Overground biomass calculation method and system based on deep learning

Abstract

The invention provides an aboveground biomass calculation method and system based on deep learning. The method comprises the following steps: acquiring geographic position information of a forest region and unmanned aerial vehicle and satellite image data at corresponding positions; randomly sampling in the forest to obtain data of all arbor species, plant height and diameter at breast height in a quadrat and a different-speed growth equation corresponding to each tree; calculating the average AGB of each ground quadrat; constructing a deep learning calculation network of unmanned aerial vehicle multi-source data and a ground AGB; inverting AGB data of the flight area of the unmanned aerial vehicle; constructing a satellite image-biomass calculation model; estimating large-range AGB distribution of the target area; and carrying out weight visualization analysis on the unmanned aerial vehicle image-biomass calculation network. According to the invention, accurate estimation of the AGB can be realized by using the high-resolution image of the unmanned aerial vehicle. Compared with a traditional method, the method avoids the defects that a traditional measurement method is time-consuming and labor-consuming, and a traditional machine learning method is insufficient in representativeness.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular, to a method and system for calculating above-ground biomass based on deep learning. Background Art

[0002] Forest ecosystems are important carbon sinks, and forest carbon is mainly stored in the form of biomass. Therefore, accurately estimating forest biomass is crucial. Since above-ground biomass (AGB) is considered the most dynamic, visible, and important among all forms of biomass, most biomass estimations focus on above-ground biomass.

[0003] Traditionally, large-scale estimation of AGB is achieved by dividing each region and sampling to statistically calculate the average AGB. However, this method has low accuracy and cannot adapt to scenarios that require fine AGB distribution. With the development of remote sensing technology, the use of remote sensing data to achieve large-spatial-scale estimation of forest AGB has been widely applied. Most existing methods are to construct a model between satellite remote sensing data and AGB, so as to use large-scale satellite images for large-scale estimation of AGB. However, the satellite image coverage is wide, and the AGB data collected manually is narrowly distributed and does not have sufficient representativeness, resulting in low accuracy of this method. UAV images have the advantages of high resolution and relatively wide coverage. Using UAV images, a more accurate remote sensing data-biomass model can be constructed. Therefore, using the high-resolution images of UAVs can create AGB data distributed in a relatively wide area and weaken the problem of insufficient representativeness. The high resolution of UAV images results in a large number of pixel inputs with spatial distribution characteristics corresponding to each AGB output in the model, which makes it necessary to analyze a large amount of data with spatial characteristics in an innovative way.

[0004] Deep learning methods are excellent feature extractors. Among them, Convolutional Neural Networks (CNN) use convolution and pooling operations to achieve powerful image spatial feature extraction functions and are widely used in the field of machine vision. Compared with manually extracting spatial features in images (such as calculating mean, variance, information entropy, etc.), convolutional neural networks can automatically extract features through multiple hidden layers.

[0005] Currently, there is no effective method for combining deep learning methods with AGB estimation. Summary of the Invention

[0006] The present invention provides a method and system for calculating above-ground biomass based on deep learning, which is used to solve the defect that deep learning is not applied in AGB estimation in the prior art and realize high-resolution AGB estimation on a large spatial scale.

[0007] In a first aspect, the present invention provides a method for calculating aboveground biomass based on deep learning, including: Obtain the geographical location information of the target forest area, and collect the multi-source data of unmanned aerial vehicles (UAVs) and satellite image data of the target forest area; Randomly sample within the target forest area, obtain the set of all tree information within the quadrat, and calculate the average AGB of each quadrat using the set of tree information; Based on the UAV multi-source data, the satellite image data, and the set of tree information, construct a UAV image-biomass calculation network; Train the UAV image-biomass calculation network, and invert the AGB data of the UAV flight area; Use the UAV image-biomass calculation network and the satellite image data, and perform fitting through the random forest algorithm to construct a satellite image-biomass calculation model; Train the satellite image-biomass calculation model to estimate the preset large-scale AGB data of the target forest area.

[0008] According to the method for calculating aboveground biomass based on deep learning provided by the present invention, it further includes: Conduct weight visualization analysis on the UAV image-biomass calculation network.

[0009] According to the method for calculating aboveground biomass based on deep learning provided by the present invention, randomly sample within the target forest area, obtain the set of all tree information within the quadrat, and calculate the average AGB of each quadrat using the set of tree information, including: Determine that the set of tree information includes all tree species, tree height, diameter at breast height data, and the allometric growth equation corresponding to each tree; Estimate the average AGB of each quadrat using the tree height, diameter at breast height data, and the allometric growth equation:

[0010] Wherein, represents aboveground biomass, represents diameter at breast height, represents tree height, represents the allometric growth equation.

[0011] According to the method for calculating aboveground biomass based on deep learning provided by the present invention, based on the UAV multi-source data, the satellite image data, and the set of tree information, construct a UAV image-biomass calculation network, including: Obtain the digital surface elevation model (DSM) point data of the bare soil area near a large number of forest areas, and calculate the canopy height of the forest area:

[0012]

[0013] Among them, represents an unknown point to a known point the L2 distance of, represents the weight coefficient of the i-th known point, represents the digital surface elevation point data of the extracted bare soil area, represents the canopy height at the k position, represents the digital surface elevation at the k position; Taking the forest canopy height, the multi-spectral bands of the unmanned aerial vehicle, the derived vegetation indices, and the digital surface elevation model as the inputs of the unmanned aerial vehicle image-biomass calculation network; It is determined that the unmanned aerial vehicle image-biomass calculation network includes a plurality of convolution-pooling modules and a fully connected layer, extracts different-scale information through convolutions of different sizes, splices different-scale information through max pooling, and converts the extracted feature information into AGB results through the fully connected layer.

[0014] According to a method for calculating aboveground biomass based on deep learning provided by the present invention, it is determined that the unmanned aerial vehicle image-biomass calculation network includes a plurality of convolution-pooling modules and a fully connected layer, extracts different-scale information through convolutions of different sizes, splices different-scale information through max pooling, and converts the extracted feature information into AGB results, including: The convolution-pooling module includes a convolutional layer, a normalization layer, an activation layer, and a max pooling layer, and the output result is:

[0015] Among them, represents the input of the module at channel, and the position is ; represents the output of the module at channel, and the position is ; represents the convolution operation, represents the normalization operation, is the activation function, and the Relu function is adopted, represents the downsampling operation; The structure of the convolutional layer is:

[0016] Among them, represents that the convolutional layer is connected to the l th layer and the j th feature map position is The output, indicates that the m position of the th convolutional kernel is l the value of, and this convolutional kernel is connected to the j th feature map in the Hl ( Wl -1)th layer, bl,j and l represent the height and width of the convolutional kernel respectively, j represents the bias of the th feature map in the

[0017]

[0018]

[0019] where represents the th sample, the th channel, and the input at position , represents the mean of each channel in this batch, represents the variance of each channel in this batch, represents a small positive value used to avoid a zero denominator, is the scaling parameter, is the offset parameter, and both are learnable parameters, represents the output of the normalization layer; After the normalization layer, a skip connection operation is performed:

[0020] where represents the input of the module at channel and position , where represents the output of the module at channel and position , represents the convolution operation, represents the normalization operation; The max pooling layer is used to implement the downsampling operation:

[0021] where S, K represent the stride of the pooling window asS , with a size of (K, K); denotes the position under the channel is the input of, denotes the position under the channel is the output of; The extracted features are transformed into AGB results through a fully connected layer:

[0022] Among them, is the extracted feature, and are the weight matrix and bias parameters to be learned.

[0023] According to a method for calculating aboveground biomass based on deep learning provided by the present invention, the UAV image-biomass calculation network is trained, and the AGB data of the UAV flight area is inverted, including: Determine the loss function MSE according to the UAV multi-source data and the ground AGB data, and train the UAV image-biomass calculation network; By rotating the picture matrix of the input picture, the position of each reflectance or elevation pixel in all channel image matrices is obtained:

[0024] Among them, represents the rotated pixel position, , represent the center point of the image matrix, represents the rotation angle; The loss function MSE is:

[0025] Among them, is the true value, is the corresponding predicted value, is the number of samples; The training results are analyzed using a preset evaluation index.

[0026] According to a method for calculating aboveground biomass based on deep learning provided by the present invention, using the UAV image-biomass calculation network and the satellite image data, a satellite image-biomass calculation model is constructed through fitting by the random forest algorithm, including:

[0027] Among them, represents the predicted value generated by a single decision tree of the random forest algorithm, Represents the number of decision trees in the random forest algorithm. Represents the final output value of the random forest algorithm.

[0028] According to a method for calculating aboveground biomass based on deep learning provided by the present invention, a weight visualization analysis is performed on the UAV image-biomass calculation network, including:

[0029] Among them, the observed value Represents the predicted value of the model for observation i. Is the average predicted value of all observed values. Refers to the SHAP value of the j-th feature of the observed value i, which represents the marginal contribution of the feature to the prediction.

[0030] In a second aspect, the present invention also provides a system for calculating aboveground biomass based on deep learning, including: An acquisition module for acquiring the geographical location information of the target forest area and collecting the UAV multi-source data and satellite image data of the target forest area. A sampling module for randomly sampling within the target forest area to obtain a set of all tree information within the quadrat, and calculating the average AGB of each quadrat using the set of tree information. A construction module for constructing a UAV image-biomass calculation network based on the UAV multi-source data, the satellite image data, and the set of tree information. A training module for training the UAV image-biomass calculation network and inverting the AGB data of the UAV flight area. A fitting module for using the UAV image-biomass calculation network and the satellite image data to perform fitting through the random forest algorithm to construct a satellite image-biomass calculation model. A calculation module for training the satellite image-biomass calculation model and calculating the preset large-range AGB data of the target forest area.

[0031] In a third aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the method for calculating aboveground biomass based on deep learning as described in any one of the above.

[0032] The method and system for calculating aboveground biomass based on deep learning provided by the present invention raise the level of data information in estimating biomass from simply using unmanned aerial vehicle (UAV) information to using high-resolution UAV images across spatial positions. Compared with directly using the average spectral reflectance and average elevation, the present invention employs a convolutional neural network in the spatial domain of the image, and uses multiple hidden layers to achieve automatic extraction of spatial features and estimation of AGB. Compared with the ordinary method of directly fitting satellite data, multi-source fusion of ground-air-space data is carried out, and high-resolution UAV data is used to enhance the representativeness of the model. The present invention presents the model results in a visual way to extract the model's input features and two-dimensional spatial features, and verifies the interpretability of the model. Description of the Drawings

[0033] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0034] Figure 1 is a schematic flowchart of the method for calculating aboveground biomass based on deep learning provided by the present invention; Figure 2 is a structural diagram of the UAV image-biomass calculation neural network for calculating AGB using UAV data provided by the present invention; Figure 3 is an accuracy graph comparing the UAV image-biomass calculation neural network provided by the present invention with the traditional network; Figure 4 is a schematic diagram comparing the calculation results of the UAV image-biomass calculation neural network provided by the present invention with the calculation results of the traditional network; Figure 5 is a scatter plot of the satellite image-biomass calculation model provided by the present invention on the test set; Figure 6 is the final large-scale 20m resolution AGB distribution map provided by the present invention; Figure 7 is a SHAP importance analysis graph of each input parameter level of the UAV image-biomass calculation neural network provided by the present invention; Figure 8 is a SHAP importance analysis graph of each input of the UAV image-biomass calculation neural network at the spatial level provided by the present invention; Figure 9 is a structural diagram of the system for calculating aboveground biomass based on deep learning provided by the present invention; Figure 10It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0035] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0036] Figure 1 It is a schematic flowchart of the above-ground biomass calculation method based on deep learning provided by an embodiment of the present invention. As Figure 1 shown, it includes: Step 100: Obtain the geographical location information of the target forest area, and collect the multi-source data of the unmanned aerial vehicle (UAV) and satellite image data of the target forest area; Step 200: Randomly sample within the target forest area, obtain the set of all tree information within the quadrat, and calculate the average AGB of each quadrat by using the set of tree information; Step 300: Based on the multi-source data of the UAV, the satellite image data and the set of tree information, construct a UAV image-biomass calculation network; Step 400: Train the UAV image-biomass calculation network, and invert the AGB data of the UAV flight area; Step 500: Use the UAV image-biomass calculation network and the satellite image data, and perform fitting through the random forest algorithm to construct a satellite image-biomass calculation model; Step 600: Train the satellite image-biomass calculation model to estimate the preset large-range AGB data of the target forest area.

[0037] Specifically, as Figure 2 shown, the process of the embodiment of the present invention includes: Step 1: Obtain the geographical location information of the forest area and the corresponding UAV and satellite image data at the corresponding location. Obtain the distribution map of the forest area by obtaining the division of the forest area by the local government. Use an agricultural multi-spectral UAV to take pictures in random areas of the forest area. The reflectance information obtained by the UAV includes five spectral bands of red, green, blue, near-infrared and red edge. At the same time, obtain the digital surface model (DSM) data of the flight area by the UAV. The satellite data acquisition includes the global canopy height data ETH Global Sentinel-2 10m Canopy Height (2020), and log in to the official website of the European Space Agency to obtain the sentinel-2 satellite L2A product.

[0038] Step 2: Randomly sample within the forest area, divide the plots of 20 * 20 m, record the geographical location information of the plots, record the tree species of each tree within the plots, measure its tree height and diameter at breast height, and use the Internet to obtain the allometric equations of each tree species in similar areas. Use arcgis software to crop the drone images to obtain the drone image data of each plot.

[0039] Step 3: Calculate the AGB for each tree using the tree species, tree height, diameter at breast height, and allometric equations collected in Step 2. To avoid large-scale destructive sampling, the present invention estimates the ground AGB using the tree height, diameter at breast height, and the corresponding allometric equations, which can be expressed by the following formula:

[0040] Where, represents the aboveground biomass, represents the diameter at breast height, represents the tree height, represents the allometric equation.

[0041] Add the AGB of each tree within the plot to obtain the average AGB within the plot.

[0042] Step 4: Combine the data collected in Steps 1 and 2 to construct a deep learning calculation network for drone multi-source data and ground AGB. The drone image-biomass calculation network accepts the drone multispectral bands and their derived vegetation indices, digital surface model (DSM), and canopy height (CH) data as inputs. According to the resolution of the drone equipment, bilinearly resample the drone image data of each plot to a size of 360×360 as the input of the model.

[0043] For the DSM digital surface elevation image collected by the drone, obtain a large amount of DSM point data in the bare soil area near the forest area, and calculate the canopy height (CH) of the forest area based on this, which can be expressed by the following formula:

[0044]

[0045] Where, represents the unknown point to the known point the L2 distance of, represents the weight coefficient of the i-th known point, represents the extracted digital surface elevation point data of the bare soil area, represents the canopy height at the k position, represents the digital surface elevation at the k position; The UAV image-biomass calculation network consists of multiple convolution-pooling modules and a fully connected layer. Different scale information is extracted through convolutions of different sizes, and then the information of different scales is spliced through max pooling. Finally, the extracted feature information is converted into AGB results through a fully connected layer.

[0046] The convolution-pooling module consists of a convolutional layer, a normalization layer, an activation layer, and a max pooling layer. The output result can be expressed by the following formula:

[0047] where represents the input of the module at channel and position , represents the output of the module at channel and position ; represents the convolution operation, represents the normalization operation, is the activation function, using the Relu function, represents the downsampling operation; The convolution operation is implemented using a convolutional layer. The structure of the convolutional layer can be expressed by the following formula:

[0048] where represents the output of the convolutional layer at the position of the l th feature map in the j th layer, represents the value at the position of the m th convolutional kernel, which is connected to the th feature map in the ( l - 1)th layer, j Hl and Wl are the height and width of the convolutional kernel respectively, bl,j l represents the bias of the j th feature map in the th layer; The normalization layer operation is implemented using batch normalization (BN), and each feature is normalized within the current batch. For an input matrix with a batch size of N, the number of input features (i.e., channels) is C, and the length and width are H and W respectively. The normalization layer can be expressed by the following formula:

[0049]

[0050]

[0051] Among them, represents the th sample, the th channel, and the input at position ; represents the mean of each channel in this batch, represents the variance of each channel in this batch, represents a small positive value used to avoid a zero denominator, is the scaling parameter, is the offset parameter, and both are learnable parameters, represents the output of the normalization layer; After the normalization layer, a skip connection operation is performed. It is implemented by means of a residual connection, that is, the output of the previous module is directly added to the output of the normalization layer, thereby improving the overfitting and gradient vanishing problems caused by the excessive depth of the model. The skip connection can be expressed by the following formula:

[0052] Among them represents the input of the module in the channel at position , where represents the output of the module in the channel at position ; represents the convolution operation, represents the normalization operation; The downsampling operation is implemented using a max pooling layer (maxpool), which can be expressed as:

[0053] Among them, S, K respectively represent that the stride of the pooling window is S , and the size is (K, K); represents the input at position in the channel, represents the output at position in the

[0054] Finally, the extracted features are transformed into the AGB result through a fully connected layer (FNN): is the extracted feature, and are the weight matrix and bias parameter to be learned.

[0055] Step 5: Train the UAV image-biomass calculation network in Step 4 and invert the AGB data of the entire UAV flight area; in Step 5, set the loss function MSE according to the UAV multi-source data and the ground AGB data, and train the UAV image-biomass calculation network in Step 4.

[0056] Before training, perform data augmentation on the input pictures by rotating the picture matrix multiple times. For each reflectance or elevation pixel in all-channel image matrices:

[0057] where, represents the rotated pixel position, and represent the center point of the image matrix, represents the rotation angle; The loss function MSE is:

[0058] where, is the true value, is the corresponding predicted value, is the number of samples; During the network training process, the model adopts the Adam optimizer, sets the learning rate lr to 10-3, the training batch size batch-size to 20, and the number of training epochs epoch to 200. Analyze the training results after training. The comparison of the training results of the network used in the present invention with those of traditional machine learning methods is as Figure 3 shown. The machine learning algorithms participating in the comparison are least squares regression algorithm (LR), partial least squares regression algorithm (PLSR), support vector machine algorithm (SVR), lasso regression algorithm (Lasso), elastic net algorithm (ElasticNet) and random forest algorithm (RF). Using the coefficient of determination R2, root mean square error RMSE and mean absolute squared error MAE as evaluation indicators, their calculation formulas are as follows:

[0059]

[0060]

[0061] Use the trained UAV image-biomass calculation network in Step 5 to invert the AGB distribution map covering all UAV flight ranges. The comparison of the network output used in this method with the output of traditional machine learning methods is as Figure 4 shown.

[0062] Step 6: Fit the output of the UAV image-biomass calculation network and the satellite image input through the random forest algorithm to construct a satellite image-biomass calculation model. The random forest algorithm (RF) is a machine learning method that can be used for classification and regression tasks. The random forest algorithm encompasses a variety of different algorithms, each of which is analogized to a tree. It randomly samples data and features, and each tree is trained on independent small batches of samples and the total data, and then makes a final prediction by averaging. The formula is as follows:

[0063] where ft(X) represents the predicted value generated by the corresponding tree, represents the number of decision trees in the random forest algorithm, represents the final output value of the random forest algorithm.

[0064] Step 7: Use the large amount of AGB data generated in Step 5 for training, and analyze the training results. The coefficient of determination R2, root mean square error RMSE, and mean absolute error MAE are used as evaluation indicators. The scatter plot of the satellite image-biomass calculation model on the test set is as Figure 5 shown.

[0065] Estimate the AGB distribution of the entire target area using the trained satellite image-biomass calculation model. The result is as Figure 6 shown.

[0066] Step 8: Conduct a weight visualization analysis on the UAV image-biomass calculation network in Step 4. According to the UAV image-biomass calculation network trained in Step 5, perform SHAP analysis at the parameter level and two-dimensional space level respectively to visualize the contributions of each part and analyze the internal mechanism of the model. The SHapley Additive exPlanations (SHAP) algorithm believes that the output of the model is composed of the base value and the SHAP values related to each feature added together. The basic formula is as follows:

[0067] where the observed value represents the predicted value of the model for observation i, is the average predicted value of all observed values, refers to the SHAP value of the jth feature of the observed value i, and this value represents the marginal contribution of the feature to the prediction.

[0068] The SHAP importance analysis diagrams of each input parameter level of the UAV image-biomass calculation neural network in the embodiments of the present invention are as Figure 7 shown. In Figure 7Among them, the larger the SHAP value, the greater the contribution of the parameter to the result. The ordinate represents each input of the model, specifically the multispectral input, the digital surface elevation input, and its derived inputs. It can be seen from the figure that the digital surface elevation (DSM) and the derived canopy height (CH) data have the greatest impact on the output of the model, followed by the ratio vegetation index (RVI).

[0069] The SHAP importance analysis diagram of each input of the UAV image-biomass calculation neural network according to the embodiment of the present invention at the spatial level is as Figure 8 shown. In Figure 8 each subfigure's title represents each input feature, including the multispectral input, the digital surface elevation input, and its derived inputs. The more obvious the color area, the greater the contribution to the result. Taking subfigure (k) as an example, the top part of the tree canopy in the left figure corresponds to the dark area in the right figure, indicating that the AGB in the area with a large canopy height is large, which conforms to common sense.

[0070] The following describes the above-ground biomass calculation system based on deep learning provided by the present invention. The above-ground biomass calculation system based on deep learning described below can be mutually corresponding and referenced with the above-ground biomass calculation method described above.

[0071] Figure 9 is a schematic structural diagram of the above-ground biomass calculation system based on deep learning provided by the embodiment of the present invention, as Figure 9 shown, including: an acquisition module 91, a sampling module 92, a construction module 93, a training module 94, a fitting module 95, and a calculation module 96, wherein: The acquisition module 91 is used to acquire the geographical location information of the target forest area, and collect the UAV multi-source data and satellite image data of the target forest area; the sampling module 92 is used to randomly sample within the target forest area, obtain the set of all tree information in the quadrat, and calculate the average AGB of each quadrat by using the set of tree information; the construction module 93 is used to construct a UAV image-biomass calculation network based on the UAV multi-source data, the satellite image data, and the set of tree information; the training module 94 is used to train the UAV image-biomass calculation network and invert the AGB data of the UAV flight area; the fitting module 95 is used to use the UAV image-biomass calculation network and the satellite image data to perform fitting through the random forest algorithm to construct a satellite image-biomass calculation model; the calculation module 96 is used to train the satellite image-biomass calculation model and calculate the preset large-range AGB data of the target forest area.

[0072] Figure 10 illustrates a schematic physical structure diagram of an electronic device, as Figure 10As shown in the figure, the electronic device may include: a processor 1010, a communications interface 1020, a memory 1030, and a communication bus 1040. Among them, the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 may call the logical instructions in the memory 1030 to execute a method for calculating above-ground biomass based on deep learning. The method includes: obtaining the geographical location information of the target forest area, collecting the multi-source data of the unmanned aerial vehicle and satellite image data of the target forest area; randomly sampling within the target forest area to obtain the set of all tree information within the quadrat, and calculating the average AGB of each quadrat using the set of tree information; based on the multi-source data of the unmanned aerial vehicle, the satellite image data, and the set of tree information, constructing an unmanned aerial vehicle image-biomass calculation network; training the unmanned aerial vehicle image-biomass calculation network and inverting the AGB data of the flight area of the unmanned aerial vehicle; using the unmanned aerial vehicle image-biomass calculation network and the satellite image data, performing fitting through the random forest algorithm to construct a satellite image-biomass calculation model; training the satellite image-biomass calculation model to estimate the preset large-scale AGB data of the target forest area.

[0073] In addition, when the logical instructions in the above-mentioned memory 1030 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0074] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative effort.

[0075] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A method for calculating aboveground biomass based on deep learning, characterized in that: include: Obtaining geographic location information of a target forest area, and collecting multi-source drone data and satellite image data of the target forest area; Random sampling is performed in the target forest area to obtain a set of information about all trees in the sample plot, and the average aboveground biomass AGB of each sample plot is calculated using the set of tree information; Based on the drone multi-source data, the satellite image data and the tree information set, a drone image-biomass calculation network is constructed; Training the UAV image-biomass calculation network and inverting AGB data of the UAV flight area; Using the UAV image-biomass calculation network and the satellite image data, a satellite image-biomass calculation model is constructed by fitting through a random forest algorithm; The satellite image-biomass calculation model is trained to estimate the preset large-scale AGB data of the target forest area.

2. The aboveground biomass calculation method based on deep learning according to claim 1, characterized in that: Also includes: A weighted visualization analysis of the UAV image-biomass calculation network was performed.

3. The aboveground biomass calculation method based on deep learning according to claim 1, characterized in that: Random sampling is performed in the target forest area to obtain a set of information about all trees in the sample plot, and the average AGB of each sample plot is calculated using the set of tree information, including: Determine that the tree information set includes all tree species, plant height, diameter at breast height data and allometric growth equations corresponding to each tree; The average AGB of each plot was estimated using plant height, DBH data and the allometric equation: in, represents aboveground biomass, Represents the breast diameter, Represents plant height, Represents the allometric growth equation.

4. The aboveground biomass calculation method based on deep learning according to claim 1, characterized in that: Based on the drone multi-source data, the satellite image data and the tree information set, a drone image-biomass calculation network is constructed, including: Obtain a large amount of digital surface elevation model (DSM) point data of bare soil areas near forest areas and calculate the canopy height of forest areas: in, Indicates unknown point To a known point The L2 distance, represents the weight coefficient of the i-th known point, Represents the extracted digital surface elevation point data of bare soil area, represents the canopy height at position k, represents the digital surface elevation at location k; The forest canopy height, drone multispectral bands, derived vegetation index and digital surface elevation model were used as inputs to the drone image-biomass calculation network; It is determined that the UAV image-biomass calculation network includes multiple convolution-pooling modules and a fully connected layer. Different scale information is extracted through convolutions of different sizes, different scale information is spliced ​​through maximum pooling, and the extracted feature information is converted into AGB results through the fully connected layer.

5. The aboveground biomass calculation method based on deep learning according to claim 4, characterized in that: It is determined that the UAV image-biomass calculation network includes multiple convolution-pooling modules and a fully connected layer, extracts information of different scales through convolutions of different sizes, splices information of different scales through maximum pooling, and converts the extracted feature information into AGB results through the fully connected layer, including: The convolution-pooling module includes a convolution layer, a normalization layer, an activation layer, and a maximum pooling layer. The output is: in, Indicates that the module is Channel, location Input, Indicates that the module is Channel, location Output: represents the convolution operation, Represents standardized operation, As the activation function, the Relu function is used. represents the downsampling operation; The convolutional layer structure is: in, Indicates that the convolutional layer is connected to the l Layer j The feature map location is The output, Indicates m The convolution kernel position is The value of the convolution kernel is connected to the first ( l -1) in the layer j feature map, Hl and Wl are the height and width of the convolution kernel, respectively. bl,j Representative l Tier j The bias of the feature map; Batch normalization is used to perform the normalization layer operation, and each feature is normalized in the current batch. For each batch size N, the number of channels C, and the input matrix with length and width H and W respectively, the normalization layer is expressed as: in, Indicates Samples, channels, located at Input, Represents the mean value of each channel in this batch, represents the variance of each channel in the batch, Represents a small positive value used to avoid a denominator of 0. is the scaling parameter, is the offset parameter, both of which are learnable parameters. represents the output of the normalization layer; The normalization layer is followed by a skip connection operation: in Indicates that the module is Channel, location The input of Indicates that the module is Channel, location The output, represents the convolution operation, Indicates standardized operation; Use the maximum pooling layer to implement downsampling operation: in, S.K. They represent the stride of the pooling window respectively. S , size is (K,K); express The position under the channel is Input, express The position under the channel is Output: The extracted features are converted into AGB results through a fully connected layer: in, To extract features, and are the weight matrix and bias parameters that need to be learned.

6. The aboveground biomass calculation method based on deep learning according to claim 1, characterized in that: The UAV image-biomass calculation network is trained and the AGB data of the UAV flight area is inverted, including: Determine the loss function MSE according to the multi-source data of the UAV and the ground AGB data, and train the UAV image-biomass calculation network; By rotating the input image matrix, we can get the position of each reflectance or elevation pixel in the image matrix of all channels: in, represents the pixel position after rotation, , represents the center point of the image matrix, Indicates the rotation angle; The loss function MSE is: in, is the true value, is the corresponding predicted value, is the number of samples; The training results are analyzed using preset evaluation indicators.

7. The aboveground biomass calculation method based on deep learning according to claim 1, characterized in that: The UAV image-biomass calculation network and the satellite image data are used to fit the satellite image-biomass calculation model through a random forest algorithm, including: in, represents the predicted value generated by the corresponding tree, represents the number of decision trees in the random forest algorithm, Represents the final output value of the random forest algorithm.

8. The aboveground biomass calculation method based on deep learning according to claim 2, characterized in that: The weighted visualization analysis of the UAV image-biomass calculation network includes: Among them, the observed value represents the model's predicted value for observation i, is the average predicted value of all observations, It refers to the SHAP value of the jth feature of observation i, which represents the marginal contribution of the feature to the prediction.

9. A system for calculating aboveground biomass based on deep learning, characterized in that: include: An acquisition module is used to acquire geographical location information of a target forest area and collect multi-source data and satellite image data of drones of the target forest area; A sampling module is used to randomly sample in the target forest area, obtain all tree information sets in the sample plot, and calculate the average aboveground biomass AGB of each sample plot using the tree information set; A construction module, for constructing a UAV image-biomass calculation network based on the UAV multi-source data, the satellite image data and the tree information set; A training module, used for training the UAV image-biomass calculation network and inverting the AGB data of the UAV flight area; A fitting module, for using the UAV image-biomass calculation network and the satellite image data to fit through a random forest algorithm to construct a satellite image-biomass calculation model; The calculation module is used to train the satellite image-biomass calculation model and calculate the preset large-scale AGB data of the target forest area.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the aboveground biomass calculation method based on deep learning as described in any one of claims 1 to 8 is implemented.

Images