Satellite solar shortwave radiation inversion method based on knowledge distillation

Abstract

The present invention provides a method for inverting satellite solar shortwave radiation based on knowledge distillation, which relates to the field of solar shortwave radiation inversion, including: collecting polar-orbiting satellite data, geostationary satellite data, geographic information data, and ground observation data; preprocessing the data to obtain input features and labels for training teacher models and student models; using a random forest model to rank the importance of input features, select key features, and eliminate non-key features; using the key features to construct a deep learning model to generate a teacher model for solar shortwave radiation inversion; using the output of the teacher model as a soft label, introducing ground observation data as a hard label for knowledge distillation, and obtaining a student model for solar shortwave radiation inversion. The technical solution of the present invention overcomes the problems of data scarcity and high data acquisition costs faced by the prior art when building models based solely on ground observation data, while the inversion accuracy based entirely on satellite data is limited.

Description

Technical Field

[0001] The present invention relates to the technical field of solar shortwave radiation inversion, and in particular to a satellite solar shortwave radiation inversion method based on knowledge distillation. Background Art

[0002] Against the backdrop of global climate change and environmental protection, accurately monitoring and assessing solar shortwave radiation is crucial for Earth's energy balance and ecosystems. Solar shortwave radiation is the primary source of surface energy, influencing key processes such as atmospheric circulation, temperature distribution, and the water cycle. Accurately inverting and predicting the distribution and changes of solar shortwave radiation has significant scientific and applied value in fields such as climate research, agricultural production, solar energy utilization, and environmental monitoring.

[0003] The development of satellite observation technology has made it possible to observe solar shortwave radiation at high temporal and spatial resolution worldwide. However, the accuracy of inversions of solar shortwave radiation is often limited by interference from factors such as atmospheric attenuation, cloud cover, and surface reflectivity. Furthermore, satellite data are indirect and rely heavily on inversion algorithms to extract true ground radiation information. Therefore, the accuracy of these algorithms directly determines the reliability of the inversion results.

[0004] Deep learning methods have made breakthrough progress in the field of remote sensing inversion in recent years. By learning features from a large amount of historical data, the solar shortwave radiation inversion algorithm based on deep learning can overcome the limitations of traditional physical methods to a certain extent and improve the accuracy of solar shortwave radiation inversion. The current mainstream method for deep learning inversion of solar shortwave radiation is to directly use ground observation data as training labels. However, the number of ground observation sites is limited, which makes it difficult to meet the requirements of high temporal and spatial resolution over a large area, and it is impossible to provide real-time radiation information under rapidly changing weather conditions. Chinese patent: CN117371316A discloses a geostationary satellite solar shortwave radiation inversion method and readable storage medium based on deep learning. This method uses polar-orbiting satellite solar shortwave radiation data with higher accuracy and spatial resolution as training labels to construct a model, combining the advantages of geostationary satellite data and polar-orbiting satellite data, and can perform high-frequency monitoring of solar radiation to meet the needs of dynamic change monitoring. However, due to certain differences in the data distribution of satellite observations and ground observations, the solar shortwave radiation results inverted using only satellite data may not fully reflect the actual situation of ground radiation.

[0005] In summary, building models solely based on ground observation data will face problems of data scarcity and high data acquisition costs, while the inversion accuracy based entirely on satellite data is limited.

[0006] Therefore, there is a need for a satellite solar shortwave radiation inversion method based on knowledge distillation that can organically combine multi-source satellite and ground observation data to improve inversion accuracy while taking into account wide-area coverage and high-frequency dynamic monitoring. Summary of the Invention

[0007] The main purpose of the present invention is to provide a satellite solar shortwave radiation inversion method based on knowledge distillation to solve the problems in the existing technology of data scarcity and high data acquisition cost when building models based solely on ground observation data, while the inversion accuracy based entirely on satellite data is limited.

[0008] To achieve the above objectives, the present invention provides a satellite solar shortwave radiation inversion method based on knowledge distillation, which specifically includes the following steps:

[0009] S1, collects polar-orbiting satellite data, geostationary satellite data, geographic information data and ground observation data.

[0010] S2 preprocesses geostationary satellite data, polar-orbiting satellite data, ground observation data, and latitude and longitude data and surface elevation data in geographic information data to obtain input features and labels for teacher model and student model training.

[0011] S3, uses the random forest model to rank the importance of input features, selects the key features of satellite solar shortwave radiation inversion, and eliminates non-key features.

[0012] S4, uses key features to build a deep learning model to generate a teacher model for solar shortwave radiation inversion.

[0013] S5, the output results of the teacher model are used as soft labels, and the ground observation data are introduced as hard labels for knowledge distillation to obtain the student model of solar shortwave radiation inversion.

[0014] Furthermore, the geostationary satellite data includes: spectral channel data and geometric feature data; the geometric feature data includes: satellite zenith angle, satellite azimuth angle, solar zenith angle and solar azimuth angle.

[0015] Geographic information data includes: latitude and longitude data, surface elevation data and surface type data.

[0016] Polar-orbiting satellite data include: solar shortwave radiation data.

[0017] Ground observation data include: ground solar shortwave radiation observation data.

[0018] Furthermore, step S2 specifically includes the following steps:

[0019] S2.1. Standardize the spectral channel data from geostationary satellite data, the solar shortwave radiation data from polar-orbiting satellite data, and the latitude and longitude data and surface elevation data from geographic information data:

[0020] (1);

[0021] in, is the normalized eigenvalue, is the original data, is the mean of the original data, is the standard deviation of the original data.

[0022] S2.2, normalize the data after standardization in S2.1:

[0023] (2);

[0024] in, is the normalized eigenvalue, and are the minimum and maximum values ​​of the original data respectively.

[0025] S2.3, perform sine and cosine transforms on the geometric characteristic data of the geostationary satellite.

[0026] S2.4, perform one-hot encoding on the surface type data in the geographic information data and convert the surface type data into a numerical format:

[0027] (3);

[0028] in, is the feature vector after one-hot encoding, is the number of surface types, Indicates that if the surface type characteristics belong to Class is , otherwise 0, where .

[0029] Furthermore, step S3 specifically includes the following steps:

[0030] S3.1. For each decision tree, the input features include preprocessed geostationary satellite data and geographic information data. When an input feature is split at a node, the reduction in the contribution of the node to the prediction error is calculated:

[0031] (4);

[0032] in, is the error reduction when the node is split, 、 、 are the errors of the parent node, left node and right node respectively, and They are respectively the proportion of the number of left and right node samples in the total number of parent node samples.

[0033] S3.2, each feature will have multiple split nodes in each tree, and the sum of all nodes The sum is then averaged across the forest to get the overall importance of the feature:

[0034] (5);

[0035] in, is the importance of the feature, is the number of trees in the random forest, Indicates the A tree, Represents all nodes in the tree.

[0036] Furthermore, step S4 specifically includes the following steps:

[0037] S4.1, the neural network includes: an input layer, a fully connected layer and an output layer connected in sequence, and the key features are input into the input layer of the neural network.

[0038] S4.2 is then input to the fully connected layer, and nonlinearity is introduced through the activation function. The calculation formula of the first layer of the fully connected layer is:

[0039] (6);

[0040] in, is the output latitude of the first layer, Is the activation function for nonlinear mapping, is the weight matrix of the first layer, is the bias vector of the first layer.

[0041] S4.3, the fully connected layer The calculation formula of the layer is:

[0042] (7);

[0043] in, It is The output of the layer, It is The weight matrix of the layer, is the bias term, It is The output dimension of the layer.

[0044] S4.4 is finally output through the output layer. The calculation formula of the output layer is:

[0045] (8);

[0046] in, is the predicted value of the model, which represents the solar shortwave radiation data at the corresponding moment. is the weight matrix of the output layer, is the output dimension of the penultimate layer, is the bias term of the output layer.

[0047] S4.5, set the loss function, the calculation formula is as follows:

[0048] (9);

[0049] in, is the loss function of the pre-trained model, It is The predicted value of the sample, It is The true value of the sample, is the number of samples.

[0050] Furthermore, step S5 specifically includes the following steps:

[0051] S5.1. Use bilinear spatial interpolation to find the geostationary satellite data and geographic information data that are consistent with the ground observation station position. Estimate the value of the target point using four adjacent points. First, perform horizontal interpolation:

[0052] (10);

[0053] (11);

[0054] Then perform longitudinal interpolation:

[0055] (12);

[0056] in, , , , are the values ​​of four known adjacent points, , , , are the coordinates of the known data points, are the estimated target point coordinates.

[0057] S5.2, polar-orbiting satellite data is used to train the teacher model, and the solar shortwave radiation data predicted and output by the teacher model is used as soft labels. The ground observation data is used as hard labels to train the student model. The soft label and hard label training are optimized using soft target loss and hard target loss respectively. The student model prediction results are denormalized to obtain the optimized results of solar shortwave radiation inversion.

[0058] The hard target loss is used to measure the difference between the student model output and the hard label. The loss value is calculated using the mean square error, and the expression is:

[0059] (13);

[0060] in, refers to hard target loss, is the total number of samples in the ground observation data, It is The real solar shortwave radiation data of samples, The student model is The predicted value of solar shortwave radiation on samples.

[0061] The soft target loss is used to measure the difference between the output of the student model and the teacher model. The loss value is also calculated using the mean squared error:

[0062] (14);

[0063] in, is the soft objective function, is the number of training samples used in the knowledge distillation process, For the teacher model The predicted value on samples, that is, the soft label or the output of the teacher model, For the student model The predicted value on the sample is the value obtained by the student model in the process of imitating the output of the teacher model.

[0064] Combining the soft target loss and the hard target loss defines the total loss function of distillation, which is expressed as follows:

[0065] (15);

[0066] in, is the total loss function during the distillation process, is a hyperparameter.

[0067] Furthermore, step S5 further includes the following steps:

[0068] S5.3, set the initial learning rate to 10 -5And gradually reduce the learning rate, using the learning rate decay strategy of exponential decay, the expression is as follows:

[0069] (16);

[0070] in, is the current learning rate, is the initial learning rate, is the decay rate, is the number of cycles of current training, is the attenuation step length.

[0071] Furthermore, the generation of the student model includes the following steps:

[0072] The geostationary satellite data and geographic information data obtained by S5.1 that are consistent with the position of the ground observation station are used as input data, and the ground-observed solar shortwave radiation data are used as labels.

[0073] Perform data cleaning, feature transformation, and feature selection on the input data.

[0074] This is then input into the neural network in step S4.1 to generate a student model.

[0075] The present invention has the following beneficial effects:

[0076] High spatial coverage and regional adaptability: Satellite observation data covers a wide range and contains more spatial features. After further training through knowledge distillation combined with ground observation data, the student model can better adapt to the characteristics of different regions, especially for areas with scarce observation stations, and can provide high-precision solar shortwave radiation estimates.

[0077] High spatiotemporal resolution: Using a teacher model based on multi-source satellite observation data, it is possible to obtain solar shortwave radiation data with high spatiotemporal resolution. After fine-tuning with ground observation data, the student model can maintain high resolution while enhancing its response to the actual ground radiation conditions, thereby more accurately reflecting the characteristics of instantaneous radiation changes and supporting high-frequency and dynamic monitoring needs.

[0078] Improved Accuracy and Robustness: Knowledge distillation can fully leverage the high precision of ground-based observation data, improving the model's inversion accuracy in specific regions and meteorological conditions. Furthermore, by fine-tuning the model with ground-based data, the robustness and generalization capabilities of the student model are enhanced, ensuring that it maintains good performance across different seasons and climate conditions.

[0079] Reducing reliance on ground-based observations while improving model accuracy: Unlike relying solely on ground-based observations for model training, fine-tuning using satellite data combined with some ground-based data can achieve higher inversion accuracy with less ground-based observations. This approach can, to a certain extent, alleviate the limitations of insufficient ground-based observations on model training, reducing data collection costs while fully leveraging the potential of satellite observations. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0081] Figure 1 A flow chart of a satellite solar shortwave radiation inversion method based on knowledge distillation of the present invention is shown.

[0082] Figure 2 A comparison chart of the effects of the knowledge distillation optimization model is shown. DETAILED DESCRIPTION

[0083] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0084] like Figure 1 The satellite solar shortwave radiation inversion method based on knowledge distillation shown in FIG1 includes the following steps:

[0085] S1, collects polar-orbiting satellite data, geostationary satellite data, geographic information data and ground observation data.

[0086] S2 preprocesses geostationary satellite data, polar-orbiting satellite data, ground observation data, and latitude and longitude data and surface elevation data in geographic information data to obtain input features and labels for teacher model and student model training.

[0087] S3, uses the random forest model to rank the importance of input features, selects the key features of satellite solar shortwave radiation inversion, and eliminates non-key features.

[0088] S4, uses key features to build a deep learning model to generate a teacher model for solar shortwave radiation inversion.

[0089] S5, the output results of the teacher model are used as soft labels, and the ground observation data are introduced as hard labels for knowledge distillation to obtain the student model of solar shortwave radiation inversion.

[0090] The present invention first uses multi-source satellite data to construct a teacher model for solar shortwave radiation inversion based on deep learning technology to extract solar shortwave radiation information with high temporal and spatial resolution; based on the teacher model, ground solar radiation data is introduced for knowledge distillation to optimize the parameters of the student model, so that the inversion results of the student model are closer to the actual situation on the ground and the inversion accuracy is improved.

[0091] Specifically, geostationary satellite data includes spectral channel data and geometric feature data; the latter includes satellite zenith angle, satellite azimuth angle, solar zenith angle, and solar azimuth angle. Geostationary satellite data and polar-orbiting satellite data with the same observation time are selected. Due to the different observation frequencies of the two types of satellites, the closest time range (e.g., ±3 minutes) is selected. The time-matched geostationary satellite data are clipped to the observation range of the polar-orbiting satellite to ensure spatial alignment.

[0092] Geographic information data includes: longitude and latitude data, surface elevation data, and surface type data. Ensure that all information is consistent with the temporal and spatial scope of satellite data.

[0093] Polar-orbiting satellite data, including solar shortwave radiation data, serves as labeled data for teacher model training. Only regions scanned simultaneously with geostationary satellites are extracted to ensure temporal and spatial alignment.

[0094] Ground observation data include: ground solar shortwave radiation observation data, which is used as high-precision label data for the subsequent knowledge distillation stage, i.e., the label for student model training.

[0095] Geostationary satellite data and geographic information data are used as input features for subsequent teacher model and student model training, and polar orbit satellite data and ground observation data are used as labels.

[0096] Specifically, all data are cleaned, outliers are removed, and missing values ​​are filled.

[0097] Step S2 specifically includes the following steps:

[0098] S2.1. Standardize the spectral channel data from geostationary satellite data, the solar shortwave radiation data from polar-orbiting satellite data, and the latitude and longitude data and surface elevation data from geographic information data:

[0099] (1);

[0100] in, is the normalized eigenvalue, is the original data, is the mean of the original data, is the standard deviation of the original data.

[0101] S2.2, normalize the data after standardization in S2.1:

[0102] (2);

[0103] in, is the normalized eigenvalue, and are the minimum and maximum values ​​of the original data respectively.

[0104] S2.3, perform sine and cosine transforms on the geometric characteristic data of the geostationary satellite.

[0105] S2.4 processes geometric angles such as the satellite zenith angle through sine and cosine transforms to preserve their periodic characteristics. The surface type is converted to a numerical format through one-hot encoding to improve the model's adaptability to different surface types.

[0106] Perform one-hot encoding on the surface type data in the geographic information data and convert the surface type data into numerical format:

[0107] (3);

[0108] in, is the feature vector after one-hot encoding, is the number of surface types, Indicates that if the surface type characteristics belong to Class is , otherwise 0, where .

[0109] Specifically, step S3 includes the following steps:

[0110] S3.1. For each decision tree, the input features include preprocessed geostationary satellite data and geographic information data. When an input feature is split at a node, the reduction in the contribution of the node to the prediction error is calculated:

[0111] (4);

[0112] in, is the error reduction when the node is split, 、 、 are the errors of the parent node, left node and right node respectively, and They are respectively the proportion of the number of left and right node samples in the total number of parent node samples.

[0113] S3.2, each feature will have multiple split nodes in each tree, and the sum of all nodes The sum is then averaged across the forest to get the overall importance of the feature:

[0114] (5);

[0115] in, is the importance of the feature, is the number of trees in the random forest, Indicates the A tree, Represents all nodes in the tree. The importance of features is sorted from high to low based on their values, and the contribution of each feature to the target variable is determined. Features with small contributions are removed. For example, if the first n features are key features, the features after the nth feature are removed.

[0116] Specifically, each data point in the training data is ensured to correspond to a spatiotemporal point scanned simultaneously by a geostationary satellite and a polar-orbiting satellite, using key features as input and the level-3 shortwave radiation data provided by the polar-orbiting satellite as labels.

[0117] The purpose of this invention is to use the same pixel points, that is, the various spectral channels and geographic information of geostationary satellites at the same longitude and latitude, to invert the solar shortwave radiation value at a certain moment. The spatial relationship and time series information between adjacent pixels do not need to be explicitly considered. Therefore, a multi-layer perceptron is used as the model architecture.

[0118] Step S4 specifically includes the following steps:

[0119] S4.1, the neural network includes: an input layer, a fully connected layer and an output layer connected in sequence, and the key features are input into the input layer of the neural network.

[0120] S4.2 is then input to the fully connected layer, and nonlinearity is introduced through the activation function. The calculation formula of the first layer of the fully connected layer is:

[0121] (6);

[0122] in, is the output latitude of the first layer, Is the activation function for nonlinear mapping, is the weight matrix of the first layer, is the bias vector of the first layer.

[0123] The fully connected layer maps the input features to a new space through linear transformation and introduces nonlinearity through the activation function, aiming to introduce higher-level feature expression capabilities to the teacher model while enabling the teacher model to capture complex input data patterns.

[0124] S4.3, each subsequent layer, i.e., the hidden layer, extracts and combines features through similar fully connected calculations. Each layer performs nonlinear transformations through linear transformations (weight matrices and bias terms) and activation functions, gradually increasing the complexity of the model to better capture the underlying patterns in the data. The calculation formula of the layer is:

[0125] (7);

[0126] in, It is The output of the layer, It is The weight matrix of the layer, is the bias term, It is The output dimension of the layer.

[0127] S4.4 is finally output through the output layer. The calculation formula of the output layer is:

[0128] (8);

[0129] in, is the predicted value of the model, which represents the solar shortwave radiation data at the corresponding moment. is the weight matrix of the output layer, is the output dimension of the penultimate layer, is the bias term of the output layer.

[0130] The output layer generates the predicted value of the teacher model through linear transformation. For each input data point, the output layer calculates a predicted value, which represents the solar shortwave radiation value at the corresponding moment.

[0131] S4.5, the loss function is used to measure the gap between the model prediction results and the hard label. The solar shortwave radiation inversion in this invention is actually a regression problem, so the loss function is the mean square error. Set the loss function and calculate it as follows:

[0132] (9);

[0133] in, is the loss function of the pre-trained model, It is The predicted value of the sample, It is The true value of the sample, is the number of samples. The mean squared error calculates the average of the squares of the differences between the predicted value and the true value. The goal is to minimize the loss function so that the predicted value is as close to the true value as possible.

[0134] Specifically, step S5 includes the following steps:

[0135] S5.1. Geostationary satellite data typically covers a wide area, while ground observation stations are relatively sparse, and the longitude and latitude of the two may not be completely consistent. Therefore, to ensure that each ground observation point has corresponding satellite data, a bilinear spatial interpolation method is used to obtain the geostationary satellite data and geographic information data that are consistent with the ground observation station location. The value of the target point is estimated using four adjacent points. First, horizontal interpolation is performed:

[0136] (10);

[0137] (11);

[0138] Then perform longitudinal interpolation:

[0139] (12);

[0140] in, , , , are the values ​​of four known adjacent points, , , , are the coordinates of the known data points, are the estimated target point coordinates.

[0141] S5.2, knowledge distillation enables a complex large model (the teacher model) to guide the training of a simplified small model (the student model) through the probability distribution of its outputs (soft labels). In this paper, the teacher model is a solar shortwave radiation inversion model pre-trained on large-scale geostationary and polar-orbiting satellite data, based on previous work. The teacher model learns the complex features of this large amount of geostationary and polar-orbiting satellite data and is applicable to solar radiation inversion tasks under a wide range of spatiotemporal conditions. The student model learns the output distribution of the teacher model and the hard labels of ground observation data through a distillation loss function.

[0142] Polar-orbiting satellite data is used to train the teacher model. The solar shortwave radiation data predicted by the teacher model is used as soft labels, and the ground observation data is used as hard labels to train the student model. Soft target loss and hard target loss are used for optimization during soft label training, respectively. The prediction results of the student model are denormalized to obtain the optimized results of solar shortwave radiation inversion.

[0143] The hard target loss is used to measure the difference between the student model output and the hard label. The hard label is the ground solar shortwave radiation observation data. The loss value is calculated using the mean square error. The expression is:

[0144] (13);

[0145] in, refers to hard target loss, is the total number of samples in the ground observation data, It is The real solar shortwave radiation data of samples (ground observations), The student model is The predicted value of solar shortwave radiation on samples.

[0146] The soft target loss is used to measure the difference between the output of the student model and the teacher model. The loss value is also calculated using the mean squared error:

[0147] (14);

[0148] in, is the soft objective function, is the number of training samples used in the knowledge distillation process, For the teacher model The predicted value on samples, that is, the soft label or the output of the teacher model, For the student model The predicted value on the sample is the value obtained by the student model in the process of imitating the output of the teacher model.

[0149] Combining the soft target loss and hard target loss to define the total loss function of distillation can ensure that the student model not only learns the knowledge of the teacher model but also accurately predicts the radiation value of the ground observation data. The expression is as follows:

[0150] (15);

[0151] in, is the total loss function during the distillation process, is a hyperparameter. It controls the weight between soft target loss and hard target loss. Its value range is between [0, 1], and its initial value is usually set to 0.5. If the teacher model has high accuracy, you can increase If the distribution of ground observation data is significantly different from that of the teacher model, then the value of , allowing the student model to focus more on hard targets.

[0152] Specifically, step S5 further includes the following steps:

[0153] S5.3, low learning rate is crucial for model stability and generalization ability. Set the initial learning rate to 10 -5 And gradually reduce the learning rate to stabilize the training process and avoid overfitting. The learning rate decay strategy is exponential decay, and the expression is as follows:

[0154] (16);

[0155] in, is the current learning rate, is the initial learning rate, Is the decay rate, usually takes a value less than 1. The smaller the value, the faster the learning rate decays. is the number of cycles of current training, is the attenuation step length, which refers to the number of cycles after which attenuation is performed.

[0156] Specifically, the generation of the student model includes the following steps:

[0157] Step 1: The geostationary satellite data and geographic information data obtained by S5.1 that are consistent with the position of the ground observation station are used as input data, and the ground-observed solar shortwave radiation data are used as labels.

[0158] Step 2: Perform data cleaning, feature conversion, and feature selection on the input data. The feature conversion method is as in step S2, and the feature selection method is as in step S3.

[0159] Step 3: Then input the neural network in step S4.1 to generate the student model.

[0160] Through the above steps, the solar shortwave radiation inversion optimization algorithm based on knowledge distillation proposed in the present invention can effectively combine satellite data with ground observation data, and provide more accurate estimation results in large-scale, high-precision solar shortwave radiation inversion tasks.

[0161] Figure 2 The comparison chart of the knowledge distillation optimization model effect is shown. Figure 2 The pre-trained model result in is the prediction result of the teacher model. Figure 2The result of knowledge distillation in is the prediction result of the student model. Figure 2 It can be seen that the results after knowledge distillation are close to the ground observations.

[0162] The present invention combines ground observation data with satellite data through a knowledge distillation strategy, achieving a balance between the accuracy and spatial coverage of solar shortwave radiation inversion. Using ground observation data to optimize the student model significantly improves the accuracy of the student model, reduces dependence on ground observations, and reduces data collection costs. On this basis, the student model has stronger regional adaptability and generalization capabilities. At the same time, the present invention supports dynamic change monitoring with high temporal and spatial resolution, meets the needs of real-time and refined environmental monitoring, and provides more reliable technical support for fields such as climate research and solar resource assessment.

[0163] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A satellite solar shortwave radiation inversion method based on knowledge distillation, characterized in that: The specific steps include: S1, collects polar-orbiting satellite data, geostationary satellite data, geographic information data, and ground observation data; S2, preprocessing geostationary satellite data, polar-orbiting satellite data, ground observation data, and latitude and longitude data and surface elevation data in geographic information data to obtain input features and labels for teacher model and student model training; S3, using the random forest model to rank the importance of input features, select the key features of satellite solar shortwave radiation inversion, and eliminate non-key features; S4, uses key features to build a deep learning model to generate a teacher model for solar shortwave radiation inversion; S5: The output of the teacher model is used as a soft label, and the ground observation data is introduced as a hard label for knowledge distillation to obtain the student model of solar shortwave radiation inversion; Step S3 specifically includes the following steps: S3.

1. For each decision tree, the input features include preprocessed geostationary satellite data and geographic information data. When an input feature is split at a node, the reduction in the contribution of the node to the prediction error is calculated: （4）； in, is the error reduction when the node is split, 、 、 are the errors of the parent node, left node and right node respectively, and are the proportions of the number of left and right node samples in the total number of parent node samples; S3.2, each feature will have multiple split nodes in each tree, and the sum of all nodes The sum is then averaged across the forest to get the overall importance of the feature: （5）； in, is the importance of the feature, is the number of trees in the random forest, Indicates the A tree, Represents all nodes in the tree; Step S4 specifically includes the following steps: S4.1, the neural network includes: an input layer, a fully connected layer, and an output layer connected in sequence, and the key features are input into the input layer of the neural network; S4.2 is then input to the fully connected layer, and nonlinearity is introduced through the activation function. The calculation formula of the first layer of the fully connected layer is: （6）； in, is the output latitude of the first layer, Is the activation function for nonlinear mapping, is the weight matrix of the first layer, is the bias vector of the first layer; S4.3, the fully connected layer The calculation formula of the layer is: （7）； in, It is The output of the layer, It is The weight matrix of the layer, is the bias term, It is The output dimension of the layer; S4.4 is finally output through the output layer. The calculation formula of the output layer is: （8）； in, is the predicted value of the model, which represents the solar shortwave radiation data at the corresponding moment. is the weight matrix of the output layer, is the output dimension of the penultimate layer, is the bias term of the output layer; S4.5, set the mean square error loss function ; Step S5 specifically includes the following steps: S5.

1. Use bilinear spatial interpolation to find the geostationary satellite data and geographic information data that are consistent with the ground observation station position. Estimate the value of the target point using four adjacent points. First, perform horizontal interpolation: （10）； （11）； Then perform longitudinal interpolation: （12）； in, , , , are the values ​​of four known adjacent points, , , , are the coordinates of the known data points, is the estimated target point coordinate; S5.2: Use polar-orbiting satellite data to train the teacher model. The solar shortwave radiation data predicted by the teacher model is used as soft labels, and the ground observation data is used as hard labels to train the student model. The soft label and hard label training are optimized using soft target loss and hard target loss, respectively. The student model prediction results are denormalized to obtain the optimized results of solar shortwave radiation inversion. The hard target loss is used to measure the difference between the student model output and the hard label. The loss value is calculated using the mean square error, and the expression is: （13）； in, refers to hard target loss, is the total number of samples in the ground observation data, It is The real solar shortwave radiation data of samples, The student model is The predicted value of solar shortwave radiation on samples; The soft target loss is used to measure the difference between the output of the student model and the teacher model. The loss value is also calculated using the mean squared error: （14）； in, is the soft objective function, is the number of training samples used in the knowledge distillation process, For the teacher model The predicted value on samples, that is, the soft label or the output of the teacher model, For the student model The predicted value on the sample, that is, the value obtained by the student model in the process of imitating the output of the teacher model; Combining the soft target loss and the hard target loss defines the total loss function of distillation, which is expressed as follows: （15）； in, is the total loss function during the distillation process, is a hyperparameter; Step S5 also includes the following steps: S5.3, set the initial learning rate to 10 -5 And gradually reduce the learning rate, using the learning rate decay strategy of exponential decay, the expression is as follows: （16）； in, is the current learning rate, is the initial learning rate, is the decay rate, is the number of cycles of current training, is the attenuation step length.

2. The satellite solar shortwave radiation inversion method based on knowledge distillation according to claim 1 is characterized in that: Geostationary satellite data include: spectral channel data and geometric feature data; geometric feature data include: satellite zenith angle, satellite azimuth angle, solar zenith angle and solar azimuth angle; Geographic information data includes: longitude and latitude data, surface elevation data and surface type data; Polar-orbiting satellite data include: solar shortwave radiation data; Ground observation data include: ground solar shortwave radiation observation data.

3. The satellite solar shortwave radiation inversion method based on knowledge distillation according to claim 2 is characterized in that: Step S2 specifically includes the following steps: S2.

1. Standardize the spectral channel data from geostationary satellite data, the solar shortwave radiation data from polar-orbiting satellite data, and the latitude and longitude data and surface elevation data from geographic information data: （1）； in, is the normalized eigenvalue, is the original data, is the mean of the original data, is the standard deviation of the original data; S2.2, normalize the data after standardization in S2.1: （2）； in, is the normalized eigenvalue, and are the minimum and maximum values ​​of the original data respectively; S2.3, perform sine and cosine transforms on the geometric characteristic data of the geostationary satellite; S2.4, perform one-hot encoding on the surface type data in the geographic information data and convert the surface type data into a numerical format: （3）； in, is the feature vector after one-hot encoding, is the number of surface types, Indicates that if the surface type characteristics belong to Class is , otherwise 0, where .

4. The satellite solar shortwave radiation inversion method based on knowledge distillation according to claim 1 is characterized in that: The generation of the student model includes the following steps: The geostationary satellite data and geographic information data obtained by S5.1 that are consistent with the position of the ground observation station are used as input data, and the ground-observed solar shortwave radiation data are used as labels; Perform data cleaning, feature conversion, and feature selection on input data; This is then input into the neural network described in step S4.1 to generate a student model.

Images