Forest biomass prediction method based on differentiable physical modeling and application thereof

Abstract

The invention relates to the technical field of deep learning, in particular to a forest biomass prediction method based on differentiable physical modeling and application thereof. Key parameters in the physical model are predicted through the deep full-connection neural network, and the insufficient precision caused by parameter uncertainty and simplified hypothesis of a traditional physical model is effectively relieved; by introducing the physical constraint, the physical consistency and interpretability of the inversion result are enhanced, and the generalization of the deep learning model during forest biomass estimation is improved while the precision is considered. The objective of the invention is to solve the problem of how to give consideration to precision and improve generalization of a deep learning model during forest biomass estimation.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a forest biomass prediction method based on differentiable physical modeling and its application. Background Technology

[0002] Aboveground biomass (AGB) is an important indicator of forest carbon sequestration capacity and a key parameter for assessing forest carbon balance.

[0003] Among relevant technical solutions, physical modeling methods based on radiative transfer theory are common methods for forest AGB (Air Quality Barrier) inversion. This method derives forest structure parameters and further estimates forest AGB by establishing the radiative transfer relationship between electromagnetic waves and the vegetation canopy and land surface. While physical modeling methods have clear physical mechanisms and a certain degree of interpretability, their accuracy is limited by factors such as model simplification assumptions, parameter sensitivity, input data uncertainty, and the complexity of parameter adjustment methods. Furthermore, they struggle to fully utilize information from remote sensing data such as large-scale synthetic aperture radar (SAR), resulting in inversion accuracy generally lagging behind data-driven methods.

[0004] Currently, some studies have applied data-driven methods based on deep learning to forest AGB estimation, which has higher fitting ability and estimation accuracy compared to traditional physical modeling methods. However, current forest AGB estimation schemes using deep learning are heavily dependent on training data and can generally only predict forest data under certain specific conditions such as fixed regions and fixed natural conditions. In reality, forest data collected will vary significantly with regional and seasonal conditions. When deep learning models make predictions outside the specific conditions corresponding to the current training data, they are prone to overfitting, resulting in insufficient generalization ability of deep learning models in forest AGB estimation.

[0005] In view of this, this application proposes a new method for forest biomass prediction, which aims to improve the generalization ability of deep learning models when estimating forest biomass while maintaining accuracy. Summary of the Invention

[0006] The main purpose of this application is to provide a forest biomass prediction method based on differentiable physical modeling, which aims to solve the problem of how to improve the generalization of deep learning models when estimating forest biomass while maintaining accuracy.

[0007] To achieve the above objectives, this application provides a forest biomass prediction method based on differentiable physical modeling, applied to a training model with a differentiable physical model as the network skeleton and incorporating a deep fully connected neural network. The method includes the following steps: Obtain the input parameter features from the collected dataset; The input parameter features are preprocessed and then input into the deep fully connected neural network to predict the physical model parameters of the variables. Obtain the variable physical model parameters and input the variable physical model parameters into the differentiable physical model through forward propagation; Obtain the forest biomass prediction results output by the differentiable physical model.

[0008] Optionally, the variable physics model parameters include a first variable physics model parameter, a second variable physics model parameter, and a third variable physics model parameter, each with the following expression:

[0009]

[0010]

[0011] In the formula, The first variable is the physical model parameter. The second variable is the physical model parameter. The third variable is the physical model parameter. Represents the theoretical value of backscattering from a forest volume. Represents the theoretical value of surface backscattering. Represents the extinction coefficient. It represents the incident angle of scattering from the Earth's surface.

[0012] Optionally, preprocessing the input parameter features specifically includes: After converting the input parameter features into tensor form, normalization is performed. The normalization expression includes:

[0013] In the formula, This represents the normalized characteristics of the input parameters; Represents the input feature tensor; This represents the minimum value of the input feature tensor X in the sample dimension; This represents the maximum value of the input feature tensor X in the sample dimension.

[0014] Optionally, the input parameter features include parameter attribute values, synthetic aperture radar parameter features, optical parameter features, and global digital elevation model data and sample plot survey data from spaceborne thermal emission and reflectance radiometers.

[0015] Optionally, the differentiable physical model is a semi-empirical water cloud model, which includes the following steps in calculating the forest biomass prediction results: Obtain the backscattering coefficient, incident angle, and extinction coefficient of the polarization channel; The variable physical model parameters, polarization channel backscattering coefficient, incident angle, and extinction coefficient are input into the computation graph of the semi-empirical water cloud model through forward propagation. The computation graph consists of continuously differentiable basic operational functions and includes automatic differentiation functions.

[0016] Optionally, the training process of the deep fully connected neural network includes: Determine the mean squared error loss of the forest biomass prediction results; Calculate the gradient between the mean squared error loss and the parameters of each layer in the deep fully connected neural network; The deep fully connected neural network is trained by backpropagation based on the gradients of each layer.

[0017] Optionally, the expression for the forest biomass prediction result is:

[0018] In the formula, Represents the forest AGB prediction value; Represents the polarization backscattering coefficient. The first variable is the physical model parameter. The second variable is the physical model parameter. The third variable is the physical model parameter.

[0019] In addition, to achieve the above objectives, this application also provides a forest biomass prediction method based on differentiable physical modeling as described in any of the preceding claims, and its application in forest biomass measurement.

[0020] In addition, to achieve the above objectives, this application also provides a computer system comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the forest biomass prediction method based on differentiable physical modeling as described in any of the preceding claims.

[0021] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the forest biomass prediction method based on differentiable physical modeling as described in any of the preceding claims.

[0022] This application has at least the following beneficial effects: 1. A fusion framework is proposed that integrates a differentiable physical model and a deep fully connected neural network. The neural network is responsible for dynamically parameterizing the key parameters in the physical model and simultaneously optimizing the network weights and biases through backpropagation and gradient descent, thereby achieving a deep fusion of physics-driven and data-driven approaches. 2. Introducing physical constraints into deep fully connected neural networks enhances the physical consistency and interpretability of the inversion results; 3. By dynamically estimating physical parameters through deep learning, the insufficient accuracy of traditional physical models due to parameter uncertainty and simplification assumptions is effectively alleviated; 4. Through simulation experiments, it effectively overcomes the insufficient accuracy of traditional physical models caused by parameter uncertainty and simplification assumptions, and improves the generalization ability and robustness of the model, outperforming traditional physical models, deep learning models and ensemble machine learning models. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the architecture of the training model involved in the embodiments of this application; Figure 2 This is a schematic diagram of forest backscattering in a semi-empirical water cloud model involved in an embodiment of this application; Figure 3 This is a graph showing the trend of the mean gradient of parameters in each layer of the deep fully connected neural network involved in the embodiments of this application as a function of the number of training rounds; Figure 4 This is a graph of the AGB loss function involved in the embodiments of this application; Figure 5 This is a scatter plot of the observed AGB and predicted AGB of the DPM model involved in the embodiments of this application; Figure 6 This is a scatter plot of the observed AGB and predicted AGB of the GRNN model involved in the embodiments of this application; Figure 7 This is a scatter plot of the observed AGB and predicted AGB of the FNN model involved in the embodiments of this application; Figure 8 This is a scatter plot of the observed AGB and predicted AGB of the Adaboost model involved in the embodiments of this application; Figure 9 This is a scatter plot of the observed AGB and predicted AGB of the RF model involved in the embodiments of this application; Figure 10 This is a scatter plot of the observed AGB and predicted AGB of the WCM model involved in the embodiments of this application; Figure 11 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application; Figure 12This is a flowchart illustrating the forest biomass prediction method based on differentiable physical modeling involved in the embodiments of this application.

[0024] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0026] First Embodiment This embodiment provides a forest biomass prediction method based on differentiable physical modeling, referring to... Figure 1 The diagram illustrates the architecture of the training model. Solid lines represent the forward propagation path of the model, used for AGB prediction; dashed lines represent the backpropagation path of the neural network during training, used for gradient calculation and parameter updates. This method is applied to training models that use a differentiable physical model as the network skeleton and incorporate a deep fully connected neural network.

[0027] In this embodiment, Differentiable Physical Modeling (DPM) refers to a physical model with automatic differentiation capabilities. Its function is to impose physical constraints on the training model, thereby enhancing its interpretability, describing the physical processes of canopy scattering and surface scattering after canopy attenuation, and generating predicted values ​​for forest biomass. In some optional implementations, a semi-empirical water cloud model is used as the differentiable physical model.

[0028] Deep Fully Connected Neural Network (DFCNN) is a purely data-driven deep learning network used to parameterize physical model parameters into neural networks.

[0029] It should be noted that traditional physical model methods possess physical interpretability, but are limited in accuracy and applicability due to difficulties in parameter acquisition and model simplification assumptions; while deep learning has strong feature learning and fitting capabilities, but suffers from poor stability and interpretability due to a lack of physical constraints. Therefore, this embodiment integrates physical models with deep learning to construct a forest AGB inversion method that combines the advantages of physical mechanism constraints and data-driven approaches.

[0030] Based on radiative transfer theory, a traditional physical semi-empirical water cloud model is selected as the physical framework. Under the Tensor operation framework of deep learning PyTorch, the semi-empirical water cloud model is seamlessly coupled with a deep fully connected neural network using automatic differentiation technology, realizing the differentiability of the semi-empirical water cloud model. The mean squared error is used as the loss function, and the weights and biases of the neural network are optimized by backpropagation chain rule algorithm and gradient descent. An end-to-end trainable DPM inversion model is constructed, which achieves high-precision inversion of forest ground biomass while taking into account physical interpretability and generalization ability.

[0031] Reference Figure 12 In this embodiment, the method includes the following steps: S10, Obtain the input parameter features from the collected dataset; In this embodiment, the collected datasets include various types of datasets, including but not limited to attribute values, synthetic aperture radar (SAR) satellite imagery, optical satellite imagery, the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM), plot survey data measured per tree, or plot survey data measured by angle gauge.

[0032] For example, in some specific implementations, a single-look complex (SLC) GF-3 PolSAR image and a GF-4 optical multispectral image are used as data sources. Tables 1 and 2 show the details of obtaining the GF-3 PolSAR and GF-4 data, which are sampled from 52 plots surveyed in the study area in December 2020, using the standard 20m × 20m plot survey method. Table 1. Specific parameter information of GF-3 PolSAR images

[0033] Table 2. Main parameters of the GF-4 satellite PMS sensor

[0034] S20, the input parameter features are preprocessed and then input into the deep fully connected neural network to predict the physical model parameters of the variables; In this embodiment, the input parameter features are preprocessed and then input into the physical model parameters of the variable generated by the deep fully connected neural network.

[0035] Furthermore, and optionally, in this embodiment, the preprocessing process includes tensor transformation and normalization to scale each feature to a fixed range in order to eliminate dimensional differences and ensure training stability.

[0036] Unlike traditional normalization methods, the normalization expression in this embodiment is:

[0037] In the formula, This represents the normalized characteristics of the input parameters; Represents the input feature tensor; This represents the minimum value of the input feature tensor X in the sample dimension; This represents the maximum value of the input feature tensor X in the sample dimension.

[0038] In some alternative implementations, the deep fully connected neural network comprises four linear layers: the first linear layer has 128 nodes and employs the LeakyReLU activation function (negative slope 0.0001) and Dropout (dropout rate 0.3); the second linear layer has 64 nodes and uses the LeakyReLU activation function (negative slope 0.0001); the third linear layer has 32 nodes and uses the LeakyReLU activation function (negative slope 0.0001); and the fourth linear layer has 3 nodes, which train and output intermediate variable physical model parameters as input to the differentiable physical module. During this process, the output layer nodes are mapped using a smooth nonlinear function mapping via tanh and Softplus to avoid non-differentiable points, thus ensuring the non-negativity and differentiability of the physical parameters.

[0039] S30, obtain the variable physics model parameters output by the deep fully connected neural network, and input the variable physics model parameters into the differentiable physics model through forward propagation; Further and optionally, the variable physics model parameters include a first variable physics model parameter, a second variable physics model parameter, and a third variable physics model parameter, each with the following expressions:

[0040]

[0041]

[0042] In the formula, The first variable is the physical model parameter. The second variable is the physical model parameter. The third variable is the physical model parameter. Represents the theoretical value of backscattering from a forest volume. Represents the theoretical value of surface backscattering. Represents the extinction coefficient. It represents the incident angle of scattering from the Earth's surface.

[0043] Furthermore, refer to Figure 2This embodiment is based on radiative transfer theory and uses the semi-empirical water cloud model WCM as its physical framework. The total backscattering coefficient of the forest is assumed to be... :

[0044]

[0045] In the formula, Representing forest biomass prediction results, This represents the polarization backscattering coefficient.

[0046] It should be noted that in traditional semi-empirical water cloud models, the three variable physical model parameters are all empirical values. However, in this embodiment, a deep learning model is used to predict the variable physical model parameters to overcome the unreliability of the selection of human empirical values ​​in traditional physical models.

[0047] In this embodiment, after obtaining the variable physical model parameters output by the deep fully connected neural network, the variable physical model parameters are further and optionally input into the computation graph of the semi-empirical water cloud model through forward propagation, wherein the computation graph is composed of continuously differentiable basic operational functions and includes automatic differentiation functions.

[0048] In some alternative implementations, the computation process of the computation graph is based on tensor operations and the computation graph is constructed using PyTorch automatic differentiation (AD) technology. The computation graph consists of continuously differentiable basic operation functions such as addition, subtraction, multiplication, division and logarithm.

[0049] S40, Obtain the forest biomass prediction results output by the differentiable physical model.

[0050] Finally, the forest biomass prediction results output by the differentiable physical model are obtained to complete the forest biomass prediction.

[0051] In the technical solution provided in this embodiment, the key parameters in the physical model are predicted by a deep fully connected neural network, which effectively alleviates the insufficient accuracy of the traditional physical model caused by parameter uncertainty and simplification assumptions. Then, the key parameters are input into the differentiable physical model to obtain the forest biomass prediction result. By introducing physical constraints, the physical consistency and interpretability of the inversion result are enhanced, so as to improve the generalization of the deep learning model when estimating forest biomass while taking into account accuracy.

[0052] Second Embodiment Based on the first embodiment, in order to ensure the prediction accuracy of the deep fully connected neural network for key parameters, this embodiment provides a training method for the deep fully connected neural network: during the training process, forward propagation is first performed and the loss function is calculated.

[0053] Subsequently, backpropagation is performed using automatic differentiation (AD). By calling the `loss.backward()` function, the AD framework applies the chain rule along the computation graph, calculating the gradient of the loss function with respect to the parameters (weights and biases) of each layer of the neural network layer by layer. The optimizer Adam updates the model parameters based on these gradients, achieving end-to-end optimization.

[0054] In the formula, the change in the loss function is given by... The gradient expansion consists of three chained gradient terms: the partial derivative of the loss function with respect to the estimated forest AGB, the partial derivative of AGB with respect to the key parameters of the physical model, and the partial derivative of the physical parameters with respect to the key parameters of the physical model. The partial derivatives of the gradient are used during the training phase. Following the chain rule algorithm, the gradient propagates backward from the output to the input layer by layer, sequentially from the loss function to the key parameters in the physical model, and then further propagates to the neural network, thus achieving [the desired effect]. Iterative updates.

[0055] The optimization process is as follows Figure 3 ,from Figure 3 As can be seen from the model structure, the mean gradient magnitude of each hidden layer parameter (weights and biases) changes during training. Overall, it shows a gradual decay trend with increasing training epochs, reflecting the gradual stabilization process of parameter updates. This process involves iterative forward propagation, loss calculation, and backpropagation until the loss function converges or reaches the preset number of iterations, thus completing the end-to-end differentiable physical modeling of the forest AGB and achieving forest AGB estimation.

[0056] This process involves iterative forward propagation, loss calculation, and back propagation until the loss function converges or reaches the preset number of iterations, thereby completing the end-to-end differentiable physical modeling of forest AGB and realizing forest AGB estimation.

[0057] Third Embodiment Based on any of the above embodiments, in this embodiment, in order to ensure the accuracy of the forest biomass prediction results output by the differentiable physical model, we need to compare the predicted AGB with the measured AGB after forward propagation, calculate the loss function of the two, and use the change value of the loss function as the quantification value of the accuracy of the prediction result.

[0058] Optionally, in this embodiment, the mean squared error (MSE) is selected as the loss function, and the calculation formula is as follows:

[0059] In the formula, The change in the loss function is represented by N, which represents the number of training data samples. Indicates the first Predicted biomass for the j-th polarization channel in a given sample. No. The measured biomass value of each sample.

[0060] The results are as follows Figure 4 As shown, from Figure 4 As can be seen from the data, the dynamic trajectory of the training and validation loss is as follows: the loss value decreases rapidly in the first 200 rounds and then tends to stabilize, indicating that the model is gradually fitting the data distribution and converging to a better solution, thus completing the end-to-end differentiable physical modeling of forest AGB and realizing forest AGB estimation.

[0061] Fourth embodiment Based on any of the above embodiments, this embodiment verifies the performance of a training model that uses a differentiable physical model as the network skeleton and integrates a deep fully connected neural network. After completing the DPM model training, the model is validated through random sampling. The sample data is divided into a training set and a validation set in a 6:4 ratio. Statistical indicators are used for quantitative evaluation, including the coefficient of determination (R²), bias, root mean square error (RMSE), and unbiased root mean square error (ubRMSE). The specific formulas are as follows:

[0062]

[0063]

[0064]

[0065] In the formula, This refers to the number of validation sample points; This refers to the measured value of biomass; This refers to the predicted biomass value. Represents the predicted values ​​of each biomass The average value.

[0066] Verification of Examples In this embodiment, to verify the superior performance of the forest biomass prediction method based on differentiable physical modeling (hereinafter referred to as DPM) proposed in this application in forest AGB inversion, a multi-model comparison experiment was designed. The comparison methods include: (1) Physical model comparison: The classic semi-empirical water cloud model WCM is adopted, and the physical parameters are estimated by nonlinear least squares method as the benchmark comparison model under the condition of no machine learning participation. (2) Comparison of deep learning models: Fully connected neural network (FNN) and generalized regression neural network (GRNN) are used. The former is a typical deep learning structure with nonlinear approximation ability, while the latter is a non-parametric method based on statistics and has good generalization performance under appropriate sample conditions. (3) Comparison of ensemble learning models: Random Forest (RF) and Adaptive Boosting (AdaBoost) ensemble learning models are used. The former completes the prediction by integrating multiple regression trees, while the latter improves the prediction performance by iteratively optimizing weak learners. The methods described above represent different paradigms, including physical models, deep learning, and classical machine learning. To ensure fairness in the comparison, each model was modeled using the same feature variables and training dataset as DPM, and necessary parameter adjustments and optimizations were made to suit their respective characteristics. Finally, performance comparisons were conducted under the same four evaluation metrics system.

[0067] Table 3. Statistical indicators for the random sampling validation of six AGB inversion models in the study area.

[0068] Table 3 shows the statistical indicators of the six AGB inversion models validated by random sampling in the study area. The table displays four indicators: coefficient of determination (R²), bias, root mean square error (RMSE), and unbiased root mean square error (ubRMSE). The table shows that the DPM framework constructed in this study performed excellently in forest AGB inversion in the study area. The DPM model achieved an R² of 0.60, significantly higher than FNN, GRNN, RF, AdaBoost, and the traditional WCM. Both RMSE and ubRMS were 24.23 t / hm², the lowest among all models, indicating a high degree of consistency between the predicted and measured values ​​with minimal dispersion, demonstrating high accuracy and stability. The bias of 0.4 t / hm², close to 0, indicates that DPM exhibits almost no systematic overestimation or underestimation, while FNN, GRNN, RF, and AdaBoost models showed significantly lower bias. The bias of st is close to or exceeds 10 t / hm², mainly due to the lack of physical constraints leading to overfitting to the noise of the training data; although WCM has a lower bias (-0.19 t / hm²), its R² is only 0.31 and its RMSE is 31.57 t / hm², indicating that its overly simplified model structure cannot fully reflect the spatial heterogeneity and environmental differences of the forest; as typical ensemble learning algorithms, RF and AdaBoost usually outperform GRNN and FNN in forest AGB inversion studies, while neural networks are often slightly inferior to ensemble learning models when faced with small-scale training data.

[0069] See Figures 5 to 10 Scatter plots of observed AGB and predicted AGB for DPM, GRNN, FNN, Adaboost, RF and WCM models are shown in sequence. Figures 5-10 The correlation between the predictions and experimental results of the six models is described. In the figure, the solid line represents the fitted linear regression line, and the dashed line represents the 1:1 ideal reference line. The results show that the scatter plot of the DPM model is closer to the 1:1 reference line, indicating that the predicted values ​​are relatively consistent with the experimental values, with the smallest deviation. The scatter plots of the RF and AdaBoost models deviate from the 1:1 reference line with an overall upward trend, indicating the existence of systematic bias. The scatter plot of the WCM model is relatively scattered, with even negative biases. The scatter plots of the FNN and GRNN models are relatively random, indicating insufficient fitting ability, with GRNN showing the most significant bias. The scatter plots intuitively reflect the prediction accuracy and systematic error of each model in forest AGB inversion, further verifying the superior performance of the proposed DPM method.

[0070] In summary, the DPM modeling method proposed in this invention outperforms the comparative models in terms of R², RMSE, Bias, and ubRMSE, demonstrating higher prediction accuracy and lower system bias. This invention organically integrates deep learning with physical models to construct an end-to-end differentiable physical modeling framework, significantly improving the accuracy and robustness of forest AGB inversion while ensuring physical consistency. It effectively addresses the limitations of existing methods, such as the difficulty in accurately obtaining physical model parameters and insufficient generalization ability of data-driven models.

[0071] As one implementation scheme, Figure 11 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.

[0072] like Figure 11 As shown, the computer system may include: a processor 1001, such as a CPU; a memory 1005; a user interface 1003; a network interface 1004; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0073] Those skilled in the art will understand that Figure 11 The computer system architecture shown does not constitute a limitation on the computer system and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0074] like Figure 11 As shown, the memory 1005, as a storage medium, may include an operating system, a network communication module, a user interface module, and computer programs. The operating system is a program that manages and controls the hardware and software resources of the computer system, as well as the operation of the computer programs and other software or programs.

[0075] exist Figure 11 In the computer system shown, the user interface 1003 is mainly used to connect to the terminal and communicate with the terminal; the network interface 1004 is mainly used to communicate with the backend server; and the processor 1001 can be used to call the computer program stored in the memory 1005.

[0076] In this embodiment, the computer system includes: a memory 1005, a processor 1001, and a computer program stored in the memory and executable on the processor, wherein: When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: Obtain the input parameter features from the collected dataset; The input parameter features are preprocessed and then input into the deep fully connected neural network to predict the physical model parameters of the variables. Obtain the variable physical model parameters and input the variable physical model parameters into the differentiable physical model through forward propagation; Obtain the forest biomass prediction results output by the differentiable physical model.

[0077] Furthermore, those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in a computer system to implement the process steps of the embodiments of the above methods.

[0078] Therefore, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the various steps of the forest biomass prediction method based on differentiable physical modeling as described in the above embodiments.

[0079] The computer-readable storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0080] It should be noted that, since the storage medium provided in the embodiments of this application is the storage medium used to implement the methods of the embodiments of this application, those skilled in the art can understand the specific structure and variations of the storage medium based on the methods described in the embodiments of this application, and therefore will not be repeated here. All storage media used in the methods of the embodiments of this application fall within the scope of protection of this application.

[0081] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0086] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A forest biomass prediction method based on differentiable physical modeling, characterized in that, The method, applied to a training model with a differentiable physical model as the network skeleton and incorporating a deep fully connected neural network, includes the following steps: Obtain the input parameter features from the collected dataset; The input parameter features are preprocessed and then input into the deep fully connected neural network to predict the physical model parameters of the variables. Obtain the variable physical model parameters and input the variable physical model parameters into the differentiable physical model through forward propagation; Obtain the forest biomass prediction results output by the differentiable physical model.

2. The forest biomass prediction method based on differentiable physical modeling as described in claim 1, characterized in that, The variable physical model parameters include a first variable physical model parameter, a second variable physical model parameter, and a third variable physical model parameter, each with its own expression: ； ； ； In the formula, The first variable is the physical model parameter. The second variable is the physical model parameter. The third variable is the physical model parameter. Represents the theoretical value of backscattering from a forest volume. Represents the theoretical value of surface backscattering. Represents the extinction coefficient. It represents the incident angle of scattering from the Earth's surface.

3. The forest biomass prediction method based on differentiable physical modeling as described in claim 1, characterized in that, Preprocessing the input parameter features specifically includes: After converting the input parameter features into tensor form, normalization is performed. The normalization expression includes: ； In the formula, This represents the normalized characteristics of the input parameters; Represents the input feature tensor; This represents the minimum value of the input feature tensor X in the sample dimension; This represents the maximum value of the input feature tensor X in the sample dimension.

4. The forest biomass prediction method based on differentiable physical modeling as described in claim 1, characterized in that, The input parameter features include parameter attribute values, synthetic aperture radar parameter features, optical parameter features, and global digital elevation model data and sample plot survey data from spaceborne thermal emission and reflectance radiometers.

5. The forest biomass prediction method based on differentiable physical modeling as described in claim 1, characterized in that, The differentiable physical model is a semi-empirical water cloud model. In calculating the forest biomass prediction results, the semi-empirical water cloud model includes: Obtain the backscattering coefficient, incident angle, and extinction coefficient of the polarization channel; The variable physical model parameters, polarization channel backscattering coefficient, incident angle, and extinction coefficient are input into the computation graph of the semi-empirical water cloud model through forward propagation. The computation graph consists of continuously differentiable basic operational functions and includes automatic differentiation functions.

6. The forest biomass prediction method based on differentiable physical modeling as described in claim 1, characterized in that, The training process of the deep fully connected neural network includes: Determine the mean squared error loss of the forest biomass prediction results; Calculate the gradient between the mean squared error loss and the parameters of each layer in the deep fully connected neural network; The deep fully connected neural network is trained by backpropagation based on the gradients of each layer.

7. The forest biomass prediction method based on differentiable physical modeling as described in claim 1, characterized in that, The expression for the forest biomass prediction result is as follows: ； In the formula, Represents the forest AGB prediction value; Represents the polarization backscattering coefficient. The first variable is the physical model parameter. The second variable is the physical model parameter. The third variable is the physical model parameter.

8. The application of a forest biomass prediction method based on differentiable physical modeling as described in any one of claims 1 to 7 in forest biomass measurement.

9. A computer system, characterized in that, The computer system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the forest biomass prediction method based on differentiable physical modeling as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the forest biomass prediction method based on differentiable physical modeling as described in any one of claims 1 to 7.

Images