecological protection.

[0005] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for inverting forest canopy height, comprising the following steps: Acquire spaceborne lidar data, UAV lidar data, and multi-source auxiliary data; perform quality screening and preprocessing on various types of data to obtain basic data that meets the modeling requirements. Using the preprocessed UAV lidar data as a truth reference, the system deviation of different spaceborne lidar data is calculated, and the spaceborne lidar data is calibrated based on the system deviation to obtain calibrated data with a unified scale. Based on the preprocessed multi-source auxiliary data, a feature set integrating multiple types of features is constructed, and the feature set is optimized and filtered to obtain the modeling input features; Based on the calibrated spaceborne lidar data, a fusion model is constructed that includes image feature extraction branch and table feature extraction branch. The modeling input features are input into the corresponding branches to complete feature extraction. After fusing the extracted features, a regression model is trained to generate a canopy height inversion model. The canopy height inversion model is applied to the target area, and a spatially continuous high-resolution canopy height product is generated through sliding window inference.

[0006] Secondly, embodiments of this application also provide a forest canopy height inversion device, the device comprising: The preprocessing module is used to acquire spaceborne lidar data, UAV lidar data, and multi-source auxiliary data, and to perform quality screening and preprocessing on various types of data to obtain basic data that meets the modeling requirements. The calibration module is used to calculate the system deviation of different spaceborne lidar data using the preprocessed UAV lidar data as a truth reference, and calibrate each spaceborne lidar data based on the system deviation to obtain calibrated data with a unified scale. The construction module is used to construct a feature set that integrates multiple types of features based on preprocessed multi-source auxiliary data, and to optimize and filter the feature set to obtain the modeling input features; The modeling module is used to construct a fusion model containing image feature extraction branch and table feature extraction branch based on the calibrated spaceborne lidar data. The modeling input features are input into the corresponding branches to complete feature extraction. After fusing the extracted features, the regression model is trained to generate the canopy height inversion model. The generation module is used to apply the canopy height inversion model to the target area and generate a spatially continuous high-resolution canopy height product through sliding window inference.

[0007] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the forest canopy height inversion method according to any one of the first aspects.

[0008] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the forest canopy height inversion method according to any one of the first aspects.

[0009] The embodiments of this application have the following beneficial effects: By integrating spaceborne lidar data, UAV lidar data, and multi-source auxiliary data, and combining targeted quality screening and preprocessing operations, the standardization and optimization of basic data for the inversion of cold-temperate forests were achieved. This effectively eliminated noise, redundancy, and low-quality data, solving the problems of sparse coverage and inconsistent quality of single spaceborne data, and providing reliable data support for subsequent steps. By using UAV lidar data as a ground truth reference, differentiated system bias calibration was performed on different spaceborne data, achieving a unified spaceborne data scale and accurately correcting systematic underestimation bias, reducing the overall bias from -1.15m to -0.35m, significantly improving the accuracy of core modeling data. By constructing a feature set that integrates multiple types of features and optimizing the selection process, a comprehensive characterization of the spatial heterogeneity of cold-temperate forests and the factors influencing canopy growth was achieved, avoiding the limitations of insufficient single feature dimensions and providing efficient and concise feature support for model input. Meanwhile, by constructing a dual-branch fusion model of imagery and tables, accurate extraction and deep fusion of different types of features were achieved. Compared with traditional single models, the generalization ability and prediction accuracy of canopy height inversion were significantly improved, with an inversion R² of 0.57, far exceeding the inversion effect of traditional machine learning models. Through sliding window inference technology, multi-source accuracy verification, and uncertainty quantification, a 10m high-resolution spatial continuous canopy height product was generated, taking into account the product's fineness, spatial integrity, and reliability. It can fully meet the needs of fine monitoring scenarios such as carbon storage estimation and resource surveys in cold temperate forests, forming a closed-loop optimization from data processing to product output. It is adapted to the complex terrain and forest characteristics of cold temperate zones and has strong practical application value. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating steps S101-S105 provided in the embodiments of this application; Figure 2 This is a schematic diagram illustrating the principle of forest canopy height inversion provided in an embodiment of this application; Figure 3 This is a schematic diagram of the forest canopy height inversion device provided in this application embodiment; Figure 4 This is a schematic diagram of the composition structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0013] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0014] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0015] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0016] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application and is not intended to limit this application.

[0018] See Figure 1 , Figure 1This is a flowchart illustrating steps S101-S105 of the forest canopy height inversion method provided in this application embodiment, which will be combined with... Figure 1 Steps S101-S105 are explained below.

[0019] In step S101, data from spaceborne lidar, UAV lidar, and multi-source auxiliary data are acquired. The data is then subjected to quality screening and preprocessing to obtain basic data that meets the modeling requirements. In step S102, the preprocessed UAV lidar data is used as the truth reference to calculate the system deviation of different spaceborne lidar data. Based on the system deviation, each spaceborne lidar data is calibrated to obtain calibrated data with a unified scale. In step S103, based on the preprocessed multi-source auxiliary data, a feature set integrating multiple types of features is constructed, and the feature set is optimized and filtered to obtain the modeling input features; In step S104, a fusion model containing image feature extraction branch and table feature extraction branch is constructed based on the calibrated spaceborne lidar data. The modeling input features are input into the corresponding branches to complete feature extraction. After fusing the extracted features, a regression model is trained to generate a canopy height inversion model. In step S105, the canopy height inversion model is applied to the target area, and a spatially continuous high-resolution canopy height product is generated through sliding window inference.

[0020] This application addresses the pain points of existing technologies in this region, such as the cold-temperate coniferous forests, complex terrain, and high biomass characteristics, including the cold-temperate forests of the Greater Khingan Mountains. It constructs a closed-loop technical system encompassing "data processing - bias calibration - feature construction - intelligent modeling - product generation," balancing inversion accuracy, spatial continuity, and ecological applicability. It should be noted that this application's embodiments are customized solutions deeply adapted to cold-temperate coniferous forests, complex terrain, and high biomass characteristics, while also reserving space for expanded analysis of ecological driving mechanisms, achieving the dual value of "inversion + application."

[0021] First, multi-source data acquisition and preprocessing are carried out. For example, taking the cold temperate forest of the Greater Khingan Mountains (117°–125°E, 41°–53°N) as the study area, the data system covers UAV-LiDAR ground truth data, GEDI and ICESat-2 satellite data, and multi-source auxiliary data. The preprocessing stage includes targeted screening rules (such as GEDI requiring quality_flag=1 and sensitivity>0.95, and ICESat-2 removing data with an elevation difference >50m) and spatial sparsity (60m KD-tree sampling) to remove noise and redundancy from the source, providing a standardized data foundation for subsequent stages and solving the problems of insufficient coverage and inconsistent quality of single satellite data.

[0022] Next, onboard data bias calibration was performed, using UAV-LiDAR data (0.1m resolution, RMSE=0.08m) as the core truth reference. This broke through the traditional unified calibration mode and addressed the differential bias (GEDI median bias) between the two types of onboard data. 0.74m, ICESat-2 Source shift correction was performed at 0.92m to adjust the overall deviation from... 1.15m down A 0.35m scale was used to unify the scale of spaceborne data, providing high-precision target variables for modeling and addressing the core pain point of systematically underestimating spaceborne data of cold-temperate forests. Then, multi-dimensional feature engineering was carried out, based on a 32-dimensional fusion feature set design, integrating multiple types of data such as spectral, SAR, topographic, and climatic data. Redundancy was removed through correlation analysis to avoid information leakage, providing sufficient feature support for adapting to the heterogeneity of cold-temperate forests.

[0023] Next, we performed multi-branch fusion modeling, using a CNN+MLP dual-branch structure to adapt to image and table features respectively. Combined with targeted training strategies (spatial grouping cross-validation, mixed precision training, etc.), the inversion performance was significantly improved (R²=0.57) compared to the traditional random forest model (R²=0.35), solving the problem of weak generalization ability of existing models in cold temperate forests.

[0024] Finally, product generation and validation are performed. A 10m resolution continuous product is generated through sliding window inference to meet the needs of fine estimation of regional carbon reserves. At the same time, multi-source validation (independent UAV-CHM, forest survey plots, and global product comparison) and uncertainty quantification are adopted to ensure product reliability.

[0025] In some embodiments, the spaceborne lidar data includes at least two types of spaceborne lidar sensor data with different orbital characteristics, the quality screening includes screening based on quality identifier, sensitivity threshold, elevation difference and canopy height range, and the preprocessing also includes spatial sparsification to eliminate spatial autocorrelation.

[0026] Here, GEDI (International Space Station orbit, inclination 51.6°) and ICESat-2 (global coverage, 3km gauge) data are selected. Their complementary orbital characteristics fill the coverage gap in the high-latitude region of northern Greater Khingan Mountains, solving the problems of sparse coverage and significant banding effects of single-satellite data, and improving the spatial density and uniformity of the training samples. This application's embodiments specifically eliminate low-quality data: quality labels and sensitivity thresholds ensure data reliability; elevation difference screening (ICESat-2 and SRTM DEM elevation difference >50m is eliminated) removes terrain interference; and canopy height range screening (<2m or >50m is eliminated) adapts to the canopy height distribution patterns of cold-temperate forests.

[0027] This application's embodiments eliminate spatial autocorrelation through spatial sparsity processing, such as the 60m KD-tree sampling operation. Because spaceborne LiDAR data is prone to spatial clustering in temperate forest regions, leading to sample redundancy and model overfitting, this processing can make the data evenly distributed in space, ensuring the objectivity and generalization ability of subsequent model training. It is especially suitable for the spatial heterogeneity problem caused by the topographic relief (200–2100m) of the Greater Khingan Mountains, laying a data foundation for accurate modeling.

[0028] In some embodiments, the true reference is generated by denoising and classifying ground points on UAV lidar point cloud data, and the true value of the canopy height corresponding to the footprint of the spaceborne lidar is determined by the height quantile method; the deviation calibration adopts a sensor source-differentiated translation correction method to eliminate the systematic underestimation deviation and scale inconsistency problem between different spaceborne sensors.

[0029] Here, the UAV-LiDAR 0.1m point cloud data is first denoised and ground point classified to generate a 1m resolution CHM (vertical accuracy RMSE=0.08m). The accuracy of this data far exceeds that of traditional ground surveys and airborne data, and can be used as a precise true value benchmark for the height of the cold temperate forest canopy. The "98th percentile (CHM_p98)" is used to determine the true value of the spaceborne footprint, which can effectively avoid the errors caused by the mixing of cold temperate forest canopy and local outliers, and ensure the accurate matching between the true value and the spaceborne LiDAR footprint.

[0030] The core innovation of this application's embodiments lies in sensor source-specific translation correction. The processing of the differential bias between GEDI and ICESat-2 includes: firstly, calculating the median systematic bias (GEDI) of the two types of satellite data relative to UAV-CHM. 0.74m, ICESat-2 (0.92m), and then a targeted constant translation correction is applied to generate canopy_debiased data with a unified scale. This method can effectively eliminate the systematic underestimation bias of the two types of sensors and solve the problem of inconsistent scales between sensors, so that the overall bias is reduced from... 1.15m down The height of the forest canopy is 0.35m, while maintaining the relative structure between samples, providing core target variables with consistent accuracy for subsequent modeling, and significantly improving the absolute accuracy of the canopy height inversion in cold temperate forests.

[0031] In some embodiments, the multi-type features include spectral features, texture features, radar polarization features, topographic factors, climate factors, forest attribute features, and spatial coordinate coding features; the optimization screening includes correlation analysis to remove redundancy and feature importance assessment screening, eliminating information leakage features and low contribution features.

[0032] Here, the embodiments of this application enrich feature dimensions and improve feature quality to adapt to the characteristics of complex terrain, mixed forest types, and strong environmental heterogeneity in cold and temperate forests, providing efficient input for subsequent deep learning models and solving the problems of coarse feature selection and poor adaptability in existing models.

[0033] Spectral features are derived from Sentinel-2 multispectral bands and vegetation indices such as NDVI and EVI, reflecting canopy growth and coverage; texture features (such as GLCM) supplement canopy spatial structure information; radar polarization features are derived from Sentinel-1 VV / VH data, and the ratio and difference calculations enhance the penetration perception of complex canopies; topographic factors are calculated based on ALOS PALSAR DEM (12.5m), including slope, aspect, and altitude, adapting to the topographic relief characteristics of the Greater Khingan Mountains; climatic factors are derived from 19 bioclimatic variables from WorldClim2.1, reflecting environmental conditions such as cold temperate temperature and precipitation; forest attribute features include 2020 China 30m forest age products and land cover types, providing prior information on canopy growth; spatial coordinate coding features compensate for the lack of location information and adapt to regional spatial heterogeneity.

[0034] This application employs a dual strategy of "correlation analysis for redundancy removal + feature importance assessment": first, correlation analysis is used to eliminate features with a correlation |r| ≥ 0.97 with the target variable, completely avoiding information leakage that could lead to a decline in the model's generalization ability; then, feature importance assessment is combined to remove low-contribution features, ultimately retaining concise and efficient modeling input features. This screening method reduces model computation and avoids overfitting, while ensuring that the retained features accurately characterize the core influencing factors of canopy height in cold-temperate forests, providing support for accurate model inversion.

[0035] In some embodiments, the spatial coordinate encoding features are constructed by normalizing latitude and longitude, combined with multinomial expansion and low-frequency Fourier encoding, to enhance the model's location awareness capability.

[0036] Here, this application implements a spatial coordinate encoding feature construction scheme, which addresses the characteristics of strong spatial heterogeneity in cold temperate forests (significant spatial differentiation in forest type, topography, and climate), and solves the problem of weak traditional feature location perception ability and inability to adapt to regional spatial patterns, thereby further improving the model's ability to characterize the spatial variation of canopy height in cold temperate forests.

[0037] Because the latitude and longitude values ​​in the Greater Khingan Mountains region span a large range, directly inputting them into the model would lead to an imbalance in the weights of location features and other features (such as vegetation index and slope). Therefore, latitude and longitude normalization is used to convert latitude and longitude to a uniform numerical range, ensuring that location features have equal weight and influence with other features, and avoiding interference from numerical differences in model training. Polynomial expansion, by constructing terms such as x², y², and xy, accurately captures the nonlinear locational correlations of latitude and longitude, adapting to the nonlinear spatial distribution of forest types and topography in cold temperate forests. Low-frequency Fourier encoding extracts regional-scale locational patterns through sin / cos terms, effectively avoiding high-frequency noise interference and focusing on the macro-environmental differences in different regions of the Greater Khingan Mountains (such as the northern high-latitude region and the southern low-latitude region). The combination of these two methods ultimately enhances the model's location awareness capability, enabling it to accurately identify differences in canopy height at different spatial locations, especially suitable for the complex terrain and mixed forest types of the Greater Khingan Mountains, solving the problem of insufficient inversion accuracy of traditional models in areas with strong spatial heterogeneity.

[0038] In some embodiments, the image feature extraction branch is constructed using a convolutional neural network, including an attention module and a generalized average pooling layer, for extracting spatial and semantic features of multi-channel remote sensing image patches; the table feature extraction branch is constructed using a multilayer perceptron, which extracts nonlinear correlation features of non-image features through fully connected layers and regularization mechanisms.

[0039] The image feature extraction branch is constructed using a convolutional neural network (CNN). The input is a 64×64 pixel multi-channel remote sensing image patch centered on the spaceborne LiDAR footprint (containing spectral, SAR, and topographic features). The network structure includes a Conv–BN–SiLU–SE attention module, 2×2 max pooling, and multiple sets of ConvBlocks. The output layer uses Gaussian weights to highlight the central region and employs generalized average pooling (GeM with a learnable exponent) to generate a 256-dimensional image representation. The SE attention module allows the model to focus on image areas that significantly affect the canopy height of cold temperate zones (such as dense coniferous forests and gently sloping terrain), avoiding interference from irrelevant areas. The Gaussian weight design aligns with the higher accuracy of the spaceborne LiDAR footprint center, further enhancing the targeting of feature extraction.

[0040] The tabular feature extraction branch is constructed using a multilayer perceptron (MLP) to adapt to the structured characteristics of non-image features (coordinate encoding, climate, forest age, slope, etc.). The network structure consists of two fully connected layers (512→256), coupled with the SiLU activation function and a Dropout regularization mechanism with a ratio of 0.3, generating a 256-dimensional tabular feature representation. The fully connected layers can deeply explore the nonlinear correlations between non-image features (such as the synergistic effect of forest age and annual mean temperature on canopy height), while the Dropout mechanism effectively suppresses overfitting, ensuring that the model still has stable feature extraction capabilities under limited sample data of cold-temperate forests. The two branches are adapted to different types of features, and complementary advantages are achieved through feature concatenation, providing high-quality fused features for subsequent regression models.

[0041] In some embodiments, the training strategy of the fusion model includes using a loss function that combines smooth L1 loss and L2 regularization, the AdamW optimizer, and spatial grouping cross-validation, while also incorporating data augmentation, mixed precision training, gradient pruning, and exponential moving average techniques to improve the model's generalization ability and training stability.

[0042] The loss function combines smoothed L1 loss (Huber loss) with L2 regularization. Smoothed L1 loss balances the advantages of ordinary L1 loss and L2 loss, avoiding the excessive influence of abnormally high / low canopy height samples in temperate forests on training, and effectively suppressing gradient explosion. The AdamW optimizer adds 1e-4 weight decay to the Adam optimizer to further suppress overfitting and adapt to multi-feature input scenarios. The validation method uses 10km grid GroupKFold 5-fold cross-validation to accurately avoid overfitting caused by spatial autocorrelation of remote sensing data, ensuring more accurate evaluation of the model's generalization ability and conforming to the spatial heterogeneity of the Greater Khingan Mountains region.

[0043] Data augmentation employs 90° rotation and horizontal / vertical flipping (training set only) to expand the sample size without altering the fundamental nature of the data, enhancing the model's adaptability to different forest types and terrain scenarios in temperate and cold-temperate forests. Mixed-precision training significantly improves training speed and reduces computational resource consumption while maintaining training accuracy, adapting to the training needs of multi-channel imagery and high-dimensional features. Gradient clipping (norm 1.0) effectively prevents gradient explosion during training, while exponential moving average of parameters (decay coefficient 0.999) makes model parameter updates smoother, improving the stability and inversion accuracy of the final model. This multi-strategy synergy ensures efficient and stable model training, ultimately achieving an inversion performance of R²=0.57, far exceeding traditional random forest models.

[0044] In some embodiments, the method further includes: using interpretability analysis to quantify the marginal contribution of each input feature to the canopy height prediction results, grouping and statistically analyzing by forest type, ecological zone and topographic gradient to reveal the ecological driving force of spatial variation in canopy height.

[0045] The core of this implementation is the quantification of marginal contributions using interpretability analysis methods. Addressing the "black box" nature of deep learning models, this approach breaks down the barrier of model ininterpretability, clarifying the impact of each input feature (such as forest age, aspect, and average annual temperature) on the prediction results of canopy height in cold-temperate forests. This not only traces the source of model inversion errors (e.g., errors in a certain region stemming from insufficient accuracy of climate factor data) but also provides precise data support for subsequent ecological pattern analysis. It can be flexibly selected according to actual needs, balancing basic inversion requirements (e.g., generating only canopy height products) with in-depth ecological research needs (e.g., revealing driving patterns). The Greater Khingan Mountains region has complex forest types (including coniferous forests, broad-leaved forests, and mixed forests), clear ecological zoning, and significant topographic gradient differences (altitude 200–2100m). The influencing factors and mechanisms of canopy height vary significantly under different scenarios, making overall statistical analysis prone to ambiguity. Grouped statistics can accurately capture the regular differences under different forest types, ecological zones, and topographic conditions, and ultimately reveal the core ecological driving laws of spatial variation in canopy height in cold temperate forests, providing a more comprehensive basis for decision-making in forest resource management and ecological restoration.

[0046] In some embodiments, the interpretability analysis method is the SHAP method, and the ecological driving laws include the nonlinear relationship between forest age and canopy height, the regulatory effect of slope aspect on canopy height, and the limiting effect of altitude on canopy height.

[0047] Here, the interpretability analysis method used in this application embodiment is the SHAP method (Shapley AdditiveExplanations), which has the advantages of quantifying the feature contribution of a single sample and ranking the global feature importance. Compared with other interpretability methods (such as LIME), it can more accurately and comprehensively characterize the correlation between features and canopy height prediction results, effectively support the quantitative analysis of marginal contributions, and is especially suitable for scenarios with the synergistic influence of multiple features in cold temperate forests, ensuring the reliability and accuracy of the analysis results.

[0048] The nonlinear relationship between forest age and canopy height reflects the characteristics of long growth cycles and age-dependent growth rates in cold-temperate coniferous forests, providing a basis for forest cultivation planning. The regulatory effect of slope aspect on canopy height is specifically manifested in the higher canopy height on shady / semi-shady slopes, adapting to the distribution characteristics of sunlight and moisture conditions in the high-latitude region of the Greater Khingan Mountains, revealing the mechanism by which topography influences canopy growth. The limiting effect of altitude on canopy height corresponds to the large altitude range of the Greater Khingan Mountains; low temperatures and insufficient air pressure at high altitudes inhibit canopy growth. These patterns not only explain the inversion results but also provide targeted guidance for cold-temperate forest cultivation and ecological restoration, further expanding the application scenarios of the technical solutions.

[0049] The embodiments of this application will now be explained in detail with reference to specific applications.

[0050] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the principle of forest canopy height inversion provided in an embodiment of this application, such as... Figure 2 As shown, this application focuses on the canopy height inversion problem in the cold-temperate forests of the Greater Khingan Mountains (117°–125°E, 41°–53°N), and provides solutions to address the multiple shortcomings of existing technologies in this region. In existing technologies, ground surveys are inefficient, costly, and prone to errors; optical remote sensing is susceptible to weather conditions and signal saturation; SAR technology lacks sufficient penetration into complex canopies; and while spaceborne LiDAR can acquire vertical structure information, it suffers from systematic underestimation, inter-sensor bias, and sparse spatial coverage. Furthermore, traditional machine learning models (such as random forests) have weak generalization ability in complex terrain and high biomass environments in cold-temperate zones (R² only 0.35). Existing canopy height products typically have a resolution of 30m or less, which is insufficient for detailed characterization and lacks a clear driving mechanism.

[0051] This application embodiment uses UAV-LiDAR to collect 0.1m resolution point cloud data in five typical sample areas, and processes it to generate a 1m resolution canopy height model (CHM) with RMSE=0.08m as a ground truth reference. GEDI L2A and ICESat-2 ATL08 data from 2021–2022 are downloaded and rigorously screened (e.g., GEDI requires quality_flag=1 and sensitivity>0.95, while ICESat-2 excludes anomalous elevation differences and extreme canopy height data), and spatial sparsity is applied to avoid autocorrelation. Multi-source auxiliary data such as Sentinel-1 / 2, ALOS PALSAR DEM, and WorldClim climate data are also collected to provide a foundation for subsequent modeling. Subsequently, using the 98th percentile height of the UAV-CHM as the ground truth, the median systematic bias (GEDI: 0.74m, ICESat-2: (0.92m) and perform translation correction to reduce the overall system deviation from 1.15m down 0.35m, effectively eliminating differences between sensors and improving absolute height accuracy.

[0052] In the feature engineering stage, this application constructs a 32-dimensional fused feature set including spectral, texture, SAR, topography, climate, forest attributes, and spatial coordinate encoding. Highly correlated features (|r|≥0.97) are removed to avoid information leakage. Simultaneously, RF feature importance pre-screening and SHAP interpretability analysis are combined to optimize variable combinations. In the modeling stage, a CNN+MLP multi-source fusion deep learning architecture is adopted: the image branch crops image patches centered on the spaceborne LiDAR footprint, generating a 256-dimensional image representation through a convolutional network containing an SE attention module and GeM pooling; the table feature branch processes non-image features through two fully connected layers, outputting a table representation of equal dimension. The two are then concatenated and regressed through a fully connected layer to regress canopy height. During training, the SmoothL1+L2 regularized loss function and AdamW optimizer are used, along with data augmentation, 10km grid GroupKFold 5-fold cross-validation, and other strategies to effectively avoid overfitting and spatial autocorrelation.

[0053] Finally, the trained model was applied to the entire study area, generating a 10m resolution, spatially continuous forest canopy height (FCH) product. This product underwent multi-source validation using independent UAV-CHM (RMSE=1.85m), forest inventory plot data, and existing global products, while quantifying pixel-level uncertainty. Furthermore, the SHAP method was used to analyze the environmental driving mechanism, and statistical analysis was performed by forest type, ecological zone, and topographic gradient. This revealed unique patterns in cold-temperate forests, such as a stronger positive correlation between coniferous forest height and age, and higher canopies on shady / semi-shady slopes, achieving an integrated assessment of "inversion + driving forces." This application not only improved the inversion performance to R²=0.57, compensating for the insufficient coverage of a single satellite-borne LiDAR, but also provided high-quality data support for fine-grained estimation of regional carbon storage, solving the core problems of weak applicability and insufficient product accuracy and resolution in existing technologies in cold-temperate high-latitude forest regions.

[0054] In summary, the embodiments of this application have the following beneficial effects: By employing a high-precision UAV-LiDAR ground truth reference and source-specific translation correction from spaceborne sensors, the systematic underestimation bias of spaceborne data was accurately corrected, reducing the overall bias from -1.15m to -0.35m and significantly improving the reliability of the basic data. Furthermore, by constructing a 32-dimensional multi-dimensional fusion feature set and combining it with a CNN+MLP dual-branch deep learning model, the spatial heterogeneity of temperate forests was accurately characterized, achieving an inversion R² of 0.57, significantly outperforming the generalization ability and inversion accuracy of traditional machine learning models (such as random forest R²=0.35). Finally, through sliding window inference technology, a 10m high-resolution spatial connection was achieved. The generation of continuous canopy height products balances product precision and spatial integrity, fully meeting the needs of precise monitoring such as carbon storage estimation in cold-temperate forests. By adding an environmental-driven mechanism analysis step based on the SHAP method, a breakthrough in the "black box" characteristics of deep learning models is achieved, extending the embodiments of this application from simple canopy height inversion to the revelation of ecological laws, achieving the dual value of "inversion + application". Through customized design for the characteristics of cold-temperate forests such as complex terrain, concentrated forest types, and high biomass, it achieves efficient adaptation to typical areas such as the Greater Khingan Mountains, while also having the ability to flexibly migrate to other forest types, thus having a wider range of applications.

[0055] Based on the same inventive concept, this application also provides a forest canopy height inversion device corresponding to the forest canopy height inversion method in the first embodiment. Since the principle of the device in this application is similar to the above-mentioned forest canopy height inversion method, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0056] like Figure 3 As shown, Figure 3 This is a schematic diagram of the forest canopy height inversion device 300 provided in this application embodiment. The forest canopy height inversion device 300 includes: The preprocessing module 301 is used to acquire spaceborne lidar data, UAV lidar data and multi-source auxiliary data, and to perform quality screening and preprocessing on various types of data to obtain basic data that meets the modeling requirements. The calibration module 302 is used to calculate the system deviation of different spaceborne lidar data using the preprocessed UAV lidar data as a truth reference, and to calibrate each spaceborne lidar data based on the system deviation to obtain calibrated data with a unified scale. The construction module 303 is used to construct a feature set that integrates multiple types of features based on preprocessed multi-source auxiliary data, and to optimize and filter the feature set to obtain the modeling input features; The modeling module 304 is used to construct a fusion model containing image feature extraction branch and table feature extraction branch based on the calibrated spaceborne lidar data. The modeling input features are input into the corresponding branches to complete feature extraction. After fusing the extracted features, the regression model is trained to generate the canopy height inversion model. The generation module 305 is used to apply the canopy height inversion model to the target area and generate a spatially continuous high-resolution canopy height product through sliding window inference.

[0057] Those skilled in the art should understand that Figure 3 The functions of each unit in the forest canopy height inversion device 300 shown can be understood by referring to the relevant description of the aforementioned forest canopy height inversion method. Figure 3 The functions of each unit in the forest canopy height inversion device 300 shown can be realized by a program running on a processor or by specific logic circuits.

[0058] In one possible implementation, the spaceborne lidar data includes at least two types of spaceborne lidar sensor data with different orbital characteristics, the quality screening includes screening based on quality identifiers, sensitivity thresholds, elevation differences, and canopy height ranges, and the preprocessing further includes spatial sparsification to eliminate spatial autocorrelation.

[0059] In one possible implementation, the true reference is generated by denoising and classifying ground points from UAV lidar point cloud data, and the true value of the canopy height corresponding to the footprint of the spaceborne lidar is determined by the height quantile method; the deviation calibration adopts a sensor source-differentiated translation correction method to eliminate the systematic underestimation deviation and scale inconsistency problem between different spaceborne sensors.

[0060] In one possible implementation, the multi-type features include spectral features, texture features, radar polarization features, topographic factors, climate factors, forest attribute features, and spatial coordinate coding features; the optimization screening includes correlation analysis to remove redundancy and feature importance assessment screening, eliminating information leakage features and low contribution features.

[0061] In one possible implementation, the spatial coordinate encoding features are constructed by normalizing latitude and longitude, combined with multinomial expansion and low-frequency Fourier encoding, to enhance the model's location awareness capability.

[0062] In one possible implementation, the image feature extraction branch is constructed using a convolutional neural network, including an attention module and a generalized average pooling layer, for extracting spatial and semantic features of multi-channel remote sensing image patches; the table feature extraction branch is constructed using a multilayer perceptron, which extracts nonlinear correlation features of non-image features through fully connected layers and regularization mechanisms.

[0063] In one possible implementation, the training strategy of the fusion model includes using a loss function that combines smooth L1 loss and L2 regularization, the AdamW optimizer, and spatial grouping cross-validation, while also incorporating data augmentation, mixed precision training, gradient pruning, and exponential moving average techniques to improve the model's generalization ability and training stability.

[0064] In one possible implementation, the generation module is also used to: quantify the marginal contribution of each input feature to the canopy height prediction results using interpretability analysis methods, group and statistically analyze the data by forest type, ecological zone and topographic gradient, and reveal the ecological driving force of spatial variation in canopy height.

[0065] In one possible implementation, the interpretability analysis method is the SHAP method, and the ecological driving laws include the nonlinear relationship between forest age and canopy height, the regulatory effect of slope aspect on canopy height, and the limiting effect of altitude on canopy height.

[0066] The aforementioned forest canopy height inversion device, through high-precision UAV-LiDAR ground truth reference and source-specific translation correction of spaceborne sensors, accurately corrected the systematic underestimation bias of spaceborne data, reducing the overall bias from -1.15m to -0.35m, significantly improving the reliability of the basic data. By constructing a 32-dimensional multidimensional fusion feature set and combining it with a CNN+MLP dual-branch deep learning model, it achieved accurate characterization of the spatial heterogeneity of cold-temperate forests, with an inversion R² of 0.57, significantly outperforming the generalization ability and inversion accuracy of traditional machine learning models (such as random forest R²=0.35). Through sliding window inference technology, it achieved a height of 10m. The generation of high-resolution spatial continuous canopy height products balances product precision and spatial integrity, fully meeting the needs of precise monitoring such as carbon storage estimation in cold-temperate forests. By adding an environmental-driven mechanism analysis step based on the SHAP method, a breakthrough in the "black box" characteristics of deep learning models is achieved, extending the embodiments of this application from simple canopy height inversion to the revelation of ecological laws, achieving the dual value of "inversion + application". Through customized design for the characteristics of cold-temperate forests such as complex terrain, concentrated forest types, and high biomass, efficient adaptation to typical areas such as the Greater Khingan Mountains is achieved, while also having the ability to flexibly migrate to other forest types, making it more widely applicable.

[0067] like Figure 4 As shown, Figure 4 This is a schematic diagram of the composition structure of the electronic device 400 provided in the embodiments of this application. The electronic device 400 includes: The device includes a processor 401, a storage medium 402, and a bus 403. The storage medium 402 stores machine-readable instructions that can be executed by the processor 401. When the electronic device 400 is running, the processor 401 communicates with the storage medium 402 via the bus 403. The processor 401 executes the machine-readable instructions to perform the steps of the forest canopy height inversion method described in this application embodiment.

[0068] In practical applications, the various components in the electronic device 400 are coupled together via a bus 403. It is understood that the bus 403 is used to achieve communication between these components. In addition to a data bus, the bus 403 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 4 The general designated all buses as Bus 403.

[0069] The aforementioned electronic equipment, through high-precision UAV-LiDAR ground truth reference and source-specific translation correction of spaceborne sensors, achieved accurate correction of the systematic underestimation bias of spaceborne data, reducing the overall bias from -1.15m to -0.35m, significantly improving the reliability of the basic data. By constructing a 32-dimensional multidimensional fusion feature set and combining it with a CNN+MLP dual-branch deep learning model, it achieved accurate characterization of the spatial heterogeneity of temperate forests, with an inversion R² of 0.57, significantly outperforming the generalization ability and inversion accuracy of traditional machine learning models (such as random forest R²=0.35). Through sliding window inference technology, a 10m high resolution was achieved. The generation of spatially continuous canopy height products balances product precision and spatial integrity, fully meeting the needs of precise monitoring such as carbon storage estimation in cold-temperate forests. By adding an environmental-driven mechanism analysis step based on the SHAP method, a breakthrough in the "black box" characteristics of deep learning models is achieved, extending the embodiments of this application from simple canopy height inversion to the revelation of ecological laws, achieving the dual value of "inversion + application". Through customized design for the characteristics of cold-temperate forests such as complex terrain, concentrated forest types, and high biomass, it achieves efficient adaptation to typical areas such as the Greater Khingan Mountains, while also having the ability to flexibly migrate to other forest types, thus broadening its applicability.

[0070] This application also provides a computer-readable storage medium storing executable instructions that, when executed by at least one processor 401, implement the forest canopy height inversion method described in this application.

[0071] In some embodiments, the storage medium may be a magnetic random access memory (FRAM), a read-only memory (ROM), or a programmable read-only memory (PROM). Erasable Programmable Read-Only Memory (EPROM) Electrically Erasable Programmable Read-Only Memory (EEPROM) Read-only memory, flash memory, magnetic surface storage, optical disc, or CD-ROM ROM, Compact Disc Read It can be a memory such as a memory only; or it can be a device that includes one or any combination of the above-mentioned memories.

[0072] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0073] As an example, executable instructions may, but do not necessarily, correspond to files in the file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0074] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0075] The aforementioned computer-readable storage medium, through UAV-LiDAR high-precision truth reference and source-specific translation correction of spaceborne sensors, achieved accurate correction of the systematic underestimation bias of spaceborne data, reducing the overall bias from -1.15m to -0.35m, significantly improving the reliability of the basic data. By constructing a 32-dimensional multi-dimensional fusion feature set and combining it with a CNN+MLP dual-branch deep learning model, it achieved accurate characterization of the spatial heterogeneity of cold-temperate forests, with an inversion R² of 0.57, significantly outperforming the generalization ability and inversion accuracy of traditional machine learning models (such as random forest R²=0.35). Through sliding window inference technology, it achieved 10m high resolution. The generation of high-resolution spatial continuous canopy height products balances product precision and spatial integrity, fully meeting the needs of precise monitoring such as carbon storage estimation in cold-temperate forests. By adding an environmental-driven mechanism analysis step based on the SHAP method, a breakthrough in the "black box" characteristics of deep learning models is achieved, extending the embodiments of this application from simple canopy height inversion to the revelation of ecological laws, achieving the dual value of "inversion + application". Through customized design for the characteristics of cold-temperate forests such as complex terrain, concentrated forest types, and high biomass, it achieves efficient adaptation to typical areas such as the Greater Khingan Mountains, while also having the ability to flexibly migrate to other forest types, thus broadening its applicability.

[0076] In the several embodiments provided in this application, it should be understood that the disclosed methods and electronic devices can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0077] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0078] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0079] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0080] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for inverting forest canopy height, characterized in that, Includes the following steps: Acquire spaceborne lidar data, UAV lidar data, and multi-source auxiliary data; perform quality screening and preprocessing on various types of data to obtain basic data that meets the modeling requirements. Using the preprocessed UAV lidar data as a truth reference, the system deviation of different spaceborne lidar data is calculated, and the spaceborne lidar data is calibrated based on the system deviation to obtain calibrated data with a unified scale. Based on the preprocessed multi-source auxiliary data, a feature set integrating multiple types of features is constructed, and the feature set is optimized and filtered to obtain the modeling input features; Based on the calibrated spaceborne lidar data, a fusion model is constructed that includes image feature extraction branch and table feature extraction branch. The modeling input features are input into the corresponding branches to complete feature extraction. After fusing the extracted features, a regression model is trained to generate a canopy height inversion model. The canopy height inversion model is applied to the target area, and a spatially continuous high-resolution canopy height product is generated through sliding window inference.

2. The method according to claim 1, characterized in that, The spaceborne lidar data includes at least two types of spaceborne lidar sensor data with different orbital characteristics. The quality screening includes screening based on quality identifiers, sensitivity thresholds, elevation differences, and canopy height ranges. The preprocessing also includes spatial sparsification to eliminate spatial autocorrelation.

3. The method according to claim 1, characterized in that, The true reference is generated by denoising and classifying ground points from UAV lidar point cloud data, and the true value of the canopy height corresponding to the footprint of the spaceborne lidar is determined by the height quantile method. The deviation calibration adopts a sensor source-differentiated translation correction method to eliminate the systematic underestimation deviation and scale inconsistency problem between different spaceborne sensors.

4. The method according to claim 1, characterized in that, The multi-type features include spectral features, texture features, radar polarization features, topographic factors, climate factors, forest attribute features, and spatial coordinate coding features; the optimization screening includes correlation analysis to remove redundancy and feature importance assessment screening, eliminating information leakage features and low contribution features.

5. The method according to claim 4, characterized in that, The spatial coordinate encoding features are constructed by normalizing latitude and longitude, combined with multinomial expansion and low-frequency Fourier encoding, to enhance the model's location awareness capability.

6. The method according to claim 1, characterized in that, The image feature extraction branch is constructed using a convolutional neural network, including an attention module and a generalized average pooling layer, to extract spatial and semantic features of multi-channel remote sensing image patches; the table feature extraction branch is constructed using a multilayer perceptron, which extracts nonlinear correlation features of non-image features through fully connected layers and regularization mechanisms.

7. The method according to claim 6, characterized in that, The training strategy of the fusion model includes using a loss function that combines smooth L1 loss and L2 regularization, the AdamW optimizer, and spatial grouping cross-validation. It also incorporates data augmentation, mixed precision training, gradient pruning, and exponential moving average techniques to improve the model's generalization ability and training stability.

8. The method according to claim 1, characterized in that, The method further includes: using interpretability analysis to quantify the marginal contribution of each input feature to the canopy height prediction results, and grouping statistics by forest type, ecological zone and topographic gradient to reveal the ecological driving force of spatial variation in canopy height.

9. The method according to claim 8, characterized in that, The interpretability analysis method is the SHAP method, and the ecological driving laws include the nonlinear relationship between forest age and canopy height, the regulatory effect of slope aspect on canopy height, and the limiting effect of altitude on canopy height.

10. A forest canopy height inversion device, characterized in that, The device includes: The preprocessing module is used to acquire spaceborne lidar data, UAV lidar data, and multi-source auxiliary data, and to perform quality screening and preprocessing on various types of data to obtain basic data that meets the modeling requirements. The calibration module is used to calculate the system deviation of different spaceborne lidar data using the preprocessed UAV lidar data as a truth reference, and calibrate each spaceborne lidar data based on the system deviation to obtain calibrated data with a unified scale. The construction module is used to construct a feature set that integrates multiple types of features based on preprocessed multi-source auxiliary data, and to optimize and filter the feature set to obtain the modeling input features; The modeling module is used to construct a fusion model containing image feature extraction branch and table feature extraction branch based on the calibrated spaceborne lidar data. The modeling input features are input into the corresponding branches to complete feature extraction. After fusing the extracted features, the regression model is trained to generate the canopy height inversion model. The generation module is used to apply the canopy height inversion model to the target area and generate a spatially continuous high-resolution canopy height product through sliding window inference.

Images