Method, device, medium and product for estimating forest biomass from airborne lidar point clouds
By processing airborne lidar point cloud data through a target deep learning network and combining global structural feature encoding and vertical structure hierarchical perception modules, the problem of low accuracy in forest biomass estimation is solved, and more efficient and accurate biomass prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- RES INST OF FOREST RESOURCE INFORMATION TECHN CHINESE ACADEMY OF FORESTRY
- Filing Date
- 2025-10-30
- Publication Date
- 2026-06-12
Smart Images
Figure CN121348276B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of forest biomass monitoring, and in particular to a method, equipment, medium, and product for estimating forest biomass using airborne lidar point cloud. Background Technology
[0002] Forest biomass (AGB) is an important indicator reflecting the status of forest resources and carbon storage. Accurate estimation of forest biomass is crucial for understanding and monitoring carbon sinks in forest ecosystems and assessing forest resource status. LiDAR (Light Detection and Ranging) technology is widely used in remote sensing because its laser pulses can penetrate the forest canopy and obtain detailed three-dimensional structural information. When LiDAR is equipped on a manned aircraft, it is usually called airborne lidar (ALS).
[0003] Area-based estimation (ABA) is an important method for biomass estimation using ALS data because it remains reliable even at low point densities, thus reducing data acquisition costs. However, it also suffers from cumbersome procedures and lower accuracy due to the loss of spatial features. The development of point cloud deep learning has provided a new direction for estimating forest biomass, but how to effectively apply deep neural networks to achieve this estimation remains a problem to be solved. Summary of the Invention
[0004] The purpose of this application is to provide a method, equipment, medium, and product for estimating forest biomass using airborne lidar point cloud data, which can achieve relatively accurate estimation of forest biomass.
[0005] To achieve the above objectives, this application provides the following solution:
[0006] Firstly, this application provides a method for estimating forest biomass using airborne lidar point clouds, including:
[0007] Acquire target airborne lidar data, wherein the target airborne lidar data is a three-dimensional point cloud obtained by the airborne lidar scanning the target area;
[0008] The target airborne lidar data is input into the trained target deep learning network, and the predicted value of forest biomass in the target area is obtained through the target deep learning network.
[0009] The target deep learning network includes an input layer, a backbone network, and an output layer stacked in sequence. The backbone network includes a global feature extraction layer and multiple local feature extraction layers stacked in sequence. The global feature extraction layer includes a global structural feature encoding module and a vertical structural hierarchical perception module stacked in sequence.
[0010] The global structural feature encoding module is configured to extract key global features of the 3D point cloud by performing feature extraction on the output of the input layer.
[0011] The vertical structure layered perception module is configured to: perform vertical layered perception on the key global features based on the relative height relationship of the three-dimensional point cloud to obtain perception features as input to the first local feature extraction layer.
[0012] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the airborne lidar point cloud forest biomass estimation method described in any one of the above.
[0013] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the airborne lidar point cloud forest biomass estimation method described above.
[0014] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the airborne lidar point cloud forest biomass estimation method described above.
[0015] According to the specific embodiments provided in this application, the following technical effects are disclosed:
[0016] This application provides a method, device, medium, and product for estimating forest biomass using airborne lidar point clouds. It predicts forest biomass in a target area based on airborne lidar data using a target deep learning network. Unlike area-based estimation methods that require compressing the three-dimensional structure into manually designed one-dimensional features, the deep learning network can directly process the original three-dimensional point cloud without dimensionality reduction, fully utilizing the rich spatial information preserved in the three-dimensional point cloud, thus improving the accuracy of forest biomass prediction. The target deep learning network innovatively integrates a global structural feature encoding module and a vertical structure hierarchical perception module on a high-resolution feature map, aiming to capture more complete structural information and provide global guidance for subsequent local feature extraction. The global structural feature encoding module can extract key global features of the three-dimensional point cloud from the preliminary features extracted from the input layer, while the vertical structure hierarchical perception module can learn the differentiated features of the three-dimensional point cloud in the vertical height dimension, thereby more accurately predicting forest biomass. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is an application environment diagram of an airborne lidar point cloud forest biomass estimation method according to one embodiment of this application;
[0019] Figure 2 A flowchart illustrating a method for estimating forest biomass using airborne lidar point clouds, provided as an embodiment of this application;
[0020] Figure 3 An architecture diagram of a target deep learning network provided in an embodiment of this application;
[0021] Figure 4 An architecture diagram of a global structural feature encoding module provided in an embodiment of this application;
[0022] Figure 5 An architecture diagram of a vertical structure hierarchical sensing module provided in an embodiment of this application;
[0023] Figure 6 This is a comparison chart of the deterministic coefficients of different network models in a verification example of this application;
[0024] Figure 7 This is a comparison chart of the root mean square error of different network models in a verification example of this application;
[0025] Figure 8 This is a comparison chart of the average deviation of different network models in a verification example of this application;
[0026] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0028] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0029] The airborne lidar point cloud forest biomass estimation method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on another server. Terminal 101 can send target airborne LiDAR data to server 102. Server 102 receives the target airborne LiDAR data and inputs it into a trained target deep learning network to obtain a predicted value of forest biomass in the target area. Server 102 can then feed back the predicted value of forest biomass in the target area to terminal 101. Furthermore, in some embodiments, the airborne LiDAR point cloud forest biomass estimation method can also be implemented independently by server 102 or terminal 101. For example, terminal 101 can directly process the target airborne LiDAR data, or server 102 can obtain the target airborne LiDAR data from the data storage system.
[0030] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, and tablets. The server 102 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.
[0031] In one specific implementation, such as Figure 2 As shown, an airborne lidar point cloud forest biomass estimation method is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this specific embodiment, the method is applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps 210 and 220.
[0032] Step 210: Obtain target airborne lidar data, which is a three-dimensional point cloud obtained by the airborne lidar scanning the target area.
[0033] Step 220: Input the target airborne lidar data into the trained target deep learning network, and obtain the predicted value of forest biomass in the target area through the target deep learning network.
[0034] The target deep learning network comprises an input layer, a backbone network, and an output layer stacked sequentially. The backbone network comprises a feature extraction layer and multiple downsampling layers stacked sequentially. The feature extraction layer comprises a global structure feature encoding module and a vertical structure layered perception module stacked sequentially. The global structure feature encoding module is configured to extract key global features of the 3D point cloud from the output of the input layer. The vertical structure layered perception module is configured to perform vertical layered perception on the key global features based on the relative height relationship of the 3D point cloud to obtain perceptual features as the input of the first downsampling layer.
[0035] The output layer performs global summation pooling on the final feature map (the output of the backbone network), and then outputs an estimated value of biomass through two layers of multilayer perceptrons.
[0036] It should be noted that stacking in sequence means that the output of the former is the input of the latter.
[0037] In the specific implementation described above, a target deep learning network is used to predict forest biomass in a target area based on airborne lidar data. Unlike area-based estimation methods that require compressing the three-dimensional structure into manually designed one-dimensional features, deep learning networks can directly process the original three-dimensional point cloud without dimensionality reduction. This allows for full utilization of the rich spatial information preserved in the three-dimensional point cloud, thereby improving the accuracy of forest biomass prediction.
[0038] More importantly, the above specific implementation provides a concrete deep learning network architecture. The target deep learning network innovatively integrates a global structural feature encoding module and a vertical structural hierarchical perception module on high-resolution feature maps, aiming to capture more complete structural information and provide global guidance for subsequent local feature extraction.
[0039] Specifically, the input layer first performs preliminary feature extraction on the input 3D point cloud and inputs the results into the global structural feature encoding module. This module can then extract key global features from the preliminary features extracted by the input layer. These key global features are the features of global key points within the 3D point cloud. Learnable parameters for determining global key points are introduced into the global structural feature encoding module. During model training, optimizing these learnable parameters improves the module's ability to capture global key points, resulting in key global features with a greater predictive contribution. Subsequently, the vertical structure hierarchical perception module performs hierarchical perception of the key global features, rather than directly perceiving the key global features as a whole. Since the features of different points in the 3D point cloud have different ecological meanings—for example, the upper layer (features of points with greater height) reflects differences in tree height, while the lower layer (features of points with smaller height) reflects variations in lower-level vegetation—hierarchical perception can learn differentiated features at each layer, thus enabling more accurate prediction of forest biomass.
[0040] As explained above, the target deep learning network provided in the specific implementation method innovatively combines a global structural feature encoding module and a vertical structural hierarchical perception module as the global feature extraction layer. The vertical structural hierarchical perception module can perform hierarchical perception on the key global features extracted by the global structural feature encoding module to learn the feature differences at different heights. Moreover, the feature hierarchical division is determined based on the relative height relationship of the 3D point cloud, that is, using height quantiles as thresholds for adaptive hierarchical division. For example, one-third and two-thirds of the height (relative to the total height of the 3D point cloud) are used as hierarchical thresholds to determine three height intervals, thereby dividing the key global features into three hierarchical features. Compared with using a fixed height threshold, it has advantages such as parameter exemption, greater robustness to outliers and missing data, and the ability to adapt to differences in tree height in different plots / stands, thus achieving higher biomass estimation results for different types of target areas.
[0041] In one embodiment, the feature extraction expression of the global structural feature encoding module is:
[0042]
[0043]
[0044]
[0045]
[0046] in, This represents the coordinates of the i-th point cloud. This represents the features of the i-th point cloud provided by the input layer. This represents the input to the global structural feature encoding module. This represents the m-th learnable parameter. Indicates by The key spatial locations (key query points) obtained through learning. express The characteristics of the place, The feature set representing all key spatial locations, , and Both represent linear transformation layers, used to perform linear transformations on their respective input features to increase dimensionality, for example... Represents a linear transformation layer right Perform a linear transformation to increase dimensionality. Represents key global features.
[0047] Specifically, the global structural feature encoding module mainly performs key query operations. In the global structural feature encoding module, let the 3D point cloud coordinates be... The corresponding point-level features (obtained from the input layer) are: A set of key query points (3D points located in key spatial positions) can be defined to summarize the most representative spatial-semantic patterns while compressing redundancy. Let the spatial position of the key query point be denoted as... The characteristics of key query points are: Introducing learnable vectors The spatial location and features of key query points are adaptively learned. The spatial location of key query points is adaptively estimated through semantic similarity with the original 3D point cloud features (features of the point cloud output from the input layer), ensuring that they fall within semantically important regions and cover representative spatial distributions, resulting in sparse but information-rich key spatial locations (key 3D points):
[0048]
[0049] in, v m It is a set of defined learnable vectors that are compared with the features of the point cloud. Perform the dot product, then... softmax The similarity weights are obtained after function normalization, and these are used as weights for the original 3D point cloud features, so that they fall into semantically important regions.
[0050] To enhance representativeness, key queries also incorporate spatial and semantic information from the neighborhood to update features:
[0051]
[0052]
[0053]
[0054] in, Represents the spatial weighting function. This represents the semantic weight function. In order to take into account the characteristics of spatial structure updates, The latter part is the feature update that takes into account semantic information. The spatial structure feature update is reflected in the introduction of point cloud coordinate information, while the semantic information feature update is reflected in the introduction of high-level semantic features learned by the network.
[0055] Spatial weighting focuses key queries on the geometric structure of spatial neighbors, while semantic weighting emphasizes semantically relevant points (capturing long-range features), thus obtaining local-global features with a global structure awareness. Using sparse yet information-dense key queries as proxies, global interactions are performed with three sets of latent vectors, calculating affinity to selectively exchange information, and then fused with the original input via residual paths.
[0056]
[0057]
[0058] In the above embodiments, a specific design of a global structural feature encoding module is provided. By using this global structural feature encoding module, sparse but information-rich key global features can be extracted from the output features of the input layer. While retaining rich information, the amount of subsequent feature calculation is reduced, which means that the prediction accuracy and prediction efficiency of the target deep learning network are guaranteed at the same time.
[0059] In one embodiment, the vertical layered perception process of the vertical structure layered perception module includes: dividing key global features into multiple hierarchical features belonging to different height intervals using height quantiles; encoding height information for each point feature in each hierarchical feature; performing feature perception on each hierarchical feature after height information encoding to obtain multiple perception results, and fusing the multiple perception results to obtain a fusion result; and performing feature perception on the fusion result to obtain perception features. This step involves reweighting and aggregating the feature information of different height intervals from a global vertical perspective.
[0060] In this embodiment, the vertical structure layering perception module can be used to implement adaptive vertical layering.
[0061] Based on the relative height relationships of 3D point clouds, height quantiles are used as thresholds for adaptive layering. When the 3D point cloud can fully describe the vertical structure of the target region, this layering method can effectively distinguish different layer structural features of the target region, such as upper, middle, and lower layer structural features (corresponding to different hierarchical features).
[0062] After vertical layering is completed, height and position encoding is performed.
[0063] Different layers have different ecological meanings (for example, the upper structure reflects differences in tree height, while the lower structure reflects variations in the vegetation below). Therefore, hierarchical perception is needed to learn differentiated features at each layer while maintaining awareness of its vertical hierarchical position. Height information is encoded to represent the vertical layer to which each 3D point belongs (different vertical layers correspond to different height ranges) and its relative position within that vertical layer. The lightweight multilayer perceptron maps this information to a high-dimensional embedding, helping the model understand the order and context within and between layers.
[0064] The feature perception process for each hierarchical feature is as follows: the hierarchical feature is normalized and the first intermediate feature is obtained through gating adaptive selection; the first intermediate feature is subjected to two-layer linear transformation and single-injection kernel mapping to obtain the second intermediate feature, realizing adaptive weighting based on similarity; the first intermediate feature is subjected to one-layer linear transformation to obtain the third intermediate feature; the second and third intermediate features are dynamically weighted through a forget gate, and then the hierarchical features are merged through a linear transformation and residual connection to obtain the fourth intermediate feature; the normalized fourth intermediate feature is perceived through a multilayer perceptron to obtain the fifth intermediate feature, and the fifth and fourth intermediate features are merged through residual connection to obtain the perception result.
[0065] In the above embodiments, a specific structure of a vertical structure hierarchical perception module is provided, which guides feature learning through fine-grained vertical representation.
[0066] In one embodiment, the input layer includes a first sparse 3D convolution and a saliency aggregation module stacked sequentially, with the output of the saliency aggregation module serving as the output of the input layer. The first sparse 3D convolution has a size of 7*7*7. The input layer employs sparse convolution with large kernels in conjunction with saliency aggregation to integrate neighborhood context within a larger receptive field, while suppressing isolated noise and enhancing informative structures such as the canopy.
[0067] In one embodiment, there are four local feature extraction layers, each including a downsampling layer. The global feature extraction layer and multiple downsampling layers are all residually connected to the input and output through a local feature aggregation module. The local feature aggregation includes a second sparse 3D convolution, a first normalization layer, an activation function, a third sparse 3D convolution, and a second normalization layer stacked sequentially. The second and third sparse 3D convolutions are both 3x3x3 in size. In this embodiment, the backbone network adopts a five-layer hierarchical structure (one global feature extraction layer + four local feature extraction layers), progressively downsampling to expand the receptive field, and using 3×3×3 sparse convolutions to efficiently aggregate local features. The local aggregation uses a residual network-style bottleneck design to enhance depth and robustness.
[0068] It should be noted that the various embodiments described above can be combined to form new embodiments. For example, in one embodiment, the airborne lidar point cloud forest biomass estimation method includes:
[0069] Acquire target airborne lidar data, which is a three-dimensional point cloud obtained by the airborne lidar scanning the target area.
[0070] The target airborne lidar data is input into the trained target deep learning network, and the predicted value of forest biomass in the target area is obtained through the target deep learning network.
[0071] The target deep learning network includes an input layer, a backbone network, and an output layer stacked sequentially. The backbone network includes a global feature extraction layer and multiple local feature extraction layers stacked sequentially. The global feature extraction layer includes a global structural feature encoding module and a vertical structural hierarchical perception module stacked sequentially.
[0072] The input layer consists of a first sparse 3D convolution and a saliency aggregation module stacked sequentially. The output of the saliency aggregation module is the output of the input layer. The size of the first sparse 3D convolution is 7*7*7.
[0073] There are four local feature extraction layers, each including a downsampling layer. The global feature extraction layer and multiple downsampling layers all use a local feature aggregation module to perform residual connections between the input and output. The local feature aggregation module consists of a second sparse 3D convolution, a first normalization layer, an activation function, a third sparse 3D convolution, and a second normalization layer, stacked sequentially. The second and third sparse 3D convolutions are both 3*3*3 in size.
[0074] The global structural feature encoding module is configured to extract key global features of the 3D point cloud from the output of the input layer; the feature extraction expression of the global structural feature encoding module is:
[0075]
[0076]
[0077]
[0078]
[0079] in, This represents the coordinates of the i-th point cloud. This represents the features of the i-th point cloud provided by the input layer. This represents the m-th learnable parameter. Indicates by The key spatial locations obtained through learning express The characteristics of the place, The feature set representing all key spatial locations, , and All represent weights. Represents key global features.
[0080] The vertical structure hierarchical perception module is configured to: perform vertical hierarchical perception on key global features based on the relative height relationship of the 3D point cloud to obtain perceptual features, which serve as input to the first local feature extraction layer. The vertical hierarchical perception process of the module includes: dividing the key global features into multiple hierarchical features belonging to different height intervals using height quantiles; encoding height information for each point feature in each hierarchical feature; performing feature perception on each hierarchical feature after height information encoding to obtain multiple perception results; fusing the multiple perception results to obtain a fused result; and performing feature perception on the fused result to obtain the perceptual features. The feature perception process for each hierarchical feature is as follows: the hierarchical feature is normalized and a first intermediate feature is obtained through gating adaptive selection; the first intermediate feature is subjected to two-layer linear transformation and single-injection kernel mapping to obtain the second intermediate feature; the first intermediate feature is subjected to one-layer linear transformation to obtain the third intermediate feature; the second and third intermediate features are dynamically weighted through a forget gate, and then a first-layer linear transformation is performed before the hierarchical features are merged through residual connection to obtain the fourth intermediate feature; the normalized fourth intermediate feature is perceived through a multilayer perceptron to obtain the fifth intermediate feature, and the fifth and fourth intermediate features are merged through residual connection to obtain the perception result.
[0081] In this embodiment, forest biomass estimation is performed by jointly learning local features, global context, and vertical structure information. The input layer employs sparse convolutions with large kernels combined with saliency aggregation to integrate neighborhood context within a larger receptive field, while suppressing isolated noise and enhancing informative structures such as the canopy. The backbone network uses a five-layer hierarchical structure, progressively downsampling to expand the receptive field, and efficiently aggregating local features using 3×3×3 sparse convolutions; local aggregation employs a residual network-style bottleneck design to enhance depth and robustness. To further enhance the feature characterization of forest scene point clouds, a global structure feature encoding module and a vertical structure hierarchical perception module are integrated, allowing them to operate on high-resolution feature maps, capturing more complete structures and providing global guidance for subsequent local extraction. The output layer performs global summation pooling on the final feature map, and then outputs the biomass estimate after passing through two layers of multilayer perceptrons.
[0082] As can be seen from the above embodiments, the focus of this invention is to achieve forest biomass estimation based on airborne lidar point clouds using a specific target deep learning network. Before using the target deep learning network, it needs to be trained, and the training process is as follows:
[0083] 1. Data source acquisition: The data acquisition includes two parts: (1) Airborne lidar point cloud data of the study area, which can penetrate the canopy and provide detailed three-dimensional forest structure information; (2) Ground plot measured forest biomass data corresponding to the point cloud data, which serve as the true values for model training and validation.
[0084] 2. Data Preprocessing and Sample Construction: Taking ground plots as units, a subset of point clouds corresponding to each plot is extracted from the large-scale airborne lidar point cloud data. The extracted point clouds undergo denoising, ground point classification, and elevation normalization to eliminate the influence of terrain and obtain point cloud data reflecting the relative height of trees. Finally, a three-dimensional coordinate point set (N×3) consisting of N points is generated for each plot, which will serve as the direct input to the model.
[0085] 3. Dataset Partitioning: Point cloud data from all sample plots and their corresponding measured biomass data pairs were randomly divided into training, validation, and test sets according to a certain ratio (70%, 15%, 15%). The training set was used to train the network parameters, the validation set was used to adjust hyperparameters and monitor the model training process, and the test set was used to finally evaluate the model's performance and generalization ability.
[0086] 4. Train an end-to-end biomass estimation model based on deep learning.
[0087] The preprocessed sample point cloud data is input into the target deep learning network model for training and estimation. The core advantage of this network model lies in its end-to-end nature, enabling it to automatically learn high-dimensional, non-linear abstract features directly from the original point cloud without the need for manual feature design and selection. To more accurately depict complex forest scenes, the network innovatively integrates a global structural feature encoding module and a vertical structural hierarchical perception module on high-resolution feature maps, aiming to capture more complete structural information and provide global guidance for subsequent local feature extraction.
[0088] For any input point cloud data of a sample plot, the trained target deep learning network completes end-to-end processing and directly outputs a continuous value, which is the estimated value of forest biomass (AGB) of that sample plot.
[0089] After training the target deep learning network, its performance is evaluated using a reserved test set. The biomass predicted by the target deep learning network is compared with the actual biomass measured in the sample plots, and statistical indicators such as the coefficient of determination (R²), root mean square error (RMSE), and mean bias are calculated to quantify the model's estimation accuracy and reliability.
[0090] After training, the target-oriented deep learning network can be applied to large-scale airborne lidar data to achieve efficient and accurate mapping and dynamic monitoring of forest biomass at regional or national scales, providing technical support for forest resource surveys, carbon inventory assessment, and ecosystem management. This method requires no complex feature engineering; estimation can be performed simply by inputting point clouds, ensuring ease of application and providing strong technical support for forest resource surveys, carbon inventory assessment, and ecosystem management.
[0091] The above describes the training and application of the target deep learning network. The following section verifies the effectiveness of the target deep learning network.
[0092] Based on China's ecological zoning and forest types, a northern dataset was constructed to verify the model's inversion capability. The northern forest dataset covers Saihanba, Genhe, Mengjiagang, and Dagujia, with a total of over 600 sample plots and measured data.
[0093] This invention designs a target deep learning network that, compared to other methods, possesses stronger representational capabilities and enhances the ability to detect three-dimensional structural information such as the vertical and spatial distribution of forests, thereby improving the accuracy of regional biomass retrieval. To verify the effectiveness and advancement of the target deep learning network, a comprehensive comparative validation was conducted with various benchmark methods on a dataset covering typical forest regions in northern China. The selected methods include traditional machine learning methods (multiple linear regression, random forest) and state-of-the-art deep learning methods (ResNet14, PointNet, Kpconv, SENet14, SENet50, PointTransformer). Comparative experiments used the following three core metrics to evaluate the performance of all methods under different regions and conditions: Coefficient of Determinism (R²): measures the goodness of fit between the model's predicted values and the actual values. The closer R² is to 1, the better the model performance. Root Mean Square Error (RMSE): measures the deviation between the model's predicted values and the actual values. The smaller the RMSE value, the higher the model accuracy. Mean Bias: measures the systematic bias of the model's predictions, i.e., whether the predicted values are systematically overestimated or underestimated. The closer the bias is to 0, the more unbiased the model. The comparative results show that, regardless of R², RMSE, or Mean Bias, the model performs better. In terms of bias metrics, the target deep learning network proposed in this study performs best in most test regions and evaluation dimensions. Specifically, the target deep learning network has the highest R² value, the lowest RMSE value, and the average bias closest to zero, which fully demonstrates that its estimation results are superior to other similar models in terms of accuracy and reliability.
[0094] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 9 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements an airborne lidar point cloud forest biomass estimation method.
[0095] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0096] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0097] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0098] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0099] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0101] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0102] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0103] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for estimating forest biomass using airborne lidar point clouds, characterized in that, include: Acquire target airborne lidar data, wherein the target airborne lidar data is a three-dimensional point cloud obtained by the airborne lidar scanning the target area; The target airborne lidar data is input into the trained target deep learning network, and the predicted value of forest biomass in the target area is obtained through the target deep learning network. The target deep learning network includes an input layer, a backbone network, and an output layer stacked in sequence. The backbone network includes a global feature extraction layer and multiple local feature extraction layers stacked in sequence. The global feature extraction layer includes a global structural feature encoding module and a vertical structural hierarchical perception module stacked in sequence. The global structural feature encoding module is configured to extract key global features of the 3D point cloud by performing feature extraction on the output of the input layer. The vertical structure layered perception module is configured to: perform vertical layered perception on the key global features based on the relative height relationship of the three-dimensional point cloud to obtain perception features as input to the first local feature extraction layer; The feature extraction expression of the global structural feature encoding module is: in, This represents the coordinates of the i-th point cloud. This represents the features of the i-th point cloud provided by the input layer. This represents the m-th learnable parameter. Indicates by The key spatial locations obtained through learning express The characteristics of the place, The feature set representing all key spatial locations, , and Both represent linear transformation layers. F This represents the key global feature. Represents the spatial weighting function. This represents the semantic weight function; The vertical layering perception process of the vertical structure layering perception module includes: The key global features are divided into multiple hierarchical features belonging to different height intervals using height quantiles; For each point feature in each of the hierarchical features, height information is encoded; Each of the hierarchical features after high-level information encoding is subjected to feature perception to obtain multiple perception results, and the multiple perception results are fused to obtain a fusion result; The fusion result is subjected to feature perception to obtain the perceived features.
2. The airborne lidar point cloud forest biomass estimation method according to claim 1, characterized in that, The feature perception process for each of the hierarchical features is as follows: The hierarchical features are normalized and the first intermediate features are obtained through gated adaptive selection; The second intermediate feature is obtained by performing two-layer linear transformation and single-injection kernel mapping on the first intermediate feature; A third intermediate feature is obtained by performing a linear transformation on the first intermediate feature. The second and third intermediate features are dynamically weighted through a forgetting gate, and then a linear transformation is performed before the hierarchical features are merged through residual connections to obtain the fourth intermediate feature. The normalized fourth intermediate feature is perceived by a multilayer perceptron to obtain the fifth intermediate feature, and the fifth intermediate feature and the fourth intermediate feature are merged by residual connection to obtain the perception result.
3. The airborne lidar point cloud forest biomass estimation method according to claim 1, characterized in that, The input layer includes a first sparse 3D convolution and a saliency aggregation module stacked sequentially, and the output of the saliency aggregation module is the output of the input layer.
4. The airborne lidar point cloud forest biomass estimation method according to claim 1, characterized in that, There are four local feature extraction layers, each of which includes a downsampling layer. The global feature extraction layer and the multiple downsampling layers all use a local feature aggregation module to perform residual connection between the input and output.
5. The airborne lidar point cloud forest biomass estimation method according to claim 4, characterized in that, The local feature aggregation includes a second sparse three-dimensional convolution, a first normalization layer, an activation function, a third sparse three-dimensional convolution, and a second normalization layer stacked sequentially.
6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the airborne lidar point cloud forest biomass estimation method according to any one of claims 1-5.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the airborne lidar point cloud forest biomass estimation method as described in any one of claims 1-5.
8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the airborne lidar point cloud forest biomass estimation method as described in any one of claims 1-5.