Training methods and related devices based on physical mechanisms and rule constraints

Abstract

This application provides a training method and related apparatus based on physical mechanisms and rule constraints, relating to the field of machine learning. The electronic device acquires multi-source spatiotemporal remote sensing data and processes it through the model to be trained to obtain a leaf area index map to be evaluated. Subsequently, based on the map, three types of losses are calculated simultaneously: fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss. Finally, the weighted sum of the three is used as the comprehensive loss for updating model parameters, thereby improving the inversion accuracy while enhancing the scientific nature and interpretability of the results.

Description

Technical Field

[0001] This application relates to the field of machine learning, and more specifically, to a training method and related apparatus based on physical mechanisms and rule constraints. Background Technology

[0002] Leaf area index (LAI) is a key ecological parameter characterizing vegetation canopy structure and photosynthetic capacity. Its value directly reflects the regional vegetation growth status, carbon sequestration potential, and water transpiration intensity. Therefore, accurate acquisition of leaf area index plays an irreplaceable role in quantitatively assessing the impact of terrestrial ecosystems on global climate change.

[0003] However, current mainstream remote sensing methods generally rely on a single remote sensing data source when obtaining leaf area index (LAI), which has significant limitations. Specifically, while optical remote sensing images possess rich spectral information, they are susceptible to interference from cloud and rain, leading to large areas of data loss in the target area. Although Synthetic Aperture Radar (SAR) images have all-weather, all-time observation capabilities, they lack effective spectral resolution, making it difficult to distinguish vegetation types. More importantly, optical images and SAR images have inherent differences in spatiotemporal resolution, geometric distortion characteristics, and radiometric response mechanisms, making it difficult to achieve natural alignment at the pixel scale. Therefore, relying on either data source alone cannot stably, continuously, and reliably complete the LAI inversion task.

[0004] Therefore, in order to overcome the inherent defects of a single remote sensing data source, a technical approach of multi-source remote sensing data fusion has been proposed in related technologies. This approach involves simultaneously utilizing the advantages of optical image data and synthetic aperture radar image data to complement each other, and improving the spatiotemporal coverage and stability of leaf area index retrieval through joint modeling.

[0005] In practice, it has been found that when using multi-source remote sensing data to invert leaf area index, problems such as the complexity and heterogeneity of the ecosystem need to be addressed, resulting in poor inversion results. Summary of the Invention

[0006] In order to overcome at least one of the shortcomings of the prior art, one of the objectives of this application is to provide a training method and related apparatus based on physical mechanisms and rule constraints, which improves the inversion accuracy of the trained inversion model while enhancing the scientific nature and interpretability of the results.

[0007] Firstly, this application provides a training method based on physical mechanisms and rule constraints, the method comprising: Acquire multi-source spatiotemporal remote sensing data; The multi-source spatiotemporal remote sensing data is processed by the model to be trained to obtain the leaf area index map to be evaluated. Based on the leaf area index map, the fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss of the model to be trained are obtained. The fitting contrast loss represents the error between the leaf area index map and the corresponding label. The physical mechanism constraint loss represents the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance. The prior rule constraint loss represents the degree to which the leaf area index map violates the preset prior knowledge. The comprehensive loss is obtained based on the fitting comparison loss, the physical mechanism constraint loss, and the prior rule constraint loss, and the model to be trained is updated based on the comprehensive loss.

[0008] Secondly, this application provides a training device based on physical mechanisms and rule constraints, the device comprising: The sample collection module is used to acquire multi-source spatiotemporal remote sensing data; The forward inference module is used to process the multi-source spatiotemporal remote sensing data through the model to be trained to obtain the leaf area index map to be evaluated. The loss calculation module is used to obtain the fitting contrast loss, physical mechanism constraint loss and prior rule constraint loss of the model to be trained based on the leaf area index map. The fitting contrast loss represents the error between the leaf area index map and the corresponding label. The physical mechanism constraint loss represents the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance. The prior rule constraint loss represents the degree to which the leaf area index map violates the preset prior knowledge. The model update module is used to obtain a comprehensive loss based on the fitting comparison loss, physical mechanism constraint loss and prior rule constraint loss, and update the model to be trained based on the comprehensive loss.

[0009] Thirdly, this application provides a storage medium storing a computer program that, when executed by a processor, implements the training method based on physical mechanisms and rule constraints.

[0010] Fourthly, this application provides an electronic device, which includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the training method based on physical mechanisms and rule constraints.

[0011] Compared with the prior art, this application has the following beneficial effects: In the training method and related apparatus based on physical mechanisms and rule constraints provided in this embodiment, the electronic device acquires multi-source spatiotemporal remote sensing data and processes it through the model to be trained to obtain the leaf area index map to be evaluated. Then, based on the map, three types of losses are calculated simultaneously: fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss. Finally, the weighted sum of the three is used as the comprehensive loss for updating the model parameters, thereby improving the inversion accuracy while enhancing the scientific nature and interpretability of the results. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating the training method based on physical mechanisms and rule constraints provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the model to be trained provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a training device based on physical mechanisms and rule constraints provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application (hereinafter referred to as "the embodiments") clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0015] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0016] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0017] In the description of this application, it should be noted that the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0018] Furthermore, terms such as "horizontal," "vertical," and "sag" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0019] Based on the above statement, as introduced in the background technology, the current inversion of leaf area index based on multi-source remote sensing data faces problems such as the complexity and heterogeneity of the ecosystem, resulting in poor inversion results.

[0020] Specifically, in ecological ecotones, mountainous and hilly areas, and other regions with prominent spatial heterogeneity of ecosystems, traditional fusion methods still rely on the assumption of homogeneous pixels, which fails to characterize the nonlinear coupling relationship between vegetation, soil, and topography. This results in a serious distortion of the model's representation of the real physical processes on the Earth's surface and causes a significant scale effect, leading to a jump in the leaf area index inversion results in the boundary region.

[0021] Meanwhile, existing fusion methods largely employ purely data-driven deep learning models, which necessitates a high degree of reliance on large-scale, high-quality, and spatiotemporally balanced leaf area index ground truth labels for supervised training. However, in remote forest areas, alpine meadows, and other regions where on-site measurements are difficult and ground verification points are sparse, labeled samples are severely lacking, leading to a decline in the model's generalization ability. This not only results in a significant decrease in prediction accuracy but also produces results that clearly violate vegetation growth patterns and common sense in remote sensing physics, such as negative leaf area indices and high leaf area indices in aquatic areas.

[0022] It should be noted that the defects in the solutions in the prior art are the result of practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be regarded as contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0023] Based on the discovery of the above-mentioned technical problems, this embodiment provides a training method based on physical mechanisms and rule constraints. For example... Figure 1 As shown, the method includes: S1, acquire multi-source spatiotemporal remote sensing data.

[0024] S2, through the processing of multi-source spatiotemporal remote sensing data by the model to be trained, obtains the leaf area index map to be evaluated.

[0025] S3, based on the leaf area index diagram, obtain the fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss of the model to be trained.

[0026] Among them, the fitting contrast loss characterizes the error between the leaf area index map and the corresponding label, the physical mechanism constraint loss characterizes the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance, and the prior rule constraint loss characterizes the degree to which the leaf area index map violates the preset prior knowledge.

[0027] S4. The comprehensive loss is obtained based on the fitting comparison loss, the physical mechanism constraint loss, and the prior rule constraint loss, and the training model is updated based on the comprehensive loss.

[0028] Thus, this embodiment acquires multi-source spatiotemporal remote sensing data and processes it through the model to be trained to obtain the leaf area index map to be evaluated; then, based on this map, three types of losses are calculated simultaneously: fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss; finally, the weighted sum of the three is used as the comprehensive loss for updating model parameters, thereby improving the inversion accuracy while enhancing the scientific nature and interpretability of the results.

[0029] In this embodiment, the electronic device implementing the training method based on physical mechanisms and rule constraints can be, but is not limited to, a server, mobile terminal, tablet computer, desktop computer, etc., as long as it can provide sufficient computing power for the model. When serving as a service, the server can be a single server or a group of servers. The server group can be centralized or distributed (e.g., the server can be a distributed system). In some embodiments, the server can be local or remote relative to the user terminal. In some embodiments, the server can be implemented on a cloud platform; by way of example only, the cloud platform can include private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, inter-cloud, multi-cloud, etc., or any combination thereof. In some embodiments, the server can be implemented on an electronic device having one or more components.

[0030] To make the solution provided in this embodiment clearer, a server is used as the electronic device for implementing the method below, and in conjunction with... Figure 1 Each step of the method is described in detail. However, it should be understood that the operations in the flowchart may not be implemented in sequence, and steps without logical contextual relationships may be reversed in order or performed simultaneously. Furthermore, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowchart, or remove one or more operations from the flowchart. See also... Figure 1 The method includes: S1, acquire multi-source spatiotemporal remote sensing data.

[0031] In this embodiment, the server can automatically download and integrate multi-source spatiotemporal remote sensing data of the target area from the satellite data platform, including three categories: optical image data, synthetic aperture radar (SAR) image data, and true value data of ecosystem parameters.

[0032] The optical image data used is multispectral or hyperspectral optical image, such as Sentinel-2 L2A data, which includes 13 spectral bands including visible light, red edge, near infrared, and shortwave infrared.

[0033] SAR image data can be selected from synthetic aperture radar images that are contemporaneous with the aforementioned optical images and located in the same spatial region. For example, Sentinel-1 ground range detection product (Sentinel-1 GRD) data includes backscattering coefficients for both vertical-transmit vertical-receive (VV) and vertical-transmit horizontal-receive (VH) polarization modes.

[0034] The true values ​​of ecosystem parameters were obtained through ground measurements or high-precision global remote sensing products. Ground measurements were conducted using the LAI-2200 plant canopy analyzer, with 30 plots of 30×30m each set up within the protected area. The high-precision global remote sensing product was the MODIS LAI product MCD15A3H, which was resampled to 10-meter resolution and matched with optical and SAR images in both time and space.

[0035] In practical applications, the server further performs standardization processing on the acquired raw image data, including radiometric calibration, atmospheric correction, geometric registration, resampling, and sample construction.

[0036] Radiometric calibration uses calibration coefficients provided by the satellite to convert the original digital number (DN) values ​​into top atmospheric reflectance.

[0037] Atmospheric correction uses the 6S atmospheric radiative transfer model, inputting parameters such as imaging time, geographical location, aerosol type, and water vapor content to correct the reflectivity of the top atmospheric layer to the reflectivity of the Earth's surface.

[0038] Geometric registration uses optical images as a reference and employs a quadratic polynomial algorithm to spatially align SAR images, ensuring that the root mean square error does not exceed 0.25 pixels.

[0039] Resampling uniformly adjusted all images to a spatial resolution of 10 meters, and bilinear interpolation was used to preserve data continuity.

[0040] The sample construction was carried out by dividing the training set, validation set, and test set in a 7:2:1 ratio. Each sample contained optical data, SAR data, and corresponding ground truth labels within a 64×64 pixel spatiotemporal window across four time phases (16-day intervals), totaling 1200 samples: 840 in the training set, 240 in the validation set, and 120 in the test set. Abnormal samples with cloud coverage exceeding 20% ​​were excluded. The training set included the multi-source spatiotemporal remote sensing data used for model training.

[0041] Based on the above description of multi-source spatiotemporal remote sensing data in the embodiments, please refer to... Figure 1 Next, step S2 will be explained: S2, through the processing of multi-source spatiotemporal remote sensing data by the model to be trained, obtains the leaf area index map to be evaluated.

[0042] It should be understood that the multi-source spatiotemporal remote sensing data in this embodiment includes optical image data and synthetic aperture radar (SAR) image data. While optical image data contains spectral information, it is susceptible to cloud and rain obstruction, resulting in data loss. Although SAR image data can image in all weather conditions, it lacks a spectral response sensitive to vegetation biochemical parameters. These two types of data differ fundamentally in their physical characteristics, imaging mechanisms, and information dimensions. If they are directly mixed and input into a single network, the model will struggle to effectively distinguish and synergistically utilize their complementary advantages.

[0043] Therefore, as an optional implementation, the model to be trained includes a first feature extraction branch, a second feature extraction branch, and a parameter inversion layer. The server processes the optical image data through the first feature extraction branch to obtain a first feature to be fused; processes the synthetic aperture radar image through the second feature extraction branch to obtain a second feature to be fused; the first feature to be fused and the second feature to be fused are stitched together to obtain a fused feature; and the fused feature is processed through the parameter inversion layer to obtain the leaf area index map to be evaluated.

[0044] Specifically, the server processes the optical image data through a first feature extraction branch, which employs a three-dimensional convolutional structure with an input tensor dimension of: [Batch size, 4 phases, height, width, 13 spectral bands].

[0045] like Figure 2 The structure of the first feature extraction branch includes, in sequence, a first three-dimensional convolutional operation layer, a second three-dimensional convolutional operation layer, and a third and fourth three-dimensional convolutional operation layer.

[0046] In the first 3D convolutional operation layer, the convolutional kernel size is (3,3,3), the number of output channels is 64, the stride is (1,1,1), and the padding is (1,1,1). Then, 3D batch normalization and ReLU activation are connected. The first 3D max pooling operation has a pooling kernel size of (1,2,2) and only downsampling is performed on the spatial dimension.

[0047] In the second 3D convolutional operation layer, the convolutional kernel size is (3,3,3), the number of output channels is 128, the stride is (1,1,1), and the padding is (1,1,1). Then, 3D batch normalization and ReLU activation are connected. The second 3D max pooling operation has a pooling kernel size of (1,2,2).

[0048] In the third and fourth layers of 3D convolution operations, the structure is the same as that of the second layer of 3D convolution operations, with 256 and 512 output channels respectively. Finally, 3D global average pooling is performed to perform global average pooling on the spatiotemporal dimension, and outputs a 512-dimensional feature vector, which is the first feature to be fused.

[0049] See also Figure 2 The server simultaneously processes the synthetic aperture radar image data through a second feature extraction branch, which employs a two-dimensional convolutional structure with an input tensor dimension of: [Batch size, 4 phases, height, width, 2 polarization channels] See also Figure 2 The second feature extraction branch includes a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, and a third and fourth two-dimensional convolutional layer. The server processes the [H,W,2] data for each time phase separately.

[0050] In the first two-dimensional convolutional operation layer, the convolutional kernel size is (3,3) and the number of output channels is 32. Then, two-dimensional batch normalization and ReLU activation are connected. The first two-dimensional max pooling operation has a pooling kernel size of (2,2).

[0051] In the second two-dimensional convolutional operation layer, the convolutional kernel size is (3,3) and the number of output channels is 64. Then, two-dimensional batch normalization and ReLU activation are connected. The second two-dimensional max pooling operation has a pooling kernel size of (2,2).

[0052] In the third and fourth two-dimensional convolutional operation layers, the number of output channels is 128 and 256, respectively. Then, the feature maps of the four phases are concatenated in the channel dimension to form a feature map with 256×4=1024 channels. Then, the dimensionality is reduced to 512 channels by 1×1 two-dimensional convolution. Finally, a 512-dimensional feature vector is output by two-dimensional global average pooling. This feature vector is the second feature to be fused.

[0053] Before fusion, the server performs L2 normalization on both the first and second features to be fused and then concatenates them along the channel dimension to obtain a 1024-dimensional fused feature.

[0054] See also Figure 2 The server then processes the fused feature through a parameter inversion layer, which consists of three fully connected layers.

[0055] The first fully connected layer has an input dimension of 1024 and an output dimension of 512, followed by a ReLU activation function and a Dropout operation (dropout probability of 0.3). The second fully connected layer has an input dimension of 512 and an output dimension of 256, followed by a ReLU activation function. The third fully connected layer is the output layer, with an input dimension of 256 and an output dimension of H×W. A ReLU activation function is used to ensure that the output is non-negative, and then the output is reshaped into a shape of [batch size, 1, height, width], which yields the leaf area index map to be evaluated.

[0056] Based on the leaf area index diagram obtained from the above embodiments, please refer to... Figure 1 Next, we will... Figure 1 Step S3 will be explained below: S3, based on the leaf area index diagram, obtain the fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss of the model to be trained.

[0057] Among them, the fitting contrast loss characterizes the error between the leaf area index map and the corresponding label, the physical mechanism constraint loss characterizes the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance, and the prior rule constraint loss characterizes the degree to which the leaf area index map violates the preset prior knowledge.

[0058] The study found that traditional models treat the relationship between leaf area index and remote sensing reflectance as a purely statistical mapping, failing to embed the actual physical processes of vegetation-light interaction. However, the absorption, scattering, and transmission behavior of vegetation canopy to different wavelengths of light is jointly determined by factors such as leaf area, leaf tilt angle, chlorophyll content, water content, and observation geometry, exhibiting a clear intrinsic causal relationship. If the leaf area index output by the model cannot match the measured optical response, it indicates that while the result may fit the labeled value, it deviates from the physical essence. Therefore, this embodiment provides the following optional implementation method for step S3: S3-1: Input the leaf area index map into the vegetation radiative transfer model to simulate and generate canopy reflectance in at least two bands.

[0059] S3-2, calculate the error between the canopy reflectivity and the measured reflectivity of the corresponding band for each band, and obtain the physical mechanism constraint loss.

[0060] Specifically, the server inputs the leaf area index map to be evaluated into a differentiable vegetation radiative transfer model, which is based on the PROSAIL vegetation radiative transfer model, and its functional relationship is as follows:

[0061] In the formula, For canopy reflectivity, The leaf area index is represented by the leaf area index diagram. The leaf tilt angle distribution parameter is set to a spherical distribution (corresponding to a parameter value of 0.5), and the chlorophyll content is... Take 40 micrograms per square centimeter ( ), equivalent water thickness Take 0.015 grams per square centimeter ( ), dry matter content Take 0.009 grams per square centimeter ( ), Sun zenith angle Observing the zenith angle relative azimuth All data were obtained from remote sensing image metadata.

[0062] This embodiment uses the PyTorch or TensorFlow framework to replace all the non-differentiable core operations in the PROSAIL model, including exponential integrals and ray tracing processes, with differentiable tensor operations. This ensures that the gradient can be continuously propagated back from the aforementioned leaf area index map to the parameters of each layer of the model to be trained during backpropagation.

[0063] Furthermore, in this embodiment, the server further inputs the leaf area index map along with the aforementioned fixed parameters into the differentiable model to simulate and generate canopy reflectance in the red band (corresponding to Sentinel-2 B4 band) and near-infrared band (corresponding to Sentinel-2 B8 band). The server further employs smoothed L1 loss (Huber Loss, threshold loss). ) calculate separately Measured reflectivity of the corresponding band The error between them is used as a physical mechanism constraint loss.

[0064] Thus, this embodiment, without changing the structure of the model to be trained, enables the model learning process to be directly guided by the vegetation optical mechanism by embedding a differentiable real physical model.

[0065] In practice, it has also been found that when using deep learning to invert leaf area index, the model often produces predictions that defy common sense in areas lacking sufficient labeled data (e.g., mountainous areas, cloud-covered areas, or ecotones). For example, it may output negative values, extremely high values ​​far exceeding the physiological limits of vegetation (e.g., LAI > 10), drastic patchy jumps in the image, or give significant non-zero vegetation values ​​in known water areas.

[0066] Further research revealed that the purely data-driven model only learns the statistical correlation between pixels and labels. It neither understands the fundamental ecological threshold that the leaf area index must be non-negative and typically not exceed 10, nor does it possess prior knowledge of spatial continuity, and it cannot recognize basic geographical facts such as the absence of vegetation on water surfaces. When the training samples are unevenly distributed or contain noise, the model is prone to misclassifying abnormal patterns as valid patterns. Therefore, this embodiment also provides the following optional implementation methods for step S3: S3-3 generates numerical constraint loss, spatial smoothing loss, and logical consistency loss when obtaining the leaf area index map.

[0067] Among them, numerical constraint loss characterizes the degree of exceeding the limit of the leaf area index in the leaf area index map, spatial smoothing loss characterizes the degree of abrupt changes between adjacent pixels in the leaf area index map, and logical consistency loss characterizes the degree of violation of ecological common sense in the leaf area index map within a known water body area.

[0068] As an optional implementation, the server can obtain the excess amount of each pixel in the leaf area index map; average the excess amounts of all pixels to obtain the numerical constraint loss; and calculate the first-order finite difference between adjacent pixels along the horizontal and vertical directions based on the leaf area index map to obtain the horizontal gradient tensor and the vertical gradient tensor; use the mean of the horizontal gradient tensor and the vertical gradient tensor as the spatial smoothing loss; finally, determine the leaf area index located in the water body region in the leaf area index map; and use the mean of the leaf area index in the water body region as the logical consistency loss.

[0069] In practical applications, the server first performs numerical range constraint processing, defining the reasonable range for the leaf area index (LAI) as 0 to 10. The server calculates the excess amount for each pixel in the LAI image, assigning a positive value to any value exceeding this range: when the pixel value is greater than 10, the excess amount is the pixel value minus 10; when the pixel value is less than 0, the excess amount is the absolute value of the pixel value; and when the pixel value is between 0 and 10, the excess amount is 0. Subsequently, the server calculates the arithmetic mean of the excess amounts for all pixels to obtain the numerical constraint loss, the mathematical expression of which is:

[0070] In the formula, This represents a leaf area index graph.

[0071] The server then performs spatial smoothness constraint processing, that is, based on the leaf area index map, it calculates the first-order finite difference between adjacent pixels along the horizontal (i.e., image width direction) and vertical (i.e., image height direction) directions respectively, to obtain the horizontal gradient tensor. and longitudinal gradient tensor Then, the server takes the absolute values ​​of these two gradient tensors, sums them, and finally calculates the arithmetic mean of the sum as the spatial smoothing loss. Its mathematical expression is:

[0072] The server further performs logical consistency constraint processing, specifically by first calculating the water mask based on the Normalized Difference Water Index (NDWI). The NDWI index is calculated from specific bands of remote sensing imagery. When the NDWI is greater than 0.3, it is identified as a water body, and the corresponding location is assigned a value of 1 in the mask; other locations are assigned a value of 0. The server then uses this to determine the leaf area index (LAI) of areas located within water bodies in the leaf area index map, i.e., extracting all... Location Finally, the server uses the arithmetic mean of the leaf area indices within these water bodies as the logical consistency loss, the mathematical expression of which is:

[0073] Thus, this embodiment strictly adheres to prior ecological knowledge and imposes formal rationality constraints on the model output from three computable dimensions.

[0074] Based on the three types of losses obtained in the above embodiments, step S3 further includes: S3-4, weighted by the numerical constraint loss, spatial smoothing loss and logical consistency loss, yields the prior rule constraint loss.

[0075] During implementation, the server can multiply the numerical constraint loss, spatial smoothing loss, and logical consistency loss by their respective preset weight coefficients and then sum them to obtain the prior rule constraint loss. The weight coefficient for the numerical constraint loss is 1.0, the weight coefficient for the spatial smoothing loss is 0.5, and the weight coefficient for the logical consistency loss is 2.0. This set of weight coefficients is determined using a grid search method to guide the three types of constraints during model training. The mathematical expression for the weighted summation operation performed by the server is:

[0076] In the formula, , , , For numerical constraint loss, For spatial smoothing loss, This is the logical consistency loss. The prior rule constraint loss will be used as the total loss function. One of the components.

[0077] Based on the prior rule constraint loss mentioned above, the server also compares the leaf area index map with the corresponding label to obtain the fitting comparison loss.

[0078] During implementation, the server acquires the corresponding true values ​​of ecosystem parameters for each input sample as a label. This label is the true value of leaf area index (LAI) obtained through ground-based measurements or high-precision global remote sensing products. After spatiotemporal matching, it has the same spatial resolution and geographical coverage as the LAI map. The server compares the predicted value of each pixel in the LAI map with the true value in the corresponding location label one by one, calculating the pixel-by-pixel error between the two. This error is quantified using conventional regression loss functions such as Mean Squared Error (MSE) or Mean Absolute Error (MAE), and the resulting value is the fitting comparison loss. This loss characterizes the overall deviation between the LAI map and the corresponding label.

[0079] Based on the above explanation of the fitting comparison loss, physical mechanism constraint loss, and prior rule constraint loss, the following will discuss... Figure 1 Step S4 will be explained below: S4. The comprehensive loss is obtained based on the fitting comparison loss, the physical mechanism constraint loss, and the prior rule constraint loss, and the training model is updated based on the comprehensive loss.

[0080] During implementation, the server records the fitting comparison loss as... Let the physical mechanism constraint loss be denoted as Let the loss of prior rule constraints be denoted as The server performs a weighted sum of the three factors according to preset weighting coefficients to obtain the total loss function, the mathematical expression of which is:

[0081] In the formula, , , This set of weighting coefficients can balance the impact of the fitting contrast loss, the physical mechanism constraint loss, and the prior rule constraint loss on model training.

[0082] The server further minimizes this total loss function using a backpropagation algorithm. In each iteration, all learnable parameters of the first feature extraction branch, the second feature extraction branch, and the parameter inversion layer are simultaneously optimized.

[0083] During implementation, the provided training environment consisted of an Intel i9-12900K CPU, an NVIDIA RTX 3090 GPU, 64GB of RAM, Python 3.8, and PyTorch 1.12.0. Based on this training environment, the server could use the He initialization method to initialize the model parameters and then use the Adam optimizer to perform parameter updates, including the momentum decay parameter of this optimizer. The second moment is used to estimate the attenuation parameter. The numerical stability constant is 0.999. for The initial learning rate was 0.001, and a cosine annealing scheduler was used to smoothly decay the learning rate to 0 over a total of 100 training epochs. Furthermore, the server loaded data in batches of 32. When the loss value on the validation set decreased by no more than 0.001 over 10 consecutive training epochs, the model was considered converged and training was stopped. Finally, the server saved the model weights with the smallest root mean square error (RMSE) on the validation set. The saved model file contained all network weights, hyperparameter settings, training logs, and parameters required for preprocessing.

[0084] With this hardware and software configuration, experiments show that this embodiment achieves LAI inversion in mountainous and hilly areas and ecotones with frequent cloud cover and sparse ground verification points. Reaching 0.91 The value is 0.45, which is relatively low compared to purely data-driven deep learning baseline models. , ) and the traditional multi-source fusion method STARFM ( , All of them have seen substantial improvements.

[0085] Furthermore, due to the differentiable embedding of the physical model and the explicit modeling of the regular loss, the model output is no longer just a product of statistical correlation, but has a clear vegetation optical mechanism. For example, the trend of near-infrared reflectivity monotonically increasing with LAI and the enhanced red light absorption are accurately reproduced, which greatly enhances the interpretability and interdisciplinary credibility of the results.

[0086] Meanwhile, the dual-branch feature extraction structure is well adapted to the high-dimensional spectral characteristics of optical images and the polarization / scattering structure characteristics of SAR images, avoiding information distortion caused by forced alignment of heterogeneous data, and significantly improving the generalization stability under heterogeneous surface conditions while maintaining computational efficiency.

[0087] Therefore, the inversion model trained by the training method based on physical mechanisms and rule constraints provided in this embodiment can be used to support operations such as dynamic estimation of agricultural growth, quantitative assessment of interannual changes in forest carbon sinks, spatial mapping of regional drought stress, and continuous monitoring of vegetation cover in endangered habitats.

[0088] Based on the same inventive concept as the training method based on physical mechanisms and rule constraints provided in this embodiment, this embodiment also provides a training device based on physical mechanisms and rule constraints. This device includes at least one software functional module that can be stored in a memory or embedded in an electronic device. A processor in the electronic device executes the executable module stored in the memory. For example, the software functional modules and computer programs included in this device. Please refer to... Figure 3 Functionally, the device may include: Sample collection module 11 is used to acquire multi-source spatiotemporal remote sensing data; Forward inference module 12 is used to process multi-source spatiotemporal remote sensing data through the model to be trained to obtain the leaf area index map to be evaluated. The loss calculation module 13 is used to obtain the fitting comparison loss, physical mechanism constraint loss and prior rule constraint loss of the model to be trained based on the leaf area index map. The fitting comparison loss represents the error between the leaf area index map and the corresponding label, the physical mechanism constraint loss represents the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance, and the prior rule constraint loss represents the degree to which the leaf area index map violates the preset prior knowledge. The model update module 14 is used to obtain the comprehensive loss based on the fitting comparison loss, physical mechanism constraint loss and prior rule constraint loss, and update the model to be trained based on the comprehensive loss.

[0089] In this embodiment, the sample collection module 11 is used to implement Figure 1 In step S1, the forward inference module 12 is used to implement Figure 1 In step S2, the loss calculation module 13 is used to implement... Figure 1 In step S3, the model update module 14 is used to implement... Figure 1 Step S4 in the above process. Therefore, for a detailed description of each of the above modules, please refer to the specific implementation of the corresponding steps.

[0090] Optionally, the loss calculation module 13 obtains the physical mechanism constraint loss of the model to be trained based on the leaf area index plot, including: Input the leaf area index map into the vegetation radiative transfer model to simulate and generate canopy reflectance in at least two bands; The error between the canopy reflectivity and the measured reflectivity of the corresponding band is calculated separately to obtain the physical mechanism constraint loss.

[0091] Optionally, the vegetation radiative transfer model is obtained by differentiably reconstructing the PROSAIL vegetation radiative transfer model so that the vegetation radiative transfer model can undergo backpropagation along with the model to be trained.

[0092] Optionally, the loss calculation module 13 obtains the prior rule constraint loss of the model to be trained based on the leaf area index plot, including: The numerical constraint loss, spatial smoothing loss, and logical consistency loss generated by the leaf area index map are obtained. The numerical constraint loss characterizes the degree of exceedance of the leaf area index in the leaf area index map, the spatial smoothing loss characterizes the degree of jump between adjacent pixels in the leaf area index map, and the logical consistency loss characterizes the degree of violation of ecological common sense by the leaf area index map in the known water body area. We obtain the prior rule constraint loss by weighting the numerical constraint loss, spatial smoothing loss, and logical consistency loss.

[0093] Optionally, the loss calculation module 13 obtains the numerical constraint loss, spatial smoothing loss, and logical consistency loss generated by the leaf area index map in the following ways: Obtain the excess amount of each pixel in the leaf area index map; The numerical constraint loss is obtained by averaging the over-limit values ​​of all pixels. Based on the leaf area index map, the first-order finite difference between adjacent pixels is calculated along the horizontal and vertical axes to obtain the horizontal gradient tensor and the vertical gradient tensor. The mean of the horizontal gradient tensor and the vertical gradient tensor is used as the spatial smoothing loss; Determine the leaf area index located within the water body area in the leaf area index map; The mean leaf area index within the water body area is used as the logical consistency loss.

[0094] Optionally, the loss calculation module 13 obtains the fitting comparison loss of the model to be trained based on the leaf area index plot, including: The leaf area index map is compared with the corresponding label to obtain the fitting contrast loss.

[0095] Optionally, the multi-source spatiotemporal remote sensing data includes optical image data and synthetic aperture radar image data; The model to be trained includes a first feature extraction branch, a second feature extraction branch, and a parameter inversion layer; The forward inference module 12 processes multi-source spatiotemporal remote sensing data through the model to be trained to obtain the leaf area index map to be evaluated, including: The optical image data is processed through the first feature extraction branch to obtain the first feature to be fused; The synthetic aperture radar image is processed by the second feature extraction branch to obtain the second feature to be fused; The first feature to be fused and the second feature to be fused are concatenated to obtain the fused feature; The fusion features are processed by the parameter inversion layer to obtain the leaf area index map to be evaluated.

[0096] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0097] It should also be understood that if the above embodiments are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0098] Therefore, this embodiment also provides a storage medium, which is a computer-readable storage medium. This storage medium stores a computer program, which, when executed by a processor, implements the training method based on physical mechanisms and rule constraints provided in this embodiment. The storage medium can be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0099] This embodiment provides an electronic device for implementing a training method based on physical mechanisms and rule constraints. Figure 4 As shown, the electronic device may include a processor 22 and a memory 21. The memory 21 stores a computer program, and the processor reads and executes the computer program corresponding to the above-described embodiments in the memory 21 to implement the training method based on physical mechanisms and rule constraints provided in this embodiment.

[0100] See also Figure 4 The electronic device also includes a communication unit 23. The memory 21, processor 22 and communication unit 23 are electrically connected to each other directly or indirectly through system bus 24 to realize data transmission or interaction.

[0101] The memory 21 can be an information recording device based on any electronic, magnetic, optical, or other physical principles, used to record execution instructions, data, etc. In some embodiments, the memory 21 can be, but is not limited to, volatile memory, non-volatile memory, memory drive, etc.

[0102] In some embodiments, the volatile memory may be random access memory (RAM); in some embodiments, the non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, etc.; in some embodiments, the storage drive may be a disk drive, solid-state drive, any type of storage disk (such as optical disc, DVD, etc.), or similar storage media, or a combination thereof.

[0103] The communication unit 23 is used to send and receive data over a network. In some embodiments, the network may include a wired network, a wireless network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network, or a near field communication (NFC) network, or any combination thereof. In some embodiments, the network may include one or more network access points. For example, the network may include wired or wireless network access points, such as base stations and / or network switching nodes, through which one or more components of the service request processing system can connect to the network to exchange data and / or information.

[0104] The processor 22 may be an integrated circuit chip with signal processing capabilities, and may include one or more processing cores (e.g., a single-core processor or a multi-core processor). By way of example only, the processor described above may include a Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), an Application Specific Instruction-set Processor (ASIP), a Graphics Processing Unit (GPU), a Physics Processing Unit (PPU), a Digital Signal Processor (DSP), a Field Programmable Gate Array (FPGA), a Programmable Logic Device (PLD), a controller, a microcontroller unit, a Reduced Instruction Set Computing (RISC) computer, or a microprocessor, or any combination thereof.

[0105] Understandable. Figure 4The structure shown is for illustrative purposes only. Electronic devices may also have more advanced features. Figure 4 Showing more or fewer components, or having with Figure 4 The different configurations shown. Figure 4 The components shown can be implemented using hardware, software, or a combination thereof.

[0106] It should be understood that the apparatus and methods disclosed in the above embodiments can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0107] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A training method based on physical mechanisms and rule constraints, characterized in that, The method includes: Acquire multi-source spatiotemporal remote sensing data; The multi-source spatiotemporal remote sensing data is processed by the model to be trained to obtain the leaf area index map to be evaluated. Based on the leaf area index map, the fitting contrast loss, physical mechanism constraint loss, and prior rule constraint loss of the model to be trained are obtained. The fitting contrast loss represents the error between the leaf area index map and the corresponding label. The physical mechanism constraint loss represents the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance. The prior rule constraint loss represents the degree to which the leaf area index map violates the preset prior knowledge. The comprehensive loss is obtained based on the fitting comparison loss, the physical mechanism constraint loss, and the prior rule constraint loss, and the model to be trained is updated based on the comprehensive loss.

2. The training method based on physical mechanisms and rule constraints according to claim 1, characterized in that, Based on the leaf area index plot, the physical mechanism constraint loss of the model to be trained is obtained, including: Input the leaf area index map into the vegetation radiative transfer model to simulate and generate canopy reflectance in at least two bands; The error between the canopy reflectivity and the measured reflectivity of the corresponding band is calculated for each band to obtain the physical mechanism constraint loss.

3. The training method based on physical mechanisms and rule constraints according to claim 2, characterized in that, The vegetation radiative transfer model is obtained by differentiably reconstructing the PROSAIL vegetation radiative transfer model, so that the vegetation radiative transfer model can undergo backgrad propagation together with the model to be trained.

4. The training method based on physical mechanisms and rule constraints according to claim 1, characterized in that, Based on the leaf area index plot, the prior rule constraint loss of the model to be trained is obtained, including: The numerical constraint loss, spatial smoothing loss, and logical consistency loss generated by the leaf area index map are obtained. The numerical constraint loss represents the degree of exceeding the limit of the leaf area index in the leaf area index map, the spatial smoothing loss represents the degree of abrupt change between adjacent pixels in the leaf area index map, and the logical consistency loss represents the degree of violation of ecological common sense by the leaf area index map in a known water body area. The prior rule constraint loss is obtained by weighting the numerical constraint loss, the spatial smoothing loss, and the logical consistency loss.

5. The training method based on physical mechanisms and rule constraints according to claim 4, characterized in that, The numerical constraint loss, spatial smoothing loss, and logical consistency loss generated by the leaf area index plot are obtained, including: Obtain the excess amount of each pixel in the leaf area index map; The numerical constraint loss is obtained by averaging the over-limit values ​​of all pixels. Based on the leaf area index map, the first-order finite difference between adjacent pixels is calculated along the horizontal and vertical directions to obtain the horizontal gradient tensor and the vertical gradient tensor. The mean of the horizontal gradient tensor and the vertical gradient tensor is used as the spatial smoothing loss; Determine the leaf area index located within the water body area in the leaf area index graph; The mean leaf area index within the water body region is used as the logical consistency loss.

6. The training method based on physical mechanisms and rule constraints according to claim 1, characterized in that, Based on the leaf area index plot, the fitting contrast loss of the model to be trained is obtained, including: The leaf area index map is compared with the corresponding label to obtain the fitting contrast loss.

7. The training method based on physical mechanisms and rule constraints according to claim 1, characterized in that, The multi-source spatiotemporal remote sensing data includes optical image data and synthetic aperture radar image data. The model to be trained includes a first feature extraction branch, a second feature extraction branch, and a parameter inversion layer; The multi-source spatiotemporal remote sensing data is processed by the model to be trained to obtain the leaf area index map to be evaluated, including: The optical image data is processed through the first feature extraction branch to obtain the first feature to be fused; The synthetic aperture radar image is processed by the second feature extraction branch to obtain the second feature to be fused; The first feature to be fused and the second feature to be fused are concatenated to obtain the fused feature; The fused features are processed by the parameter inversion layer to obtain the leaf area index map to be evaluated.

8. A training device based on physical mechanisms and rule constraints, characterized in that, The device includes: The sample collection module is used to acquire multi-source spatiotemporal remote sensing data; The forward inference module is used to process the multi-source spatiotemporal remote sensing data through the model to be trained to obtain the leaf area index map to be evaluated. The loss calculation module is used to obtain the fitting contrast loss, physical mechanism constraint loss and prior rule constraint loss of the model to be trained based on the leaf area index map. The fitting contrast loss represents the error between the leaf area index map and the corresponding label. The physical mechanism constraint loss represents the error between the remote sensing reflectance obtained based on the leaf area index map and the measured reflectance. The prior rule constraint loss represents the degree to which the leaf area index map violates the preset prior knowledge. The model update module is used to obtain a comprehensive loss based on the fitting comparison loss, physical mechanism constraint loss and prior rule constraint loss, and update the model to be trained based on the comprehensive loss.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the training method based on physical mechanisms and rule constraints as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the training method based on physical mechanisms and rule constraints as described in any one of claims 1-7.

Images