Forest resource estimation method and device based on deep learning

Abstract

The invention discloses a forest resource estimation method and device based on deep learning. Relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining initial point cloud data of a target forest region, and carrying out the preprocessing of the initial point cloud data, and obtaining target point cloud data; performing feature extraction on the target point cloud data through a target backbone network of a target deep learning model to obtain target feature information; processing the target feature information through a plurality of full connection layers of the target deep learning model to obtain an output result corresponding to each full connection layer; and obtaining the forest resource quantity of the target forest region according to the output result corresponding to each full connection layer. According to the method and the device, the problem that the efficiency of estimating the forest resources is relatively low due to the fact that the forest resources in the forest area are estimated through a sample plot checking method in the related technology is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a forest resource estimation method and device based on deep learning. Background Art

[0002] Accurately and rapidly assessing the status of forest resources has become a key requirement in ecology and forestry management. While traditional plot inventory methods provide direct measurement data, their low efficiency and high cost when applied on a large scale limit their practicality in modern large-scale forest monitoring. Plot inventory methods require field measurements of a large number of plots, a time-consuming and labor-intensive process that makes it difficult to obtain comprehensive forest resource data in a short period of time. Furthermore, due to the complexity and variability of forests, plot inventory methods are unable to fully reflect the diversity of forest areas, especially in areas with complex tree species composition and widely varying growth conditions, limiting the accuracy of estimation results.

[0003] Currently, no effective solution has been proposed to the problem that the efficiency of estimating forest resources in forest areas is relatively low due to the use of sample plot inventory methods in related technologies. Summary of the Invention

[0004] The main purpose of this application is to provide a forest resource estimation method and device based on deep learning to solve the problem in related technologies of estimating forest resources in forest areas through sample plot inventory methods, resulting in relatively low efficiency in estimating forest resources.

[0005] To achieve the above objectives, according to one aspect of the present application, a deep learning-based forest resource estimation method is provided. The method comprises: obtaining initial point cloud data of a target forest area and preprocessing the initial point cloud data to obtain target point cloud data; extracting features from the target point cloud data using a target backbone network of a target deep learning model to obtain target feature information; processing the target feature information using multiple fully connected layers of the target deep learning model to obtain output results corresponding to each fully connected layer; and obtaining the forest resource volume of the target forest area based on the output results corresponding to each fully connected layer.

[0006] Furthermore, the target backbone network includes at least a convolutional layer, a channel and spatial attention layer, multiple residual block sequences and a self-attention layer. The target point cloud data is subjected to feature extraction by the target backbone network of the target deep learning model to obtain target feature information, including: performing preliminary feature extraction on the target point cloud data by the convolutional layer to obtain first feature information; performing channel layer and spatial layer feature extraction on the first feature information by the channel and spatial attention layer to obtain second feature information; performing multi-scale feature extraction on the second feature information by the multiple residual block sequences to obtain third feature information; and processing the third feature information by the self-attention layer to obtain the target feature information.

[0007] Furthermore, the first feature information is subjected to channel layer and spatial layer feature extraction through the channel and spatial attention layer to obtain the second feature information, including: extracting the channel features of the first feature information to obtain channel feature information; extracting the spatial features of the channel feature information to obtain spatial feature information; and obtaining the second feature information based on the spatial feature information.

[0008] Furthermore, extracting the channel features of the first feature information to obtain the channel feature information includes: performing a pooling operation on the first feature information to obtain a channel descriptor; processing the channel descriptor to obtain a weight vector for each channel; and calculating the first feature information and the weight vector for each channel to obtain the channel feature information.

[0009] Furthermore, the target deep learning model is trained by the following steps: obtaining multiple training sets, wherein each training set consists of a point cloud sample and a real resource amount corresponding to the point cloud sample, and the type of the real resource amount between each training set is different; performing single-task training on the first deep learning model through each training set to obtain the trained first deep learning model; constructing a second deep learning model through the backbone network and multiple fully connected layers of the trained first deep learning model; obtaining a target training set, wherein the target training set consists of a target point cloud sample and multiple real resource amounts of multiple types corresponding to the target point cloud samples; performing multi-task training on the second deep learning model through the target training set to obtain the target deep learning model.

[0010] Furthermore, the second deep learning model is subjected to multi-task training through the target training set to obtain the target deep learning model, including: processing the target point cloud samples through the second deep learning model to obtain multiple predicted resource quantities of multiple types; obtaining the weight value of the resource quantity corresponding to each type; calculating based on the multiple predicted resource quantities, the multiple real resource quantities of the multiple types and the weight values ​​to obtain a loss function; training the second deep learning model based on the loss function to obtain the target deep learning model.

[0011] Furthermore, the initial point cloud data is preprocessed to obtain target point cloud data, including: coarsely screening the initial point cloud data to obtain first point cloud data; finely screening the first point cloud data to obtain second point cloud data; and thinning and normalizing the second point cloud data to obtain the target point cloud data.

[0012] Furthermore, finely screening the first point cloud data to obtain the second point cloud data includes: screening the first point cloud data through a density-based clustering algorithm to obtain filtered first point cloud data; obtaining a point cloud distribution reference model based on the tree species types and tree species distribution in the target forest area; and screening the filtered first point cloud data through the point cloud distribution reference model to obtain the second point cloud data.

[0013] To achieve the above-mentioned purpose, according to another aspect of the present application, a forest resource estimation device based on deep learning is provided. The device comprises: a first acquisition unit for acquiring initial point cloud data of a target forest area and preprocessing the initial point cloud data to obtain target point cloud data; an extraction unit for extracting features from the target point cloud data using a target backbone network of a target deep learning model to obtain target feature information; a processing unit for processing the target feature information using multiple fully connected layers of the target deep learning model to obtain an output result corresponding to each fully connected layer; and a determination unit for obtaining the forest resource amount of the target forest area based on the output result corresponding to each fully connected layer.

[0014] Furthermore, the target backbone network includes at least a convolutional layer, a channel and spatial attention layer, multiple residual block sequences and a self-attention layer, and the extraction unit includes: a first extraction subunit, used to perform preliminary feature extraction on the target point cloud data through the convolutional layer to obtain first feature information; a second extraction subunit, used to perform channel layer and spatial layer feature extraction on the first feature information through the channel and spatial attention layer to obtain second feature information; a third extraction subunit, used to perform multi-scale feature extraction on the second feature information through the multiple residual block sequences to obtain third feature information; a first processing subunit, used to process the third feature information through the self-attention layer to obtain the target feature information.

[0015] Furthermore, the second extraction subunit includes: a first extraction module, used to extract the channel features of the first feature information to obtain channel feature information; a second extraction module, used to extract the spatial features of the channel feature information to obtain spatial feature information; and a first determination module, used to obtain the second feature information based on the spatial feature information.

[0016] Furthermore, the first extraction module includes: an operation submodule for performing a pooling operation on the first feature information to obtain a channel descriptor; a processing submodule for processing the channel descriptor to obtain a weight vector for each channel; and a calculation submodule for calculating the first feature information and the weight vector of each channel to obtain the channel feature information.

[0017] Furthermore, the target deep learning model is trained using the following devices: a second acquisition unit, used to acquire multiple training sets, wherein each training set consists of a point cloud sample and a real resource amount corresponding to the point cloud sample, and the type of the real resource amount between each training set is different; a first training unit, used to perform single-task training on the first deep learning model through each training set to obtain the trained first deep learning model; a construction unit, used to construct a second deep learning model through the backbone network and multiple fully connected layers of the trained first deep learning model; a third acquisition unit, used to acquire a target training set, wherein the target training set consists of a target point cloud sample and multiple real resource amounts of multiple types corresponding to the target point cloud samples; a second training unit, used to perform multi-task training on the second deep learning model through the target training set to obtain the target deep learning model.

[0018] Furthermore, the second training unit includes: a processing subunit, used to process the target point cloud sample through the second deep learning model to obtain multiple predicted resource quantities of multiple types; an acquisition subunit, used to obtain the weight value of the resource quantity corresponding to each type; a calculation subunit, used to calculate based on the multiple predicted resource quantities, the multiple real resource quantities of the multiple types and the weight values ​​to obtain a loss function; a training subunit, used to train the second deep learning model based on the loss function to obtain the target deep learning model.

[0019] Furthermore, the first acquisition unit includes: a first screening subunit, used to perform coarse screening on the initial point cloud data to obtain first point cloud data; a second screening subunit, used to perform fine screening on the first point cloud data to obtain second point cloud data; and a second processing subunit, used to perform thinning and normalization on the second point cloud data to obtain the target point cloud data.

[0020] Furthermore, the second screening subunit includes: a first screening module, used to screen the first point cloud data through a density-based clustering algorithm to obtain filtered first point cloud data; a second determination module, used to obtain a point cloud distribution reference model based on the tree species types and tree species distribution in the target forest area; a third screening module, used to screen the filtered first point cloud data through the point cloud distribution reference model to obtain the second point cloud data.

[0021] According to another aspect of an embodiment of the present invention, an electronic device is provided, including: a memory storing an executable program; and a processor for running the program, wherein when the program is running, any one of the above-mentioned deep learning-based forest resource estimation methods is executed.

[0022] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, which stores a program, wherein when the program is running, the device where the storage medium is located is controlled to execute any of the above-mentioned forest resource estimation methods based on deep learning.

[0023] In an embodiment of the present application, the following steps are adopted: obtaining initial point cloud data of the target forest area, and preprocessing the initial point cloud data to obtain target point cloud data; extracting features of the target point cloud data through the target backbone network of the target deep learning model to obtain target feature information; processing the target feature information through multiple fully connected layers of the target deep learning model to obtain the output results corresponding to each fully connected layer; according to the output results corresponding to each fully connected layer, the amount of forest resources in the target forest area is obtained, thereby solving the technical problem in the related art of estimating forest resources in forest areas through sample plot inventory method, resulting in relatively low efficiency in estimating forest resources.

[0024] In this solution, after obtaining the initial point cloud data of the target forest area, the initial point cloud data is preprocessed to ensure the quality of the input data and reduce invalid information. The target backbone network of the target deep learning model is used to extract features from the preprocessed target point cloud data, which can automatically learn and capture complex spatial patterns in the point cloud data that are closely related to forest resource volume, such as crown structure, trunk size, etc., thereby improving processing efficiency and accuracy. The extracted target feature information is processed through multiple fully connected layers to obtain multi-task output results of accumulation volume, biomass and carbon storage, which can quickly and accurately estimate the forest resource volume of the target forest area. This solution overcomes the limitations of the sample plot inventory method in large-scale data collection, achieves comprehensive coverage of the entire forest area, and thus achieves the technical effect of improving the efficiency of estimating forest resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0026] Figure 1 A hardware structure block diagram of a computer terminal for implementing a forest resource estimation method based on deep learning is shown;

[0027] Figure 2 This is a flow chart of a forest resource estimation method based on deep learning provided in an embodiment of the present application;

[0028] Figure 3 This is a schematic diagram of the deep learning model provided by the embodiment of the present application. Figure 1 ;

[0029] Figure 4 This is a schematic diagram of the deep learning model provided by the embodiment of the present application. Figure 2 ;

[0030] Figure 5 This is a schematic diagram of the deep learning model provided by the embodiment of the present application. Figure 3 ;

[0031] Figure 6 is a schematic diagram of a forest resource estimation method based on deep learning provided in an embodiment of the present application;

[0032] Figure 7 is a schematic diagram of a forest resource estimation device based on deep learning provided in an embodiment of the present application;

[0033] Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0035] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0036] It should be noted that the collected information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation portals for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions to provide users with corresponding operation portals for users to choose to agree or refuse the automated decision-making results; if the user chooses to refuse, the expert decision-making process will be entered.

[0037] Example 1

[0038] According to an embodiment of the present application, a method embodiment for estimating forest resources based on deep learning is also provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] The method embodiment provided in the first embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1The hardware structure block diagram of a computer terminal (or mobile device) for implementing a forest resource estimation method based on deep learning is shown. Figure 1 As shown, the computer terminal 10 (or mobile device) may include one or more (illustrated as 102a, 102b, ..., 102n in the figure) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It will be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0040] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry." The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuitry may be a single independent processing module, or may be integrated into any of the other components of the computer terminal 10 (or mobile device) in whole or in part. As described in the embodiments of the present application, the data processing circuitry serves as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0041] Memory 104 can be used to store software programs and modules for application software, such as the program instructions / data storage device corresponding to the deep learning-based forest resource estimation method in the embodiments of the present application. Processor 102 executes the software programs and modules stored in memory 104 to perform various functional applications and data processing, thereby implementing the aforementioned deep learning-based forest resource estimation method. Memory 104 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, memory 104 may further include memory remotely located relative to processor 102, which can be connected to computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0042] The transmission device 106 is configured to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the communications provider of the computer terminal 10. In one embodiment, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, the transmission device 106 may be a radio frequency (RF) module, which is configured to communicate with the Internet wirelessly.

[0043] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0044] Under the above operating environment, this application provides Figure 2 The deep learning-based forest resource estimation method shown. Figure 2 This is a flow chart of a forest resource estimation method based on deep learning according to Example 1 of the present application. The forest resource estimation method based on deep learning includes:

[0045] Step S201 : acquiring initial point cloud data of a target forest area, and preprocessing the initial point cloud data to obtain target point cloud data.

[0046] Alternatively, initial point cloud data of the target forest area can be collected using LiDAR (Light Detection And Ranging) technology. The initial point cloud data obtained from the LiDAR device may contain a large number of noise points, including instrument errors, atmospheric scattering, non-vegetation reflections, etc., and preliminary data cleaning and quality control are required to remove obviously abnormal or invalid data points. For example, determine which data in the collected point cloud data are abnormal: set a deviation threshold, calculate whether the Z value of the collected point cloud data deviates from the height range of other points in the neighborhood (i.e., the deviation threshold), such as the sudden appearance of extremely high or extremely low isolated points on the ground; determine whether there are points in the point cloud data that have no obvious relationship with other points in three-dimensional space; determine whether there are multiple points with the same position coordinates and coordinate points that are outside the reasonable range (xy is the coordinate, z represents the spatial position, and only Z is compared for the same coordinates).

[0047] The above steps remove equipment errors, atmospheric scattering and other noises, thereby improving the reliability of the data.

[0048] Step S202: extract features from the target point cloud data through the target backbone network of the target deep learning model to obtain target feature information.

[0049] Optionally, after obtaining the target point cloud data, the target point cloud data is input into a target deep learning model, and features are extracted from the target point cloud data using the target backbone network within the target deep learning model. For example, the target backbone network is composed of three-dimensional sparse convolution kernels, which can effectively capture the distribution of point cloud data in three-dimensional space, including important features such as tree height and crown width. By adopting three-dimensional sparse convolution, ineffective calculations for empty points are greatly reduced. Compared with dense convolution, computational efficiency is significantly improved, especially when processing large-scale point cloud data, which can significantly reduce the demand for computing resources.

[0050] Step S203: Process the target feature information through multiple fully connected layers of the target deep learning model to obtain the output results corresponding to each fully connected layer.

[0051] Optionally, the target deep learning model includes not only a target backbone network for feature extraction, but also multiple fully connected layers to further process and analyze the extracted target feature information to produce prediction outputs directly related to forest resource quantity (such as standing volume, biomass, and carbon storage). Each fully connected layer independently predicts a specific factor of forest resources, such as standing volume, biomass, and carbon storage.

[0052] By designing a model structure containing multiple fully connected layers, it is possible to process different aspects of forest resources in parallel, such as the prediction of volume, biomass and carbon storage, which improves the versatility of the model and enhances its overall performance and generalization ability.

[0053] Step S204: Obtain the forest resource volume of the target forest area based on the output results corresponding to each fully connected layer.

[0054] Optionally, after the target feature information is extracted through the backbone network and passed to the fully connected layer, each layer converts this information into a numerical value directly related to the prediction target. Specifically, each fully connected layer performs a linear combination (i.e., weighted sum) of the feature vectors it accepts, and performs a nonlinear transformation through an activation function (such as ReLU, Sigmoid, or tanh), and finally outputs the predicted value of a specific forest resource quantity. For example, the first fully connected layer may focus on predicting the amount of stock, the second fully connected layer for biomass, and the third fully connected layer for carbon storage. By designing independent fully connected layers for each type of forest resource quantity (stock, biomass, carbon storage), the model can make refined predictions for each task separately, ensuring the professionalism and accuracy of resource quantity estimation.

[0055] Optionally, in the deep learning-based forest resource estimation method provided in an embodiment of the present application, the target backbone network includes at least a convolutional layer, a channel and spatial attention layer, multiple residual block sequences and a self-attention layer. The target point cloud data is subjected to feature extraction by the target backbone network of the target deep learning model to obtain target feature information, including: performing preliminary feature extraction on the target point cloud data through the convolutional layer to obtain first feature information; performing channel layer and spatial layer feature extraction on the first feature information through the channel and spatial attention layer to obtain second feature information; performing multi-scale feature extraction on the second feature information through multiple residual block sequences to obtain third feature information; and processing the third feature information through the self-attention layer to obtain target feature information.

[0056] In an optional embodiment, if Figure 3 As shown in the figure, the target backbone network consists of at least convolutional layers, channel and spatial attention layers, multiple residual block sequences, and self-attention layers. The convolutional layers capture basic geometric and spatial features from the target point cloud data, such as tree outlines and height distribution. The convolution operation transforms the target point cloud data into a higher-level abstract feature representation, namely the first feature information.

[0057] Based on the first feature information, the channel and spatial attention layers further highlight features closely related to forest resource estimation, while suppressing irrelevant or secondary features. At the channel level, the attention mechanism adjusts the weights of different feature channels, directing the model's attention to features directly related to parameters such as biomass and carbon storage. At the spatial level, the attention mechanism can locate features such as vegetation structure, crown width, and shape in the point cloud data. In other words, the channel and spatial attention layers extract channel-level and spatial-level features from the first feature information to obtain the second feature information.

[0058] The residual block sequence can effectively solve the gradient vanishing problem in deep network training, enabling it to learn more complex representations. In the present application, multiple residual block sequences are responsible for extracting multi-scale features from the second feature information to obtain third feature information. Each residual block can be facilitated by a convolutional layer, batch normalization, and an activation function. The output of each residual block serves as the input of the next residual block. Each residual block in the sequence learns and superimposes new feature information while retaining the input information, which can ensure that the network learns rich features and avoid the gradient vanishing problem caused by increased depth. In an embodiment of the present application, the use of residual block sequences enables the deep learning model to effectively extract multi-scale features from the target point cloud data.

[0059] After obtaining the third feature information, the self-attention layer dynamically allocates weights according to the relationship between features, thereby integrating information globally to obtain the above-mentioned target feature information.

[0060] Through the combined use of convolutional layers, channel and spatial attention layers, residual block sequences, and self-attention layers, irrelevant information is effectively filtered out, key features are highlighted, the model's sensitivity to noise is reduced, and the reliability and stability of the prediction results are improved.

[0061] Optionally, in the forest resource estimation method based on deep learning provided in an embodiment of the present application, channel layer and spatial layer feature extraction are performed on the first feature information through the channel and spatial attention layers to obtain the second feature information, including: extracting the channel features of the first feature information to obtain channel feature information; extracting the spatial features of the channel feature information to obtain spatial feature information; and obtaining the second feature information based on the spatial feature information.

[0062] In an optional embodiment, the channel and spatial attention layer extracts the channel features in the first feature information and identifies which feature channels are most critical for forest resource estimation (stock volume, biomass, carbon storage, etc.), that is, the above-mentioned channel feature information is obtained by adjusting the weights of each channel. Based on the channel feature information, the channel and spatial attention layer further extracts features that are closely related to the spatial distribution of forest structure. The extraction of spatial feature information helps the model understand the interactions between individual trees and their distribution patterns in the forest area. Finally, the second feature information is obtained based on the spatial feature information. The channel and spatial attention layer can intelligently screen and strengthen features that are closely related to the target resource estimation, reduce the interference of irrelevant features, and significantly improve the efficiency and accuracy of model learning and resource estimation.

[0063] The channel and spatial attention layers adjust the weights of feature channels and spatial positions to focus on features directly related to forest resource estimation (such as biomass and carbon storage), while suppressing irrelevant or secondary features. This mechanism significantly enhances the model's sensitivity to key information and improves the accuracy and robustness of resource prediction.

[0064] Optionally, in the deep learning-based forest resource estimation method provided in an embodiment of the present application, the channel features of the first feature information are extracted to obtain the channel feature information, including: performing a pooling operation on the first feature information to obtain a channel descriptor; processing the channel descriptor to obtain a weight vector for each channel; and calculating the first feature information and the weight vector of each channel to obtain the channel feature information.

[0065] In an optional embodiment, the channel and spatial attention layers perform a pooling operation on the first feature information. The purpose of the pooling operation is to compress the size of the feature map and extract a descriptor from each feature channel. This descriptor is used to represent the average response or maximum response of the channel on the entire feature map.

[0066] The channel descriptors are then processed by fully connected layers within the channel and spatial attention layers to generate a weight vector for each channel. This weight vector reflects the importance of each channel feature in forest inventory estimation, specifically, which features are most likely to be closely associated with prediction targets such as forest volume, biomass, and carbon storage. The weight vectors are calculated based on parameters learned during model training. These parameters guide the model in intelligently ranking different channel features, thereby enabling feature prioritization.

[0067] Finally, the first feature information is fused with the calculated weight vector for each channel, for example, by applying the weights to the corresponding feature channel through a multiplication operation (i.e., element-by-element multiplication). This process ensures that the model strengthens key features while weakening the contribution of less important features. The resulting channel feature information is more focused on information that is highly relevant to forest inventory estimation.

[0068] In an optional embodiment, the channel and spatial attention layers can be composed of pooling layers, fully connected layers, and activation functions. Pooling layer: Use global average pooling or global maximum pooling operations to compress the feature map into a fixed-size vector. This vector contains the information of the entire feature map and is specific to each feature channel. Fully connected layer: The vector obtained by the pooling operation is sent to one or more fully connected layers for further processing and refining information, with the aim of learning the weight of each channel. The fully connected layer determines the importance of each channel for the final prediction through weight learning. Activation function: After the fully connected layer, an activation function such as Sigmoid is usually used to map the obtained weight vector to between 0 and 1, thereby generating normalized weights for subsequent feature channel weighting operations. Weighting operation: Finally, the calculated channel weight is element-wise multiplied with the original feature map, that is, the weight vector is multiplied with the corresponding position of the feature map of each feature channel to achieve weight adjustment of the feature channel.

[0069] By calculating channel descriptors and weight vectors, the model can intelligently identify and rank the importance of features, which helps focus the learning process on the most critical information and improves the accuracy of predicting forest resource volume.

[0070] In an optional embodiment, the following steps can be used to obtain a spatial feature vector: a spatial weight map can be generated by a convolution operation. The spatial weight map has the same size and shape as the original feature map, but the value of each position represents the importance of that position. Then, the channel feature map is weighted using the obtained spatial weight map, and the weight of each position is applied to the corresponding feature through element-by-element multiplication to obtain a spatial feature vector. Channel attention can adaptively adjust the weights of different channels, while spatial attention can adaptively adjust the weights of different spatial positions. This dual adaptive mechanism enables the model to learn and use features more accurately, thereby improving the accuracy and robustness of predictions.

[0071] Optionally, in the forest resource estimation method based on deep learning provided in an embodiment of the present application, the target deep learning model is trained using the following steps: obtaining multiple training sets, wherein each training set consists of a point cloud sample and a real resource quantity corresponding to the point cloud sample, and the type of real resource quantity between each training set is different; performing single-task training on the first deep learning model through each training set to obtain the trained first deep learning model; constructing a second deep learning model through the backbone network and multiple fully connected layers of the trained first deep learning model; obtaining a target training set, wherein the target training set consists of a target point cloud sample and multiple real resource quantities of multiple types corresponding to the target point cloud samples; performing multi-task training on the second deep learning model through the target training set to obtain a target deep learning model.

[0072] In an optional embodiment, multiple training sets are obtained, each of which contains point cloud samples and the corresponding true value of forest resource quantity. It should be noted that the types of resource quantity are different between the training sets. Each training set corresponds to a specific resource quantity type. For example, the resource quantity type is divided into accumulation, carbon storage, and biomass. In an optional embodiment, the schematic diagram of the first deep learning model is as follows Figure 4 As shown, it includes: backbone network and fully connected layer.

[0073] The first deep learning model is trained on a single task using multiple training sets. For example, separate training is performed on stock volume, carbon storage, and biomass. In an optional embodiment, the global point cloud data is divided into regular blocks of 25*25 square meters, and divided into training, test, and validation sets in an 8:1:1 ratio. Each task is trained separately. Single-task training can be performed for three times the training time of a multi-task model, trading time for independent feature adaptability.

[0074] After obtaining the trained first deep learning model, the backbone network of the first deep learning model is retained, that is, the convolution layer, pooling layer, residual block, etc. responsible for extracting common features. Then, based on the backbone network and the newly added multiple fully connected layers, a second deep learning model is constructed. It should be noted that each fully connected layer corresponds to a prediction head for a resource type. In an optional embodiment, the schematic diagram of the second deep learning model is as follows Figure 5 As shown, it includes: convolutional layers, channel and spatial attention layers, residual block sequences, self-attention layers, and multiple fully connected layers.

[0075] Then obtain the target training set, which contains point cloud samples and their corresponding true values ​​of various resource quantities. Unlike the single-task training set, the target training set covers all resource quantity types, allowing the model to learn on multiple tasks simultaneously. Use the target training set to perform multi-task training on the second deep learning model. The backbone network of the model converts the point cloud samples into feature representations, and then these features are passed to different fully connected layers for resource quantity prediction to perform multi-task training on the second deep learning model and obtain the final target deep learning model. For example, three types of resource quantities, y1: accumulation, y2: biomass, and y3: carbon storage, are introduced, and accurate prediction of various task targets is achieved through multi-task training.

[0076] By combining single-task pre-training with multi-task joint training, the model is optimized for each resource quantity prediction task. At the same time, through feature sharing and cross-task learning, the overall prediction accuracy and robustness are improved.

[0077] Optionally, in the deep learning-based forest resource estimation method provided in an embodiment of the present application, the second deep learning model is multi-task trained through the target training set to obtain the target deep learning model, including: processing the target point cloud samples through the second deep learning model to obtain multiple predicted resource quantities of multiple types; obtaining the weight value of the resource quantity corresponding to each type; calculating based on the multiple predicted resource quantities, multiple real resource quantities of multiple types and the weight values ​​to obtain the loss function; training the second deep learning model based on the loss function to obtain the target deep learning model.

[0078] In an optional embodiment, the backbone network of the second deep learning model is responsible for feature extraction of point cloud samples, and the generated feature information is then input into multiple fully connected layers corresponding to different resource types. Each fully connected layer corresponds to a prediction head, which is used to predict the amount of forest resources of a specific type, that is, the multiple predicted resource amounts of the above-mentioned multiple types, such as predicted stock, carbon reserves, and biomass.

[0079] Then, a weight value is assigned to each resource type (such as accumulation, carbon storage, biomass). The setting of weight values ​​can be based on factors such as the relative importance of the task, the availability of data, or the difficulty of prediction. For example, in order to reduce training time, the Loss weight is set to Loss_y1: 0.5, Loss_y2: 0.5, including but not limited to the total value of the set weight being 1, and the number of Loss_y is determined according to the number of tasks introduced. For example, this task involves 3, which are 0.4, 0.3, and 0.3 respectively. During the training process, these weight values ​​are used to adjust the loss functions of different prediction heads to achieve balance and optimization between multiple tasks.

[0080] Based on the model's predictions of multiple types of resource quantities, the corresponding actual resource quantities, and the weight values ​​of each resource quantity type, a comprehensive multi-task loss function (i.e., the loss function mentioned above) is calculated. The loss function can be a weighted average of the loss functions of each prediction head, where the weight value reflects the importance of each resource type prediction. For example, the mean square error (MSE) loss of stock volume prediction, the absolute percentage error (MAPE) loss of carbon stock prediction, and the mean absolute error (MAE) loss of biomass prediction are calculated. Each loss is weighted by its corresponding weight to form a total loss function. For example, multiple loss functions such as Smooth L1 Loss, MSE Loss, and MAPE Loss are introduced, and the model training is optimized through a loss fusion strategy. According to certain weights, the results of multiple loss functions are weighted and summed. For the final loss descent curve, the descent curvature of each loss is analyzed, and the results of multiple loss fusion are adjusted by adjusting the learning rate and loss weights. In this process, the importance of different error measurement methods to specific application fields must be considered. At the same time, the data itself must be taken into account. By experimenting with different weight configurations and the curvature of each loss decrease, the weights are adjusted to balance the impact of different error measurement methods, improve the robustness and generalization ability of the model, and enhance the stability of the model.

[0081] Finally, the calculated multi-task loss function is used to backpropagate the second deep learning model and update its parameters. Through iterative optimization, the model learns how to simultaneously minimize the loss for all resource quantity predictions, ultimately achieving a target deep learning model that performs well across multiple resource quantity types.

[0082] By combining comprehensive feature learning with task-specific head structures, the model can more accurately predict various forest resource quantities such as stock volume, carbon storage, and biomass.

[0083] In an optional embodiment, a multi-dimensional verification system can be used to comprehensively evaluate the performance of the model. First, the accuracy is verified by selecting independent verification sets with different forest stands and different regions and uniform distribution: the training set and the independent verification set are completely separated to test the model and quantify the difference between the predicted value and the true value: then an adaptability analysis is performed for different forest types (coniferous forest, broad-leaved forest, etc.): the performance of the model is evaluated separately for different forest types to ensure that the model can accurately capture the unique characteristics of each type of forest stand; in addition, a spatial distribution consistency test is performed: the rationality of the spatial distribution pattern of the model prediction results is analyzed, that is, whether the samples are evenly distributed in the entire domain and whether the distribution of each plot is similar. Overall, the point cloud data of different forest stands and different spatial regions are input into the model to calculate the determination coefficient (R 2 ), root mean square error (RMSE) and mean absolute percentage error (MAPE) and other core indicators, by focusing on analyzing the adaptability of the model to different forest types (coniferous forest, broad-leaved forest, etc.): First, based on the quantitative analysis of core indicators, namely R 2 , RMSE, MAPE, evaluate the overall prediction performance of the model, R 2 : Ideally, close to 1; greater than 0.7 generally performs well; RMSE: The smaller the better; MAPE: Within 20%. Separate reports are conducted for different forest stand types to examine the differences in model adaptability. Second, residual analysis is performed: residual plots are drawn to determine whether there are systematic biases.

[0084] It should be noted that when comprehensively evaluating model performance through multi-dimensional verification, attention should be paid to the balance analysis of various values. For example, in actual inversion projects, customers are more concerned about R 2 , so R 2 As a priority standard, the higher the better, and other values ​​are used as a reference.

[0085] Optionally, in the deep learning-based forest resource estimation method provided in an embodiment of the present application, the initial point cloud data is preprocessed to obtain target point cloud data, including: coarse screening of the initial point cloud data to obtain first point cloud data; fine screening of the first point cloud data to obtain second point cloud data; and thinning and normalizing the second point cloud data to obtain target point cloud data.

[0086] In an optional embodiment, the initial point cloud data is roughly screened to preliminarily remove points that clearly do not belong to the forest area or invalid data points, such as noise points caused by equipment failure, abnormal signal reflection, etc. For example, by calculating the deviation between the Z value (height) of each point in the point cloud data and the Z value of the points in the neighborhood, points with deviation values ​​exceeding a set threshold are removed. For another example, data points that have no reasonable connection with other points in three-dimensional space and may be isolated points or erroneous long-distance measurement points are identified. Points outside the reasonable coordinate range are also removed to ensure that only point cloud data belonging to the forest-covered area is retained.

[0087] The first point cloud data is then carefully screened to further improve its quality and purity and reduce noise. For example, local point density calculations, such as setting the number of points within a neighborhood radius, can be used to remove noise points deemed isolated or low-density. Because point cloud data can be very dense, the second point cloud data is thinned to reduce the data volume and computational requirements for model training. Feature preservation strategies are also employed to ensure that key vegetation structural features are preserved.

[0088] Finally, the coordinates of the point cloud data are normalized into a unified reference system, and the elevation information is standardized to eliminate the elevation deviation caused by different acquisition batches or equipment, ensure the consistency of the model input, and improve the accuracy of the prediction.

[0089] The target point cloud data obtained through the preprocessing step not only removes noise points and optimizes the point cloud quality, but also improves the accuracy of resource estimation.

[0090] Optionally, in the deep learning-based forest resource estimation method provided in an embodiment of the present application, the first point cloud data is finely screened to obtain the second point cloud data, including: filtering the first point cloud data through a density-based clustering algorithm to obtain filtered first point cloud data; obtaining a point cloud distribution reference model based on the tree species types and tree species distribution in the target forest area; and filtering the filtered first point cloud data through the point cloud distribution reference model to obtain the second point cloud data.

[0091] In an optional embodiment, fine-screening the first point cloud data includes the following steps: using a density-based clustering algorithm to set a neighborhood radius and a minimum number of neighborhood points, removing noise points that are neither core points nor within the neighborhood radius of any core point, thereby obtaining filtered first point cloud data. Alternatively, the average distance of each point to its 32 / 64 neighbors is calculated, and the average distance and standard deviation of all points are calculated. If the average distance of a point exceeds a set threshold, the point is removed. This method retains points that meet the distribution characteristics and removes interference from outliers.

[0092] It is also possible to remove obviously deviated noise points based on vegetation characteristics and combined with elevation difference analysis. That is, the point cloud elevation of the vegetation area should show a certain continuity and gradual trend. By constructing a point cloud distribution reference model for the local area, the deviation between the actual Z value of each point and the Z value predicted by the reference model is calculated, and an elevation deviation threshold is set. Points exceeding this threshold are considered abnormal points. The point cloud distribution reference model can reflect the height distribution pattern of typical vegetation in forest areas. The point cloud distribution reference model can be implemented by analyzing the elevation information of a large amount of valid point cloud data using statistical methods (such as mean, median) or machine learning methods.

[0093] Through the above-mentioned fine screening steps, not only the outliers and noise found based on density analysis are removed, but also the data points that do not conform to the height distribution pattern of forest vegetation are further eliminated through elevation analysis, thereby greatly improving the quality of point cloud data.

[0094] In an alternative embodiment, the Figure 6 The schematic diagram shown implements the model training process, which specifically includes:

[0095] Step S601: Data collection and preprocessing: perform preliminary cleaning of the original LiDAR point cloud data (LAS format) and sample data to remove outliers and invalid data to ensure the relative reliability of the input data. Specifically, it includes:

[0096] Determine which data in the collected point cloud data is abnormal: set a deviation threshold and calculate whether the Z value of the collected point cloud data deviates from the height range of other points in the neighborhood (i.e., the deviation threshold), such as the sudden appearance of extremely high or extremely low isolated points on the ground; determine whether there are points in the point cloud data that have no obvious relationship with other points in three-dimensional space; determine whether there are multiple points with the same coordinates and coordinate points that exceed the reasonable range (xy is the coordinate, z represents the spatial position, and only Z is compared for the same coordinates). Clean the above abnormal points: including using clustering algorithms to eliminate outliers; deduplication of duplicate points; setting reasonable boundaries to filter out points that clearly exceed the reasonable range.

[0097] Step S602: Optimize point cloud data. Optimize the LAS data processed in S601. Through a series of operations, remove noise interference, improve data quality, reduce the pressure of training on computing power, and lay the foundation for subsequent feature extraction.

[0098] The specific operations are as follows:

[0099] a. Noise reduction:

[0100] Remove noise points caused by equipment errors, environmental interference, and other characteristics to improve the purity of point cloud data. This includes using DBSCAN to set the neighborhood radius and minimum number of neighborhood points to remove noise points that are neither core points nor within the neighborhood radius of any core points. Alternatively, calculate the average distance of each point to its 32 / 64 neighbors, then calculate the average distance and standard deviation of all points. If the average distance of a point exceeds a set threshold, it is removed. This method retains points that meet the distribution characteristics and removes interference from outliers.

[0101] Based on vegetation characteristics and combined with elevation difference analysis, we remove significantly deviated noise points. Specifically, the point cloud elevation in vegetation areas should exhibit a certain degree of continuity and gradual change. By constructing a reference model for the point cloud distribution in the local area, we calculate the deviation between the actual Z value of each point and the Z value predicted by the reference model. We then set an elevation deviation threshold, and points exceeding this threshold are considered outliers.

[0102] b. Thinning:

[0103] The density of point clouds can lead to a huge consumption of training resources. A thinning strategy can be adopted to reduce the storage consumption caused by samples and the memory / GPU memory consumption during model training while retaining the sample characteristics, which can also reduce the training time.

[0104] c. Conversion:

[0105] During point cloud data preprocessing, the elevation information (i.e., Z-direction values) of the point cloud must be standardized based on a unified reference coordinate system. This eliminates elevation deviations between measurement intervals caused by differences in acquisition equipment or reference systems, normalizing the Z values ​​of the entire point cloud dataset to the same vertical reference system for ease of subsequent model training and calculation.

[0106] Step S603: Extracting vegetation point clouds: Using a smooth surface growth algorithm, the point cloud data is separated by preparing points, retaining valid non-smooth vegetation point clouds and removing artificial point clouds.

[0107] Step S604: Model selection and construction, building a single-task model through the backbone network and the fully connected layer.

[0108] Step S605: training phase.

[0109] a. Single-task model:

[0110] Train a single-task model and train (accumulation, carbon storage, biomass) separately. Divide the global point cloud data into regular blocks of 25*25 square meters, and divide them into training set, test set and validation set in a ratio of 8:1:1. Train each task separately. You can set the training time of the multi-task model to 3 times for single-task training, and exchange time for the independent adaptability of features.

[0111] b. Multi-task model:

[0112] During training, we simultaneously introduce y1 (stock volume), y2 (biomass), and y3 (carbon storage). Through a multi-task model, we share feature parameters. This means we construct a shared feature extraction backbone within the model to extract common features applicable to multiple tasks. Subsequently, we set task-specific header structures (i.e., fully connected layers) for each task to achieve accurate prediction of each task's objectives. To reduce training time, we set the loss weights to 0.5 for Loss_y1 and 0.5 for Loss_y2.

[0113] c. Multi-Loss Fusion:

[0114] We introduce multiple loss functions, including Smooth L1 Loss, MSE Loss, and MAPE Loss, and optimize model training through a loss fusion strategy. We perform a weighted summation of the results of multiple loss functions based on specific weights. To determine the final loss descent curve, we analyze the curvature of each loss descent and fine-tune the fusion results by adjusting the learning rate and loss weights. This process considers the importance of different error metrics for specific application areas, while also taking into account the data itself. By experimenting with different weight configurations and the curvature of each loss descent, we adjust the weights to balance the impact of different error metrics, improve the model's robustness, generalization, and stability.

[0115] Step S606: In the model evaluation phase, a multi-dimensional verification system is used to comprehensively evaluate the model performance. First, accuracy verification is performed by selecting independent validation sets with uniform distribution across different forest stands and regions. The model is tested using a completely separate training set and independent validation set to quantify the difference between the predicted value and the true value. Then, adaptability analysis is performed for different forest stand types (coniferous forest, broad-leaved forest, etc.). The performance of the model is evaluated separately for different forest types to ensure that the model can accurately capture the unique characteristics of each forest stand. In addition, a spatial distribution consistency test is performed to analyze the rationality of the spatial distribution pattern of the model prediction results, that is, whether the samples are evenly distributed across the entire region and whether the distribution of each plot is similar.

[0116] In general, the determination coefficients (R 2 ), root mean square error (RMSE) and mean absolute percentage error (MAPE) and other core indicators, by focusing on analyzing the adaptability of the model to different forest types (coniferous forest, broad-leaved forest, etc.): First, based on the quantitative analysis of core indicators, namely R 2, RMSE, MAPE, evaluate the overall prediction performance of the model, R 2 : Ideally, close to 1; greater than 0.7 generally performs well; RMSE: The smaller the better; MAPE: Within 20%. Separate reports are conducted for different forest stand types to examine the differences in model adaptability. Second, residual analysis is performed: residual plots are drawn to determine whether there are systematic biases.

[0117] The deep learning-based forest resource estimation method provided in the embodiment of the present application obtains initial point cloud data of the target forest area, and preprocesses the initial point cloud data to obtain target point cloud data; performs feature extraction on the target point cloud data through the target backbone network of the target deep learning model to obtain target feature information; processes the target feature information through multiple fully connected layers of the target deep learning model to obtain the output result corresponding to each fully connected layer; and obtains the forest resource amount of the target forest area based on the output result corresponding to each fully connected layer, which solves the technical problem in the related art of estimating forest resources in forest areas through sample plot inventory method, resulting in relatively low efficiency in estimating forest resources.

[0118] In this solution, after obtaining the initial point cloud data of the target forest area, the initial point cloud data is preprocessed to ensure the quality of the input data and reduce invalid information. The target backbone network of the target deep learning model is used to extract features from the preprocessed target point cloud data, which can automatically learn and capture complex spatial patterns in the point cloud data that are closely related to forest resource volume, such as crown structure, trunk size, etc., thereby improving processing efficiency and accuracy. The extracted target feature information is processed through multiple fully connected layers to obtain multi-task output results of accumulation volume, biomass and carbon storage, which can quickly and accurately estimate the forest resource volume of the target forest area. This solution overcomes the limitations of the sample plot inventory method in large-scale data collection, achieves comprehensive coverage of the entire forest area, and thus achieves the technical effect of improving the efficiency of estimating forest resources.

[0119] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0120] Example 2

[0121] The present application also provides a deep learning-based forest resource estimation device. It should be noted that the deep learning-based forest resource estimation device of the present application can be used to execute the deep learning-based forest resource estimation method provided in the present application. The following describes the deep learning-based forest resource estimation device provided in the present application.

[0122] According to an embodiment of the present application, a device for implementing the above-mentioned forest resource estimation method based on deep learning is also provided, such as Figure 7 As shown, the device includes: a first acquiring unit 701 , an extracting unit 702 , a processing unit 703 and a determining unit 704 .

[0123] The first acquisition unit 701 is used to acquire initial point cloud data of the target forest area and pre-process the initial point cloud data to obtain target point cloud data;

[0124] An extraction unit 702 is configured to extract features from the target point cloud data using a target backbone network of a target deep learning model to obtain target feature information;

[0125] A processing unit 703 is configured to process the target feature information through multiple fully connected layers of the target deep learning model to obtain an output result corresponding to each fully connected layer;

[0126] The determination unit 704 is configured to obtain the forest resource amount of the target forest area according to the output results corresponding to each fully connected layer.

[0127] The deep learning-based forest resource estimation device provided in the embodiment of the present application obtains the initial point cloud data of the target forest area through the first acquisition unit 701, and preprocesses the initial point cloud data to obtain target point cloud data; the extraction unit 702 extracts features from the target point cloud data through the target backbone network of the target deep learning model to obtain target feature information; the processing unit 703 processes the target feature information through multiple fully connected layers of the target deep learning model to obtain the output result corresponding to each fully connected layer; the determination unit 704 obtains the forest resource amount of the target forest area based on the output result corresponding to each fully connected layer, which solves the technical problem in the related art that the forest resources of the forest area are estimated through the sample plot inventory method, resulting in relatively low efficiency in estimating forest resources.

[0128] In this solution, after obtaining the initial point cloud data of the target forest area, the initial point cloud data is preprocessed to ensure the quality of the input data and reduce invalid information. The target backbone network of the target deep learning model is used to extract features from the preprocessed target point cloud data, which can automatically learn and capture complex spatial patterns in the point cloud data that are closely related to forest resource volume, such as crown structure, trunk size, etc., thereby improving processing efficiency and accuracy. The extracted target feature information is processed through multiple fully connected layers to obtain multi-task output results of accumulation volume, biomass and carbon storage, which can quickly and accurately estimate the forest resource volume of the target forest area. This solution overcomes the limitations of the sample plot inventory method in large-scale data collection, achieves comprehensive coverage of the entire forest area, and thus achieves the technical effect of improving the efficiency of estimating forest resources.

[0129] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the target backbone network includes at least a convolutional layer, a channel and spatial attention layer, multiple residual block sequences and a self-attention layer, and the extraction unit includes: a first extraction subunit, used to perform preliminary feature extraction on the target point cloud data through the convolutional layer to obtain first feature information; a second extraction subunit, used to perform channel layer and spatial layer feature extraction on the first feature information through the channel and spatial attention layer to obtain second feature information; a third extraction subunit, used to perform multi-scale feature extraction on the second feature information through multiple residual block sequences to obtain third feature information; and a first processing subunit, used to process the third feature information through the self-attention layer to obtain target feature information.

[0130] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the second extraction subunit includes: a first extraction module, used to extract channel features of the first feature information to obtain channel feature information; a second extraction module, used to extract spatial features of the channel feature information to obtain spatial feature information; and a first determination module, used to obtain second feature information based on the spatial feature information.

[0131] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the first extraction module includes: an operation submodule for performing a pooling operation on the first feature information to obtain a channel descriptor; a processing submodule for processing the channel descriptor to obtain a weight vector for each channel; and a calculation submodule for calculating the first feature information and the weight vector of each channel to obtain channel feature information.

[0132] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the target deep learning model is trained using the following devices: a second acquisition unit, used to acquire multiple training sets, wherein each training set consists of a point cloud sample and a real resource quantity corresponding to the point cloud sample, and the type of real resource quantity between each training set is different; a first training unit, used to perform single-task training on the first deep learning model through each training set to obtain the trained first deep learning model; a construction unit, used to construct a second deep learning model through the backbone network and multiple fully connected layers of the trained first deep learning model; a third acquisition unit, used to acquire a target training set, wherein the target training set consists of a target point cloud sample and multiple real resource quantities of multiple types corresponding to the target point cloud samples; a second training unit, used to perform multi-task training on the second deep learning model through the target training set to obtain a target deep learning model.

[0133] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the second training unit includes: a processing subunit, used to process the target point cloud sample through the second deep learning model to obtain multiple predicted resource quantities of multiple types; an acquisition subunit, used to obtain the weight value of the resource quantity corresponding to each type; a calculation subunit, used to calculate based on multiple predicted resource quantities, multiple real resource quantities of multiple types and weight values ​​to obtain a loss function; a training subunit, used to train the second deep learning model based on the loss function to obtain a target deep learning model.

[0134] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the first acquisition unit includes: a first screening subunit, used to coarsely screen the initial point cloud data to obtain first point cloud data; a second screening subunit, used to finely screen the first point cloud data to obtain second point cloud data; and a second processing subunit, used to thin out and normalize the second point cloud data to obtain target point cloud data.

[0135] Optionally, in the deep learning-based forest resource estimation device provided in an embodiment of the present application, the second screening sub-unit includes: a first screening module, used to screen the first point cloud data through a density-based clustering algorithm to obtain screened first point cloud data; a second determination module, used to obtain a point cloud distribution reference model based on the tree species types and tree species distribution in the target forest area; a third screening module, used to screen the screened first point cloud data through the point cloud distribution reference model to obtain second point cloud data.

[0136] It should be noted that the first acquisition unit 701, extraction unit 702, processing unit 703, and determination unit 704 described above correspond to steps S201 to S204 in Example 1. The examples and application scenarios implemented by the four units and the corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Example 1. It should be noted that the above-mentioned modules or units can be hardware components or software components stored in a memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The above-mentioned modules can also be run as part of the device in the computer terminal 10 provided in Example 1.

[0137] Example 3

[0138] An embodiment of the present application may provide an electronic device, Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present application. Figure 8 As shown, the electronic device may include: one or more ( Figure 8 Only one is shown) processor 802, memory 804, storage controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0139] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the methods and devices in the embodiments of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, implementing the above-mentioned method. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0140] The processor can call the information and application programs stored in the memory through the transmission device to perform the following steps: obtain the initial point cloud data of the target forest area, and pre-process the initial point cloud data to obtain the target point cloud data; extract features of the target point cloud data through the target backbone network of the target deep learning model to obtain target feature information; process the target feature information through multiple fully connected layers of the target deep learning model to obtain the output results corresponding to each fully connected layer; and obtain the forest resource amount of the target forest area based on the output results corresponding to each fully connected layer.

[0141] The processor can call the information and application programs stored in the memory through the transmission device to perform the following steps: the target backbone network includes at least a convolution layer, a channel and spatial attention layer, multiple residual block sequences and a self-attention layer, and the target point cloud data is subjected to feature extraction through the target backbone network of the target deep learning model to obtain target feature information, including: performing preliminary feature extraction on the target point cloud data through the convolution layer to obtain first feature information; performing channel layer and spatial layer feature extraction on the first feature information through the channel and spatial attention layer to obtain second feature information; performing multi-scale feature extraction on the second feature information through multiple residual block sequences to obtain third feature information; processing the third feature information through the self-attention layer to obtain target feature information.

[0142] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: extracting channel layer and spatial layer features of the first feature information through the channel and spatial attention layers to obtain the second feature information, including: extracting the channel features of the first feature information to obtain channel feature information; extracting the spatial features of the channel feature information to obtain spatial feature information; and obtaining the second feature information based on the spatial feature information.

[0143] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: extract the channel features of the first feature information to obtain the channel feature information, including: performing a pooling operation on the first feature information to obtain a channel descriptor; processing the channel descriptor to obtain a weight vector for each channel; calculating the first feature information and the weight vector of each channel to obtain the channel feature information.

[0144] The processor can call the information and application programs stored in the memory through the transmission device to perform the following steps: The target deep learning model is trained using the following steps: multiple training sets are obtained, wherein each training set consists of a point cloud sample and a real resource amount corresponding to the point cloud sample, and the type of the real resource amount between each training set is different; single-task training is performed on the first deep learning model through each training set to obtain the trained first deep learning model; a second deep learning model is constructed through the backbone network and multiple fully connected layers of the trained first deep learning model; a target training set is obtained, wherein the target training set consists of a target point cloud sample and multiple real resource amounts of multiple types corresponding to the target point cloud samples; multi-task training is performed on the second deep learning model through the target training set to obtain the target deep learning model.

[0145] The processor can call the information and application programs stored in the memory through the transmission device to perform the following steps: multi-task training is performed on the second deep learning model through the target training set to obtain the target deep learning model, including: processing the target point cloud samples through the second deep learning model to obtain multiple predicted resource quantities of multiple types; obtaining the weight value of the resource quantity corresponding to each type; calculating based on the multiple predicted resource quantities, multiple real resource quantities of multiple types and the weight values ​​to obtain the loss function; training the second deep learning model based on the loss function to obtain the target deep learning model.

[0146] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: preprocessing the initial point cloud data to obtain target point cloud data, including: coarse screening of the initial point cloud data to obtain first point cloud data; fine screening of the first point cloud data to obtain second point cloud data; and thinning and normalizing the second point cloud data to obtain target point cloud data.

[0147] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: finely screen the first point cloud data to obtain the second point cloud data, including: filtering the first point cloud data through a density-based clustering algorithm to obtain the filtered first point cloud data; obtaining a point cloud distribution reference model based on the tree species types and tree species distribution in the target forest area; filtering the filtered first point cloud data through the point cloud distribution reference model to obtain the second point cloud data.

[0148] It can be understood by those skilled in the art that Figure 8 The structure shown is for illustration only, and the electronic device may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (MID), a PAD, or other terminal devices. Figure 8 It does not limit the structure of the above electronic device. For example, the electronic device may also include Figure 8 More or fewer components (such as network interfaces, display devices, etc.) shown in, or with Figure 8 Different configurations shown.

[0149] A person skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0150] Example 4

[0151] The embodiment of the present application further provides a computer-selectable storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the forest resource estimation method based on deep learning provided in the first embodiment.

[0152] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0153] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing the program steps of the forest resource estimation method based on deep learning.

[0154] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0155] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0156] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0157] The units described as separate components may or may not be physically separate, and the components shown as units 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 these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0158] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0159] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0160] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A forest resource estimation method based on deep learning, characterized in that: include: Acquiring initial point cloud data of a target forest area, and preprocessing the initial point cloud data to obtain target point cloud data; Performing feature extraction on the target point cloud data through the target backbone network of the target deep learning model to obtain target feature information; Processing the target feature information through multiple fully connected layers of the target deep learning model to obtain an output result corresponding to each fully connected layer; According to the output results corresponding to each fully connected layer, the forest resource volume of the target forest area is obtained.

2. The method according to claim 1, characterized in that The target backbone network includes at least a convolutional layer, a channel and spatial attention layer, a plurality of residual block sequences and a self-attention layer. The target point cloud data is subjected to feature extraction by the target backbone network of the target deep learning model, and the target feature information obtained includes: Performing preliminary feature extraction on the target point cloud data through the convolution layer to obtain first feature information; Performing channel-level and spatial-level feature extraction on the first feature information through the channel and spatial attention layers to obtain second feature information; Performing multi-scale feature extraction on the second feature information using the multiple residual block sequences to obtain third feature information; The third feature information is processed by the self-attention layer to obtain the target feature information.

3. The method according to claim 2, characterized in that Performing channel layer and spatial layer feature extraction on the first feature information through the channel and spatial attention layers to obtain second feature information includes: Extracting channel features from the first feature information to obtain channel feature information; Extracting spatial features of the channel feature information to obtain spatial feature information; The second feature information is obtained according to the spatial feature information.

4. The method according to claim 3, characterized in that Extracting the channel feature of the first feature information to obtain the channel feature information includes: Performing a pooling operation on the first feature information to obtain a channel descriptor; Processing the channel descriptors to obtain a weight vector for each channel; The first feature information and a weight vector of each channel are calculated to obtain the channel feature information.

5. The method according to claim 1, characterized in that The target deep learning model is trained using the following steps: Acquire multiple training sets, wherein each training set consists of a point cloud sample and a real resource quantity corresponding to the point cloud sample, and the type of the real resource quantity in each training set is different; Performing single-task training on the first deep learning model using each training set to obtain a trained first deep learning model; Constructing a second deep learning model through the backbone network and multiple fully connected layers of the trained first deep learning model; Acquire a target training set, wherein the target training set consists of a target point cloud sample and a plurality of real resource quantities of various types corresponding to the target point cloud sample; Multi-task training is performed on the second deep learning model using the target training set to obtain the target deep learning model.

6. The method according to claim 5, characterized in that Performing multi-task training on the second deep learning model using the target training set to obtain the target deep learning model includes: Processing the target point cloud sample by the second deep learning model to obtain multiple types of predicted resource quantities; Get the weight value of the resource amount corresponding to each type; Calculating based on the multiple predicted resource quantities, the multiple actual resource quantities of the multiple types, and the weight values ​​to obtain a loss function; The second deep learning model is trained according to the loss function to obtain the target deep learning model.

7. The method according to claim 1, characterized in that Preprocessing the initial point cloud data to obtain target point cloud data includes: Performing a rough screening on the initial point cloud data to obtain first point cloud data; performing fine screening on the first point cloud data to obtain second point cloud data; The second point cloud data is thinned and normalized to obtain the target point cloud data.

8. The method according to claim 7, characterized in that The first point cloud data is finely screened to obtain second point cloud data including: Filtering the first point cloud data using a density-based clustering algorithm to obtain filtered first point cloud data; Obtaining a point cloud distribution reference model based on the tree species types and tree species distribution in the target forest area; The filtered first point cloud data is filtered using the point cloud distribution reference model to obtain the second point cloud data.

9. A forest resource estimation device based on deep learning, characterized in that: include: A first acquisition unit is configured to acquire initial point cloud data of a target forest area and pre-process the initial point cloud data to obtain target point cloud data; An extraction unit, configured to extract features from the target point cloud data using a target backbone network of a target deep learning model to obtain target feature information; A processing unit, configured to process the target feature information through multiple fully connected layers of the target deep learning model to obtain an output result corresponding to each fully connected layer; The determination unit is used to obtain the forest resource amount of the target forest area according to the output results corresponding to each fully connected layer.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored executable program, wherein, when the executable program is running, the device where the computer-readable storage medium is located is controlled to execute the forest resource estimation method based on deep learning as described in any one of claims 1 to 8.

11. An electronic device, characterized in that: include: a memory storing an executable program; A processor for running the program, wherein the program, when running, executes the forest resource estimation method based on deep learning as described in any one of claims 1 to 8.

Images