Satellite chip-level algorithm construction method for atmospheric anomaly instant remote sensing detection

Abstract

The invention discloses a satellite chip level algorithm construction method for atmospheric anomaly instant remote sensing detection, relates to the technical field of atmospheric anomaly detection, and constructs angle information fusion wave band characteristics based on hyperspectral satellite remote sensing data and ground monitoring station data in combination with gas absorption spectrum characteristics. Generating a data set of remote sensing information fused with atmospheric parameters; constructing a deep learning model based on atmospheric parameter quantitative inversion, evaluating model precision in combination with ground station data, and modifying different feature weights to improve model inversion precision; lightweight operation is carried out on the constructed deep learning model based on atmospheric parameter quantitative inversion, the parameter quantity is reduced to obtain a lightweight model capable of being operated on a low-power-consumption chip, the on-satellite operation condition is simulated on a development board, and the in-orbit operation performance and precision of the model are comprehensively evaluated. The model constructed by the invention not only effectively reduces the calculation load, but also remarkably improves the satellite real-time data processing efficiency on the basis of keeping the precision.

Description

Technical Field

[0001] The present invention relates to the technical field of atmospheric anomaly detection, and more specifically to a satellite chip-level algorithm construction method for real-time remote sensing detection of atmospheric anomalies. Background Art

[0002] Monitoring and warning the distribution and severity of atmospheric anomalies are crucial for climate management and the dual carbon strategy. Traditional ground-based monitoring methods have advantages for air quality monitoring at small spatial scales, but ground-based stations are sparsely distributed, construction and operation costs are high, and monitoring efficiency and scope cannot meet existing needs; satellite remote sensing technology has the advantages of large spatial scale coverage and good continuity, which helps to make up for the disadvantages of ground-based monitoring and has gradually become a powerful tool for detecting atmospheric anomalies and quantifying severity.

[0003] Among the existing on-orbit chip-level processing technologies, the main focus is on developing on-orbit data processing technologies, with a focus on on-orbit cloud judgment, target detection, image processing, etc. Satellites equipped with on-orbit cloud judgment technology can detect clouds on-orbit and filter out images of cloud-covered areas, effectively reducing the amount of data transmitted; satellites equipped with target detection algorithms can perform on-orbit intelligent detection of typical targets such as forest fires and ships on the sea, improving data analysis efficiency; satellites equipped with image processing technology can perform image compression, radiation correction, geometric correction and other image processing operations on remote sensing images on-orbit, improving data processing and transmission efficiency.

[0004] However, there is no on-orbit chip-level algorithm for real-time remote sensing detection of atmospheric anomalies. Due to the lag of remote sensing data, the traditional methods of using remote sensing to study atmospheric anomalies are mainly divided into physical model methods and machine learning methods. The main idea of ​​the physical model method is to consider the interaction between electromagnetic wave radiation and various components in the atmosphere, and to build a physical model based on radiation transfer theory, such as the Iterative Maximum A Posteriori Differential Optical Absorption Spectroscopy (IMAP-DOAS) method, which can use a grating spectrometer to jointly estimate the concentration of different gas components in the atmosphere (such as water vapor, carbon dioxide, methane, etc.). The machine learning method constructs a deep learning model, inputs a large amount of data into the model for training, and mines its implicit features, so as to make predictions when new data is input. At present, atmospheric anomaly data is mainly simulated by meteorological models such as the Large Eddy Simulation extension of the Weather Research and Forecast Model (WRF-LES) as training data sets for input into deep learning models. This method ignores the impact of the incident angle information of remote sensing images on reflectivity, and it is difficult to construct reflectivity differences between hyperspectral bands. In addition, the constructed deep models often have a large number of parameters and high requirements for the computing platform, which makes it difficult to meet the requirements of low power consumption and low computing power at the satellite on-orbit chip level, and thus cannot achieve the construction of satellite-level algorithms.

[0005] Therefore, how to provide a satellite chip-level algorithm construction method for real-time remote sensing detection of atmospheric anomalies, deeply analyze the changes in reflected signals caused by different observation angles, improve the model inversion accuracy, explore the relationship between different gases and the absorption strength of different spectral signals according to the atmospheric radiation transmission theory, and ensure that the model accuracy loss is small by adjusting the quantitative expression to realize the construction of satellite-level algorithms is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies, constructs a deep learning model to detect atmospheric anomaly areas, and implements satellite chip-level algorithm transplantation based on the deep learning model to ensure that the inversion accuracy reaches the expected standard. The algorithm is verified for accuracy and performance evaluation by simulating on-board operating conditions, providing reliable atmospheric parameter inversion support for actual application scenarios.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: a satellite chip-level algorithm construction method for real-time remote sensing detection of atmospheric anomalies, comprising:

[0008] Acquire hyperspectral satellite remote sensing data and ground monitoring site data;

[0009] Based on the hyperspectral satellite remote sensing data and the ground monitoring site data, angle information fusion band features are constructed to generate a data set of remote sensing information fusion atmospheric parameters;

[0010] Preprocessing the data set, and dividing the preprocessed data set into a training set, a test set, and a validation set;

[0011] Based on the data set, feature importance is calculated by LightGBM or XGBoost algorithm, and features are screened according to the feature importance;

[0012] Constructing a deep learning model, inputting the training set into the deep learning model for training, and optimizing the inversion accuracy in combination with the ground monitoring station data to obtain a deep learning model based on quantitative inversion of atmospheric parameters;

[0013] The deep learning model based on the quantitative inversion of atmospheric parameters is lightweighted to obtain a lightweight model;

[0014] The lightweight model is loaded onto a development board to simulate the operating conditions on the satellite, monitor the atmospheric anomaly areas, and invert and identify the atmospheric parameters of the atmospheric anomaly areas.

[0015] Preferably, constructing angle information fusion band features includes:

[0016] Based on the gas spectral absorption characteristics, the corresponding sensitive original band information is selected for different gases;

[0017] The angle information in the hyperspectral satellite remote sensing data is fused with the original band information to obtain angle information fused band features;

[0018] The normalized difference method is used to calculate the characteristics of the fusion bands of different angle information to generate a data set of remote sensing information fusion atmospheric parameters.

[0019] Preferably, the angle information contained in the hyperspectral satellite remote sensing data includes: observation zenith angle, relative azimuth angle and solar zenith angle.

[0020] Preferably, the angle information fusion band feature is expressed as:

[0021]

[0022] Among them, band raw is the original band information, θ v is the observed zenith angle, θ ris the relative azimuth angle, θ s is the solar zenith angle.

[0023] Preferably, a gradient boosting decision tree model in the LightGBM or XGBoost algorithm is used, the data set is input into the gradient boosting decision tree model, and the feature importance is calculated.

[0024] Preferably, the deep learning model adopts DenseNet algorithm;

[0025] Use the training set to train the deep learning model, set the optimizer and loss function for back propagation, and adjust the model weights;

[0026] The output of the deep learning model is evaluated in combination with ground monitoring station data, feature weights are adjusted to optimize inversion accuracy, and hyperparameters are adjusted through the validation set to avoid model overfitting;

[0027] The model performance is evaluated through the test set to ensure that the inversion accuracy meets the expected standard. The model parameters of the deep learning model are adjusted according to the evaluation results. The deep learning model is applied to satellite-level algorithm transplantation and performance verification to provide reliable atmospheric parameter inversion support for actual application scenarios.

[0028] Preferably, the architecture of the deep learning model includes: a plurality of dense blocks and transition layers;

[0029] The dense blocks are connected via the transition layer;

[0030] The dense block is used for feature accumulation learning, and the transition layer is used for feature map compression.

[0031] Preferably, the deep learning model based on the quantitative inversion of atmospheric parameters is lightweighted to obtain a lightweight model, including:

[0032] The input, output, and weight of the deep learning model based on the quantitative inversion of atmospheric parameters are processed using different scaling factors and zero points, and the floating-point parameters of the deep learning model based on the quantitative inversion of atmospheric parameters are converted into integer parameters. The calculation method is as follows:

[0033]

[0034] Among them, q is the quantized integer value, x is the original floating-point value, scale is the scaling factor, and zeropoint is the zero point.

[0035] Preferably, the lightweight model is loaded onto a development board, the operating conditions on the satellite are simulated, the atmospheric anomaly areas are monitored, and the atmospheric parameters of the atmospheric anomaly areas are inverted and identified, including:

[0036] Load the lightweight model;

[0037] Simulate satellites taking remote sensing images in orbit;

[0038] Call the development board NPU to perform model inference and invert the corresponding atmospheric parameters;

[0039] Output the inversion results and add the corresponding layer on the remote sensing image to visualize the results and get the model accuracy index in the result panel.

[0040] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies, which relates to the technical field of atmospheric anomaly detection. Based on hyperspectral satellite remote sensing data and ground monitoring site data, combined with gas absorption spectrum characteristics, angle information fusion band characteristics are constructed to produce remote sensing information fusion atmospheric parameter data sets; a deep learning model based on quantitative inversion of atmospheric parameters is constructed, the model accuracy is evaluated in combination with ground site data, and different feature weights are modified to improve the model inversion accuracy; lightweight operations are performed on the constructed deep learning model based on quantitative inversion of atmospheric parameters to reduce the amount of parameters to obtain a lightweight model that can be operated on a low-power chip, and the operating conditions on the satellite are simulated on the development board to comprehensively evaluate the model's on-orbit performance and accuracy.

[0041] The beneficial effects of the present invention are as follows:

[0042] Construct a fusion data set: By fusing image angle information, correlating spectral reflection signals with observation angles, improving the deep learning model's learning of the relationship between satellite observation angles and reflection signals, deeply analyzing the changes in reflection signals caused by different observation angles, and effectively improving the model inversion accuracy.

[0043] Screening features to improve model inversion accuracy: Based on the theory of atmospheric radiation transfer, the relationship between different gases and the absorption strength of signals in different spectral bands is explored. The importance of features is obtained based on model reasoning, and features with higher importance are screened. The feature expressions are modified in combination with physical mechanisms. This can effectively improve the model inversion accuracy and provide a theoretical basis for model interpretability.

[0044] Model asymmetric quantization: Lightweight models can significantly reduce model storage space and improve model calculation efficiency to meet the needs of low-power and low-computing chips on board. At the same time, by adjusting the quantization expression, the model accuracy loss is ensured to be small. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0046] Figure 1 A schematic flow chart of a method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies provided by the present invention.

[0047] Figure 2 A schematic diagram of a flow chart for constructing a satellite-level chip algorithm using the inversion of abnormal PM2.5 concentration as an example provided in an embodiment of the present invention.

[0048] Figure 3 A schematic diagram of an on-satellite operating condition simulation interface provided in an embodiment of the present invention.

[0049] Figure 4 A schematic diagram of the model accuracy evaluation results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The embodiment of the present invention discloses a satellite chip-level algorithm construction method for real-time remote sensing detection of atmospheric anomalies, which not only effectively reduces the computing load, but also significantly improves the efficiency of satellite real-time data processing while maintaining accuracy, laying a solid foundation for subsequent satellite data analysis and application.

[0052] In a specific embodiment of the present invention, a satellite chip-level algorithm construction method for real-time remote sensing detection of atmospheric anomalies is provided. Figure 1 As shown, including:

[0053] Acquire hyperspectral satellite remote sensing data and ground monitoring site data;

[0054] Based on the hyperspectral satellite remote sensing data and the ground monitoring site data, angle information fusion band features are constructed to generate a data set of remote sensing information fusion atmospheric parameters;

[0055] Preprocessing the data set, and dividing the preprocessed data set into a training set, a test set, and a validation set;

[0056] Based on the data set, feature importance is calculated by LightGBM or XGBoost algorithm, and features are screened according to the feature importance;

[0057] A deep learning model is constructed based on the feature importance and the screened features, the training set is input into the deep learning model for training, and the inversion accuracy is optimized in combination with the ground monitoring site data to obtain a deep learning model based on quantitative inversion of atmospheric parameters;

[0058] The deep learning model based on the quantitative inversion of atmospheric parameters is lightweighted to obtain a lightweight model;

[0059] The lightweight model is loaded onto a development board to simulate the operating conditions on the satellite, monitor the atmospheric anomaly areas, and invert and identify the atmospheric parameters of the atmospheric anomaly areas.

[0060] S1. Information fusion dataset production

[0061] The embodiment of the present invention is based on hyperspectral satellite remote sensing data and ground monitoring site data, combined with gas absorption spectrum characteristics, to construct angle information fusion band features, produce remote sensing information fusion atmospheric parameter data sets, and divide them into training data sets, validation data sets, and test data sets. The LightGBM and XGBoost model algorithms are used to screen feature importance for subsequent deep learning model construction, accuracy, and performance evaluation. Figure 1 As shown, the information fusion dataset production includes three processes: angle information fusion band feature construction, dataset division and feature importance screening.

[0062] S101, Angle information fusion band feature construction

[0063] Specifically, based on the gas spectral absorption characteristics, the corresponding response sensitive original band information is selected for different gases;

[0064] The angle information in the hyperspectral satellite remote sensing data is fused with the original band information to obtain angle information fused band features;

[0065] The normalized difference method is used to calculate the characteristics of the fusion bands of different angle information to generate a data set of remote sensing information fusion atmospheric parameters.

[0066] In the embodiment of the present invention, the angle information fusion data set production method is: based on the gas absorption spectrum characteristics, the strong absorption bands of different gases are selected, and the sine and cosine values ​​of the angle information (view zenith angle, view zenith angle; solar zenith angle, solar zenith angle; relative azimuth angle, relative azimuth angle) of the remote sensing image are fused to obtain the angle information fusion band characteristics. And the normalized difference method is used to calculate the different angle information fusion band characteristics to produce a feature data set for subsequent deep learning model training.

[0067] Specifically, the angle information contained in the hyperspectral satellite remote sensing data includes: observation zenith angle, relative azimuth angle and solar zenith angle.

[0068] Specifically, the angle information fusion band feature is expressed as:

[0069]

[0070] Among them, band raw is the original band information, θ v is the observed zenith angle, θ r is the relative azimuth angle, θ s is the solar zenith angle.

[0071] The core idea of ​​this formula is to perform weighted fusion of each band data with its corresponding angle sine and cosine values ​​to reflect the changes in spectral characteristics under different observation angles. This fusion method can integrate the effects of different directional lighting conditions and observation angles on spectral information, enhance the ability of band data to express actual objects, and show strong robustness and adaptability, especially under complex observation conditions.

[0072] In an embodiment of the present invention, the three angles (view zenith angle, relative azimuth angle and solar zenith angle) contained in the satellite data are fused with the original band information to improve the accuracy and richness of the data. This fusion method helps to capture the changes in the spectral characteristics of surface targets under different observation conditions, thereby providing more refined information for subsequent environmental monitoring and remote sensing analysis.

[0073] S102. Dataset Division

[0074] In this embodiment of the present invention, in order to ensure the reasonable division of the data set during the machine learning model training process and the objective evaluation of the model performance, a random division method is adopted to divide the data set into a training set, a validation set and a test set, and the ratio is set to 7:2:1.

[0075] S1021. Data preprocessing

[0076] Before data segmentation, necessary preprocessing is performed on the data set, including missing value processing and outlier detection, to ensure that the data quality meets the input requirements of the machine learning model. The satellite remote sensing data is extracted and stored in a CSV table in an image-by-pixel-by-band manner, where the band is the column variable and the pixel is the row variable. The satellite remote sensing data is matched pixel by pixel with the ground monitoring site data according to the latitude and longitude information of the pixel and the ground monitoring site, and the ground monitoring site data is stored in the CSV table as a column variable.

[0077] Since some pixel values ​​in satellite data may be -999, which are missing values, bands with more than 60% missing values ​​and pixels with more than 50% missing values ​​are directly discarded, and other columns or rows with missing values ​​can be selected according to the size of the original data. When the data size is large, all rows or columns with missing values ​​can be discarded.

[0078] Infinite values ​​and values ​​greater than a reasonable range may appear in the pixel values ​​in satellite data and ground monitoring data, all of which are outliers. In the embodiment of the present invention, the appearance of outliers will seriously affect the subsequent model training effect, so all rows (pixels) with outliers are discarded.

[0079] S1022. Dataset Division

[0080] Specifically, the ratio of training set, validation set, and test set is set to 7:2:1, that is, 70% of the data is used as training set (train data), 20% of the data is used as validation set (valid data), and 10% of the data is used as test set (test data). This division method can ensure the adequacy of training data while retaining enough validation set and test set for model tuning and performance evaluation.

[0081] The embodiment of the present invention uses the Python programming language and uses the train_test_split function in the Scikit-learn library to implement the division of the data set according to the following steps:

[0082] import pandas as pd

[0083] from sklearn.model_selection import train_test_split

[0084] #Import data

[0085] #Assume the CSV file path is 'data.csv'

[0086] data=pd.read_csv('data.csv')

[0087] #Divide features and labels

[0088] #Assume that the last column is the label and the remaining columns are features

[0089] features = data.iloc[:,:-1] # Take all rows and all columns except the last column

[0090] labels = data.iloc[:,-1] # Take the last column of all rows as labels

[0091] #The first step is to divide the data set into a training set (70%) and a temporary set (30%)

[0092] train_features,temp_features,train_labels,temp_labels=train_test_split(

[0093] features,labels,test_size=0.3,random_state=42 )

[0095] #The second step is to divide the temporary set (30%) into a validation set (20%) and a test set (10%)

[0096] val_features,test_features,val_labels,test_labels=train_test_split(

[0097] temp_features,temp_labels,test_size=1 / 3,random_state=42 )

[0099] S103. Feature importance screening

[0100] By calculating the feature importance in the LightGBM or XGBoost algorithm, the features with a high impact on the model inversion accuracy are screened out. In combination with the absorption band positions of different gases in the atmospheric radiation transfer theory, the physical influence mechanism of the reflection signal is explored, and the feature expression is further modified to improve the model inversion accuracy.

[0101] Specifically, a gradient boosting decision tree model in the LightGBM or XGBoost algorithm is used, the data set is input into the gradient boosting decision tree model, and the feature importance is calculated.

[0102] In the embodiment of the present invention, in order to deeply analyze and evaluate the importance of each feature in the data set to the model prediction results, two advanced ensemble learning algorithms, LightGBM (Light Gradient Boosting Machine) and XGBoost (Extreme Gradient Boosting), are used. These two algorithms can efficiently calculate and extract the importance of features by constructing a gradient boosting decision tree model, providing a scientific and reliable basis for subsequent model optimization and feature selection.

[0103] LightGBM and XGBoost have significant advantages, which are reflected in efficient training speed, excellent generalization ability and powerful feature selection ability, especially when dealing with high-dimensional and complex data sets, which can significantly improve model performance. In the process of feature selection, these two models use indicators such as gain, frequency and coverage to quantitatively evaluate the contribution of each feature to model prediction, effectively capture the nonlinear relationship and interaction between features, and thus reveal the key factors affecting model output.

[0104] Specifically, gain: LightGBM and XGBoost measure the importance of features through the gain indicator of the tree model. Gain refers to the reduction in the loss value of the objective function each time a feature is used to split a node. By accumulating the gain of each feature in the tree structure, the contribution of the feature to the final model prediction can be quantified. A higher gain means that the feature plays a greater role in the splitting process, indicating that it is more important in the model.

[0105] Frequency: refers to the number of times a feature is used for node splitting. By counting the frequency of each feature in the model, the importance of the feature can be roughly determined. The higher the frequency of the feature, the more important it plays in different levels of multiple decision trees.

[0106] Coverage: Coverage measures the number of data samples covered by a split node when a feature is used for splitting. Coverage can help us understand how many samples a feature widely covers in the tree model, thereby quantitatively evaluating the overall impact of the feature in model prediction.

[0107] As decision tree-based algorithms, LightGBM and XGBoost are naturally suitable for handling nonlinear relationships and interactions between features. The tree model can automatically capture the nonlinear effects of different features in different intervals by splitting different features layer by layer. The construction process of each tree is essentially to continuously find the best split point of the feature, which can effectively capture the complex interactions between features.

[0108] In particular, LightGBM and XGBoost allow for cross-feature learning, that is, by building a multi-layer tree structure, the model can identify the cross-effects of different features. For example, if feature A is only valid in a certain value range of feature B in the model, such an interactive relationship can be automatically captured by the splitting rules in the tree model.

[0109] The Feature Importance Plot can be used to intuitively display the gain, frequency, or coverage of each feature, thereby intuitively and quantitatively evaluating the contribution of each feature.

[0110] S2. Deep learning model construction

[0111] The method for constructing the deep learning model is as follows: based on the angle information fusion data set produced in step S1, combined with physical mechanism analysis, remote sensing band fusion features sensitive to different gases are constructed. The constructed features and the original band information of the satellite remote sensing image are used as continuous variables, and the season and the like are used as categorical variables, and are input into the DenseNet model for training to obtain a deep learning model for quantitative inversion of target atmospheric parameters.

[0112] The embodiment of the present invention constructs a deep learning model based on the quantitative inversion of atmospheric parameters based on the training set produced in S1, evaluates the model accuracy in combination with ground station data, modifies different feature weights to improve the model inversion accuracy, and is used for subsequent satellite-level algorithm transplantation and performance evaluation.

[0113] Specifically, the deep learning model adopts the DenseNet algorithm;

[0114] Use the training set to train the deep learning model, set the optimizer and loss function for back propagation, and adjust the model weights;

[0115] The output of the deep learning model is evaluated in combination with ground monitoring station data, feature weights are adjusted to optimize inversion accuracy, and hyperparameters are adjusted through the validation set to avoid model overfitting;

[0116] The model performance is evaluated through the test set to ensure that the inversion accuracy meets the expected standard. The model parameters of the deep learning model are adjusted according to the evaluation results. The deep learning model is applied to satellite-level algorithm transplantation and performance verification to provide reliable atmospheric parameter inversion support for actual application scenarios.

[0117] Specifically, the architecture of the deep learning model includes: multiple dense blocks and transition layers;

[0118] The dense blocks are connected via the transition layer;

[0119] The dense block is used for feature accumulation learning, and the transition layer is used for feature map compression.

[0120] In an embodiment of the present invention, the data set constructed in the above steps is input into a DenseNet (Densely Connected Convolutional Networks) model for training.

[0121] The DenseNet model uses a densely connected convolutional neural network architecture. Its design feature is that the output of each layer is directly connected to all subsequent layers, thereby maximizing the efficiency of feature transfer, reducing the gradient vanishing problem, and improving the parameter utilization of the model. The model architecture includes multiple Dense Blocks and Transition Layers, which are used for feature accumulation learning and feature map compression, respectively, to enhance the model's nonlinear feature extraction capabilities. Dense Block consists of multiple convolutional layers, and the output of each layer is spliced ​​with the output of all previous layers to ensure the layer-by-layer transmission and sharing of features, thereby improving the model's ability to learn complex atmospheric parameters. Each Dense Block contains batch normalization, activation function (ReLU) and convolution operations to fully extract spatial features. Transition Layer is used to connect different Dense Blocks, including 1×1 convolutional layers and average pooling layers, which are used to reduce the number of channels and size of feature maps, control model complexity, and prevent overfitting.

[0122] The DenseNet model was trained using the training set, and the optimizer (Adam) and loss function (mean square error) were set for back propagation to adjust the model weights. The model output was evaluated using the measured data from the ground station, and the feature weights were adjusted to optimize the inversion accuracy. Hyperparameters were adjusted using the validation set to avoid overfitting of the model.

[0123] The model performance is evaluated in combination with the test set to ensure that the inversion accuracy meets the expected standards. The model parameters are fine-tuned based on the evaluation results, and the model is prepared to be applied to satellite-level algorithm transplantation and performance verification to provide reliable atmospheric parameter inversion support for actual application scenarios.

[0124] S3, satellite-level algorithm transplantation

[0125] The embodiment of the present invention performs quantization operations on the deep learning model based on the quantitative inversion of atmospheric parameters, reduces the number of parameters, and obtains a lightweight model that can be operated on a low-power chip. On this basis, the lightweight model is evaluated for accuracy, and the model accuracy is continuously optimized by adjusting the quantization parameters to reduce the accuracy loss of the model quantization. The on-board operating conditions are simulated on the development board to comprehensively evaluate the on-orbit performance and accuracy of the model.

[0126] Specifically, the RKNN-Toolkit2 technology is used to lightweight the deep learning model based on the quantitative inversion of atmospheric parameters to obtain a lightweight model;

[0127] The embodiment of the present invention uses RKNN-Toolkit2 technology to lightweight the constructed deep learning model and realize the algorithm transplantation from the server side to the satellite on-orbit chip side. The chip-level operation simulates the data processing and operation process in the on-orbit environment of the satellite, and combines the ground verification data to evaluate the accuracy and performance of the lightweight chip-level model.

[0128] S301, Model Quantization

[0129] Specifically, different scaling factors and zero points are used to process the input, output, and weight of the deep learning model based on quantitative inversion of atmospheric parameters, and the floating-point parameters of the deep learning model based on quantitative inversion of atmospheric parameters are converted into integer parameters. The calculation method is as follows:

[0130]

[0131] Where q is the quantized integer value, x is the original floating-point value, scale is the scaling factor, and zeropoint is the zero point. The scaling factor is used to scale the floating-point range to the quantized integer range. The zero point defines the zero offset of the quantized value, allowing the zero value of the floating-point number to not be exactly aligned with the integer zero value, thereby adapting to the distribution characteristics of the data.

[0132] The deep learning model quantization method described in the embodiment of the present invention is: by using an asymmetric value range in the quantization process (that is, the quantization zero point does not have to be aligned with the data zero point), the floating point number is mapped to an integer value, thereby reducing the storage and computing requirements of the model while maintaining the model accuracy as much as possible. Asymmetric quantization can significantly reduce the storage space and computational complexity of the model while maintaining the accuracy of the model, and is suitable for resource-constrained hardware platforms. The embodiment of the present invention converts the floating point parameters (float32) of the original model into low-precision integer parameters (unit8). Unlike symmetric quantization, asymmetric quantization uses different scaling factors and zero points to independently process the input, output, and weight of the model. This method is more flexible and can effectively reduce the accuracy loss after the model is quantized.

[0133] S302, on-board operating condition simulation

[0134] Specifically, the lightweight model is loaded onto a development board, the on-board operating conditions are simulated, the atmospheric anomaly areas are monitored, and the atmospheric parameters of the atmospheric anomaly areas are inverted and identified, including:

[0135] Load the lightweight model;

[0136] Simulate satellites taking remote sensing images in orbit;

[0137] Call the development board NPU to perform model inference and invert the corresponding atmospheric parameters;

[0138] Output the inversion results and add the corresponding layer on the remote sensing image to visualize the results and get the model accuracy index in the result panel.

[0139] The embodiment of the present invention utilizes a "satellite on-orbit test computing system" (hereinafter referred to as the system) to perform on-board working condition simulation testing on a lightweight model, and the system runs on a development board kit based on the Ubuntu system.

[0140] First, load the on-orbit model in the system, that is, the lightweight model constructed in the above embodiment of the present invention. Then, through the "on-orbit shooting" function, simulate the satellite shooting of remote sensing images in orbit. Through the "on-orbit calculation" function, the development board NPU can be called to perform model reasoning and invert the corresponding atmospheric parameters. Finally, through the "result push" function, the model inversion results are output and the corresponding layer is added to the remote sensing image to visualize the results, and the model accuracy indicators (Accuracy, Prediction, Recall) are obtained on the result panel.

[0141] Through the system panel, the number of called threads and sample ratio can be quickly adjusted, and information such as development board power, NPU occupancy, memory occupancy, etc. can be monitored in real time during on-orbit computing.

[0142] In a specific embodiment of the present invention, the present invention takes the PM2.5 abnormal concentration inversion as an example to construct a satellite-level chip algorithm. The specific technical roadmap is as follows: Figure 2 As shown. The multispectral satellite data used in this embodiment is the Himawari-8 (Sunflower-8) satellite, which is obtained through the P-tree System website (https: / / www.eorc.jaxa.jp / ptree / ). The ground monitoring site data is obtained through the National Urban Air Quality Real-time Publishing Platform (https: / / air.cnemc.cn:18007). The remote sensing image data and the ground monitoring site data are matched according to time and longitude and latitude to obtain the original data set, which includes satellite multi-band reflection signal data, angle information, time information, longitude and latitude information and ground monitoring site PM2.5 data.

[0143] Information fusion dataset production

[0144] Based on the above original data set, according to the information fusion data set production method proposed in the present invention, based on the gas absorption spectrum characteristics, the 1-16 bands of the image are fused with the sine and cosine values ​​of the angle information (band17-20) of the image to obtain the angle information fusion band characteristics. And the band differences (band9-band13, band9-band15, band10-band15, band11-band13, band13-band15) are added to amplify the response to changes in PM2.5 concentration. In addition, the present embodiment also utilizes the advantage of the high temporal resolution of the Himawari-8 satellite as a geostationary orbit satellite to construct the grid-by-grid feature fusion data of the current grid data / grid monthly minimum and the data 24 hours ago / monthly minimum.

[0145] In this embodiment, the data of 2023 is used as the model training data set, and the data of 2024 is used as the validation data set to verify the model accuracy. The training data set uses the Python programming language, and with the help of the train_test_split function in the Scikit-learn library, the produced information fusion data set is randomly divided into 7:2:1 in proportion, that is, 70% training set, 20% validation set, and 10% test set.

[0146] On-board operating condition simulation

[0147] The constructed PM2.5 concentration remote sensing inversion deep learning model is lightweight and loaded into the development board. The on-board operating conditions are simulated using the "Satellite On-Orbit Simulation Test System" to comprehensively evaluate the model performance and accuracy loss.

[0148] like Figure 3As shown, when the sample size is 2481 and the number of model parameters is 278.57 million, the lightweight model of this embodiment needs 6 minutes of on-orbit calculation and reasoning on the development board. It can monitor areas with atmospheric anomalies and only invert and identify the PM2.5 concentration in areas with atmospheric anomalies, greatly saving on-board computing resources.

[0149] The model accuracy of the 2024 data was evaluated and the results are as follows: Figure 4 As shown in the figure, the correlation coefficient (R value) of the model reaches 0.78, and the root mean square error (RMSE) is 24. With limited onboard computing resources, the model can still achieve high accuracy. This shows that the lightweight model constructed by the satellite chip-level algorithm construction method exhibits excellent performance in the satellite on-orbit computing process. This model not only effectively reduces the computing load, but also significantly improves the efficiency of satellite real-time data processing while maintaining accuracy, laying a solid foundation for subsequent satellite data analysis and application.

[0150] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0151] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A satellite chip-level algorithm construction method for real-time remote sensing detection of atmospheric anomalies, characterized in that: include: Acquire hyperspectral satellite remote sensing data and ground monitoring site data; Based on the hyperspectral satellite remote sensing data and the ground monitoring site data, angle information fusion band features are constructed to generate a data set of remote sensing information fusion atmospheric parameters; Preprocessing the data set, and dividing the preprocessed data set into a training set, a test set, and a validation set; Based on the data set, feature importance is calculated by LightGBM or XGBoost algorithm, and features are screened according to the feature importance; A deep learning model is constructed based on the feature importance and the screened features, the training set is input into the deep learning model for training, and the inversion accuracy is optimized in combination with the ground monitoring site data to obtain a deep learning model based on quantitative inversion of atmospheric parameters; The deep learning model based on the quantitative inversion of atmospheric parameters is lightweighted to obtain a lightweight model; The lightweight model is loaded onto a development board to simulate the operating conditions on the satellite, monitor the atmospheric anomaly areas, and invert and identify the atmospheric parameters of the atmospheric anomaly areas.

2. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 1, characterized in that: Construct angle information fusion band features, including: Based on the gas spectral absorption characteristics, the corresponding sensitive original band information is selected for different gases; The angle information in the hyperspectral satellite remote sensing data is fused with the original band information to obtain angle information fused band features; The normalized difference method is used to calculate the characteristics of the fusion bands of different angle information to generate a data set of remote sensing information fusion atmospheric parameters.

3. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 2, characterized in that: The angular information contained in the hyperspectral satellite remote sensing data includes: observation zenith angle, relative azimuth angle and solar zenith angle.

4. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 3, characterized in that: The angle information fusion band feature is expressed as: ； in, is the original band information, is the observed zenith angle, is the relative azimuth angle, is the solar zenith angle.

5. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 1, characterized in that: The gradient boosting decision tree model in the LightGBM or XGBoost algorithm is used, the data set is input into the gradient boosting decision tree model, and the feature importance is calculated.

6. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 1, characterized in that: The deep learning model adopts the DenseNet algorithm; Use the training set to train the deep learning model, set the optimizer and loss function for back propagation, and adjust the model weights; Evaluate the output of the deep learning model using ground monitoring site data, adjust feature weights, and perform hyperparameter tuning using a validation set; The model performance is evaluated through the test set, and the model parameters of the deep learning model are adjusted according to the evaluation results.

7. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 1, characterized in that: The architecture of the deep learning model includes: a plurality of dense blocks and transition layers; The dense blocks are connected via the transition layer; The dense block is used for feature accumulation learning, and the transition layer is used for feature map compression.

8. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 1, characterized in that: The deep learning model based on quantitative inversion of atmospheric parameters is lightweighted to obtain a lightweight model, including: The input, output, and weight of the deep learning model based on the quantitative inversion of atmospheric parameters are processed using different scaling factors and zero points, and the floating-point parameters of the deep learning model based on the quantitative inversion of atmospheric parameters are converted into integer parameters. The calculation method is as follows: ； Among them, q is the quantized integer value, x is the original floating-point value, scale is the scaling factor, and zeropoint is the zero point.

9. The method for constructing a satellite chip-level algorithm for real-time remote sensing detection of atmospheric anomalies according to claim 1, characterized in that: The lightweight model is loaded onto the development board to simulate the on-board operating conditions, monitor the atmospheric anomaly areas, and invert and identify the atmospheric parameters of the atmospheric anomaly areas, including: Load the lightweight model; Simulate satellites taking remote sensing images in orbit; Call the development board NPU to perform model inference and invert the corresponding atmospheric parameters; Output the inversion results and add the corresponding layer on the remote sensing image to visualize the results and get the model accuracy index in the result panel.

Images