Meteorological satellite data processing method based on deep learning

Abstract

The invention relates to the technical field of meteorological data processing, and discloses a meteorological satellite data processing method based on deep learning, and the method comprises the steps: 1, obtaining multi-modal meteorological satellite data which comprises infrared band data, visible light band data, multispectral band data, auxiliary ground observation data and meteorological radar data; and step 2, denoising the meteorological satellite data through a convolution auto-encoder. By designing a cross-modal self-attention mechanism, features of infrared, visible light and multispectral data are interactively fused, complementary information among different data sources is fully mined, deep feature expression of multimodal meteorological satellite data is realized, the problem that a single data source is difficult to comprehensively characterize a complex meteorological phenomenon is solved, and the accuracy and the reliability of the meteorological satellite data are improved. The effect of remarkably improving the perception ability of the model to complex meteorological phenomena is achieved, the collaborative analysis ability of multi-modal data is improved, and higher precision and robustness are provided for meteorological prediction and phenomenon detection.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological data processing, and specifically to a meteorological satellite data processing method based on deep learning. Background Art

[0002] With the rapid development of meteorological satellite technology, satellite remote sensing data has become one of the important means of meteorological monitoring and forecasting. Meteorological satellite data covers infrared, visible light and multi-spectral multimodal information, and can provide wide-range, high-resolution information on the state of the Earth's atmosphere and surface. However, multimodal data have different sources and characteristics, and there are differences in time, spatial resolution and data quality, which leads to many challenges in directly using data for meteorological phenomenon analysis. At the same time, complex meteorological phenomena, such as typhoon paths, severe convection areas and precipitation processes, have highly nonlinear and dynamic characteristics, and traditional methods cannot meet the requirements in terms of modeling and prediction accuracy.

[0003] Existing meteorological data processing methods usually rely on single-modal data and cannot fully express complex meteorological phenomena. On the other hand, traditional meteorological analysis based on physical models or statistical methods ignores the inherent connections between multimodal data, resulting in reduced accuracy in the characterization and prediction of meteorological phenomena. In addition, the spatiotemporal dynamic characteristics of complex meteorological phenomena make it impossible for traditional static analysis methods to capture their evolution laws, which restricts the ability to predict future meteorological changes. With the rapid increase in data volume, how to mine potential laws in large-scale meteorological data and efficiently realize real-time meteorological forecasts has become an urgent problem to be solved.

[0004] In recent years, deep learning technology has made significant progress in the fields of image processing and time series analysis, bringing new technical opportunities for meteorological satellite data processing. For example, convolutional neural networks have shown strong capabilities in image feature extraction, while long short-term memory network sequence models can effectively capture dependencies in the time dimension. In addition, the application of Transformer advanced models in time series forecasting has shown high accuracy. However, when applied to meteorological data, existing deep learning methods face problems such as multimodal data fusion, combination of physical laws, and optimization of model calculation efficiency, which limits their widespread application in the meteorological field.

[0005] Therefore, those skilled in the art provide a meteorological satellite data processing method based on deep learning to solve the above-mentioned problems. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a meteorological satellite data processing method based on deep learning to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A meteorological satellite data processing method based on deep learning, comprising:

[0008] Step 1: Acquire multimodal meteorological satellite data, which includes infrared band data, visible light band data, multispectral band data, auxiliary ground observation data and meteorological radar data;

[0009] Step 2: De-noise the meteorological satellite data through the convolutional autoencoder to generate denoised data, and then align the denoised data using the geo-reference method to unify the spatial resolution of different data;

[0010] Step 3: Use the time interpolation method to align the time axis of the multimodal meteorological satellite data to generate the multimodal meteorological satellite data after time and space alignment, and then extract the physical features in the multimodal meteorological satellite data after time and space alignment, the physical features including cloud top temperature, cloud optical thickness and cloud top height;

[0011] Step 4: extract features from the infrared band data, visible light band data, and multi-spectral band data in step 1, use a multi-input convolutional neural network to extract infrared band features, visible light band features, and multi-spectral band features, and obtain corresponding feature vectors. Then, interactively fuse the infrared band features, visible light band features, and multi-spectral band features through a cross-modal self-attention mechanism to generate a fused feature vector.

[0012] Step 5: For the fused feature vector generated in step 4, the target detection model is used to perform target detection of meteorological phenomena, and the target category and bounding box information are output. At the same time, the fused feature vector is segmented into precipitation areas and clouds based on the segmentation model to generate a segmentation result mask, wherein the extracted cloud top temperature, cloud optical thickness and cloud top height are used as auxiliary input parameters in the detection and segmentation process;

[0013] Step 6: The segmentation result mask and target detection result generated in step 5 are combined with the time series characteristics, and the time series modeling method is used to analyze the spatiotemporal dynamic characteristics of the meteorological system. Then, the fused feature vector and the hidden state of the previous time step are updated in a time recursive manner to generate the hidden state of the current time step.

[0014] Step 7: Based on the spatiotemporal dynamic characteristics analysis results in step 6, the future meteorological evolution is predicted through the time series prediction model, and then the fused feature vector of the historical time step is input into the time series prediction model to output the prediction results of multiple time steps in the future, including typhoon path, precipitation area and precipitation intensity.

[0015] Preferably, in step 2, the convolutional autoencoder performs denoising processing on the meteorological satellite data, including:

[0016] Feed the weather satellite data into the convolutional autoencoder model;

[0017] The convolutional autoencoder model includes an encoder and a decoder structure. The encoder extracts the high-dimensional features of meteorological data through multiple convolutional layers, and the decoder restores and reconstructs the high-dimensional features to generate denoised data. The denoising formula is:

[0018]

[0019] Among them, X is the input meteorological satellite data, is the denoised data, and θ is the parameter set of the convolutional autoencoder model.

[0020] Preferably, in step 3, when the time axis of the multimodal meteorological satellite data is aligned using a time interpolation method, the time interpolation formula is:

[0021]

[0022] Among them, X(t 1 ) and X(t 2 ) represents the time point t 1 and t 2 data,

[0023] X aligned (t) represents the alignment data at time T, t 2 -t 1 Indicates the time interval,

[0024] is the time scale factor.

[0025] Preferably, in step 4, the process of extracting features from infrared band data, visible light band data and multispectral band data using a multi-input convolutional neural network includes:

[0026] The feature extraction formula of infrared band data is:

[0027]

[0028] The feature extraction formula for visible light band data is:

[0029]

[0030] The feature extraction formula of multispectral band data is:

[0031]

[0032] Among them, X IR , X VIS , X MULTI For infrared, visible and multispectral data,

[0033] FIR 、F VIS 、F MULTI is the extracted feature vector,

[0034] represents the convolutional neural network model for infrared data,

[0035] represents the convolutional neural network model for visible light data,

[0036] Represents a convolutional neural network model for multimodal data.

[0037] Preferably, in step 4, the infrared band features, the visible light band features and the multi-spectral band features are interactively fused through a cross-modal self-attention mechanism, and the fused feature vector formula is:

[0038]

[0039] Among them, Q, K, V are the query matrix, key matrix and value matrix, d is the dimension scaling factor of the feature vector, and F fusion represents the fused feature matrix,

[0040] Softmax acts on the similarity score matrix to represent the importance weight of each feature.

[0041] QK T is the dot product of the query matrix and the key matrix.

[0042] Preferably, in step 5, the process of performing target detection of meteorological phenomena on the fused feature vector based on the target detection model includes:

[0043] The YOLO detection model is used to output the target category and bounding box. The input of the target detection is the fused feature vector. The target categories include the typhoon center and the severe convection area. The output of the bounding box is used to locate the detection area.

[0044] Preferably, in step 5, the process of segmenting the fused feature vector into precipitation areas and clouds based on the segmentation model includes:

[0045] The segmentation model is DeepLabv3+, and the segmentation result is optimized by the following segmentation loss function:

[0046] L=L seg +λL phy ,

[0047] Among them, L is the segmentation result, L seg is the cross entropy segmentation loss, L phy is the physical consistency loss, and λ is the weight factor of the physical consistency loss.

[0048] Preferably, in step 6, the process of performing spatiotemporal dynamic characteristic analysis in combination with time series characteristics includes:

[0049] The fused feature vector of the current time step and the hidden state of the previous time step are recursively updated through ConvLSTM. The formula is:

[0050] h t =f ConvLSTM (F fusion，t ,h t-1 ),

[0051] Among them, h t is the hidden state at time step t, h t-1 is the hidden state of the previous time step, F fusion，t is the fused feature vector at time step t, f ConvLSTM Represents the nonlinear transformation function of the convolutional long short-term memory network.

[0052] Preferably, in step 7, the process of predicting future meteorological evolution by using a time series prediction model includes:

[0053] The fused feature vector of the historical time step is input into the time series prediction model, and the prediction result of the future time step is output. The prediction formula is:

[0054]

[0055] in, represents the fused feature vector from historical time steps t to T,

[0056] is the prediction result at time step t+k,

[0057] f Transformer Represents a nonlinear mapping function based on Transformer.

[0058] Preferably, in step 7, in order to further optimize the computational efficiency of the time series prediction model, the fused feature vector of the historical time step is compressed, and the weight matrix of the time series prediction model is optimized by using pruning technology, and the pruning formula is:

[0059] W pruned =W⊙M mask ,

[0060] Among them, W represents the weight matrix of the time series prediction model, M mask represents the pruning mask matrix, W pruned represents the weight matrix after pruning, and ⊙ is the element-by-element multiplication operation.

[0061] The present invention provides a method for processing meteorological satellite data based on deep learning. It has the following beneficial effects:

[0062] 1. The present invention designs a cross-modal self-attention mechanism to interactively fuse the features of infrared, visible light and multispectral data, fully mine the complementary information between different data sources, realize the deep feature expression of multimodal meteorological satellite data, and solve the problem that a single data source is difficult to fully characterize complex meteorological phenomena. It significantly improves the model's ability to perceive complex meteorological phenomena, improves the collaborative analysis capability of multimodal data, and provides higher accuracy and robustness for meteorological forecasting and phenomenon detection.

[0063] 2. The present invention introduces important physical characteristics of cloud top temperature, cloud optical thickness and cloud top height into target detection and segmentation tasks as auxiliary input parameters, effectively combines physical knowledge and deep learning technology, achieves more accurate modeling of meteorological phenomena, and the prediction results are more in line with actual physical laws, improves the performance of key tasks such as detecting severe convective areas, segmenting precipitation areas and cloud layers, and significantly improves the accuracy and reliability of detection and segmentation, providing a scientific basis for the analysis of complex meteorological systems.

[0064] 3. The present invention adopts a convolutional long short-term memory network to recursively update the spatiotemporal dynamic characteristics of the meteorological system, establishes a deep learning model that integrates time series characteristics and spatial distribution characteristics, and realizes accurate modeling of the evolution trend of complex weather systems. It solves the problem that traditional static analysis methods cannot describe nonlinear dynamic changes, and has the effect of enhancing the system's ability to describe nonlinear dynamic changes, providing more powerful technical support for dynamic tracking and evolution analysis of weather systems.

[0065] 4. The present invention optimizes the model weight matrix based on the Transformer time series prediction model in combination with pruning technology, significantly reducing the model calculation complexity, achieving high-precision prediction of historical multi-time-step data to future multi-time-step meteorological evolution, improving prediction efficiency, and solving the problem of high computing resource requirements restricting real-time applications. The result is improved computing efficiency and applicable to real-time weather forecasting, so that it can operate efficiently in actual business and provide solid technical support for meteorological disaster warning and real-time prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0067] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0068] The present invention is described in detail below in conjunction with the accompanying drawings:

[0069] Example:

[0070] Please refer to the attached Figure 1 , an embodiment of the present invention provides a meteorological satellite data processing method based on deep learning, comprising:

[0071] Step 1: Acquire multimodal meteorological satellite data, which includes infrared band data, visible light band data, multispectral band data, auxiliary ground observation data and meteorological radar data;

[0072] Step 2: De-noise the meteorological satellite data through the convolutional autoencoder to generate denoised data, and then align the denoised data using the geo-reference method to unify the spatial resolution of different data;

[0073] Step 3: Use the time interpolation method to align the time axis of the multimodal meteorological satellite data to generate the multimodal meteorological satellite data after time and space alignment, and then extract the physical features in the multimodal meteorological satellite data after time and space alignment, the physical features including cloud top temperature, cloud optical thickness and cloud top height;

[0074] Step 4: extract features from the infrared band data, visible light band data, and multi-spectral band data in step 1, use a multi-input convolutional neural network to extract infrared band features, visible light band features, and multi-spectral band features, and obtain corresponding feature vectors. Then, interactively fuse the infrared band features, visible light band features, and multi-spectral band features through a cross-modal self-attention mechanism to generate a fused feature vector.

[0075] Step 5: For the fused feature vector generated in step 4, the target detection model is used to perform target detection of meteorological phenomena, and the target category and bounding box information are output. At the same time, the fused feature vector is segmented into precipitation areas and clouds based on the segmentation model to generate a segmentation result mask, wherein the extracted cloud top temperature, cloud optical thickness and cloud top height are used as auxiliary input parameters in the detection and segmentation process;

[0076] Step 6: The segmentation result mask and target detection result generated in step 5 are combined with the time series characteristics, and the time series modeling method is used to analyze the spatiotemporal dynamic characteristics of the meteorological system. Then, the fused feature vector and the hidden state of the previous time step are updated in a time recursive manner to generate the hidden state of the current time step.

[0077] Step 7: Based on the spatiotemporal dynamic characteristics analysis results in step 6, the future meteorological evolution is predicted through the time series prediction model, and then the fused feature vector of the historical time step is input into the time series prediction model to output the prediction results of multiple time steps in the future, including typhoon path, precipitation area and precipitation intensity.

[0078] The multimodal meteorological satellite data is denoised by convolutional autoencoders to significantly reduce noise interference. The spatial and temporal alignment is achieved by using georeferencing and time interpolation methods to unify the resolution and time axis of the data, providing a high-quality and consistent data foundation for subsequent analysis.

[0079] The infrared, visible light and multi-spectral band features are extracted through multi-input convolutional neural networks, and the features are interactively fused through the cross-modal self-attention mechanism to effectively mine the complementary information between different modalities and improve the understanding ability and modeling effect of complex meteorological phenomena;

[0080] By using the fused feature vectors, combined with the target detection model and the segmentation model, we can achieve target detection of meteorological phenomena, and at the same time, accurately segment precipitation areas and clouds, thus improving the efficiency and accuracy of multi-task processing.

[0081] By introducing key physical characteristics such as cloud top temperature, cloud optical thickness and cloud top height, the model's physical consistency description of meteorological phenomena is enhanced, making the detection, segmentation and prediction results scientific and reliable;

[0082] Through the time series modeling method and convolutional long short-term memory network, the spatiotemporal dynamic characteristics of the meteorological system can be accurately modeled, which can capture the nonlinear dynamic changes and evolution of the weather system and provide support for the analysis of complex meteorological phenomena.

[0083] Based on the time series prediction model, combined with the fusion feature vector of historical time steps, it can make high-precision predictions of future meteorological evolution, providing a reliable basis for weather forecasting and disaster warning;

[0084] By introducing pruning technology, the computational efficiency of the prediction model is optimized and the model complexity is reduced, so that the method can be used for real-time weather monitoring and forecasting, meeting the needs of meteorological services for rapid response and real-time processing.

[0085] This method is suitable for multi-modal, multi-time series, and high-dimensional meteorological data processing, and is widely used in typhoon prediction, precipitation monitoring, and meteorological disaster warning scenarios, providing technical support for meteorological research, environmental protection, and public safety.

[0086] In step 2, the convolutional autoencoder performs denoising on the meteorological satellite data including:

[0087] Feed the weather satellite data into the convolutional autoencoder model;

[0088] The convolutional autoencoder model includes an encoder and a decoder structure. The encoder extracts the high-dimensional features of meteorological data through multiple convolutional layers, and the decoder restores and reconstructs the high-dimensional features to generate denoised data. The denoising formula is:

[0089]

[0090] Among them, X is the input meteorological satellite data, is the denoised data, and θ is the parameter set of the convolutional autoencoder model.

[0091] The convolutional autoencoder extracts high-dimensional features of the data through the encoder, which can effectively capture the local structure and spatial characteristics of the data, retain key information during the denoising process, and significantly improve the quality of meteorological satellite data;

[0092] Without manual intervention, the convolutional autoencoder automatically extracts and restores high-dimensional features through multiple convolutional layers, adapts to complex meteorological data, reduces the dependence of traditional denoising methods on parameter adjustment, and improves the intelligence level of the denoising process;

[0093] The denoised data can maintain the physical and spatial integrity of the meteorological data while removing noise, ensuring that the processed data is more suitable for subsequent analysis and modeling;

[0094] The convolutional autoencoder learns the characteristic patterns of different noise types through extensive training, has good generalization ability, and is applied to meteorological satellite data of different types and resolutions, enhancing its applicability;

[0095] The denoised data is clearer and purer, providing input data for subsequent feature extraction, target detection and spatiotemporal modeling, which helps to improve the accuracy and robustness of subsequent meteorological analysis;

[0096] The convolutional autoencoder has high computational efficiency and can quickly complete the denoising task of large-scale meteorological satellite data, providing technical support for real-time meteorological monitoring and forecasting;

[0097] The model structure is flexible, and the network depth and parameters can be adjusted according to the specific application scenarios to adapt to the denoising needs of meteorological data of different complexities, providing a solid foundation for multimodal data fusion.

[0098] In step 3, when the time axis of the multi-modal meteorological satellite data is aligned using the time interpolation method, the time interpolation formula is:

[0099]

[0100] Among them, X(t 1 ) and X(t 2 ) represents the time point t 1 and t 2 data,

[0101] X aligned (t) represents the alignment data at time T, t 2 -t 1 Indicates the time interval,

[0102] is the time scale factor.

[0103] Through the time interpolation method, the multimodal data at different time intervals are aligned to a unified time axis, eliminating the inconsistency caused by sampling time differences, and providing a basis for subsequent multimodal data fusion and analysis;

[0104] The time interpolation method smoothes the original data to generate time-aligned continuous data, making full use of the existing multimodal data resources and avoiding data loss or neglect due to time axis mismatch;

[0105] Time interpolation provides the possibility of consistent modeling of data at different time intervals, making subsequent time series modeling and analysis more accurate and improving the ability to capture the dynamic evolution of meteorological phenomena;

[0106] Time interpolation synchronizes multimodal data at different time intervals, allowing multimodal features to be jointly analyzed at the same time point, optimizing the basic conditions for feature interaction and providing accurate input for deep learning models;

[0107] The flexibility of the interpolation method allows adaptation to scenarios with different temporal resolutions and data sampling densities, making it suitable for a wide range of meteorological data processing needs;

[0108] During the interpolation process, the method can smooth out abnormal fluctuations caused by insufficient sampling or changes in time intervals, making the data on the time axis more continuous and reliable.

[0109] In summary, the time axis of multimodal meteorological satellite data can be aligned using the time interpolation method to improve the consistency and completeness of the data, providing a solid data foundation for deep feature extraction, modeling and prediction of multimodal meteorological data.

[0110] In step 4, the process of extracting features from infrared band data, visible light band data and multispectral band data using a multi-input convolutional neural network includes:

[0111] The feature extraction formula of infrared band data is:

[0112]

[0113] The feature extraction formula for visible light band data is:

[0114]

[0115] The feature extraction formula of multispectral band data is:

[0116]

[0117] Among them, X IR , X VIS , X MULTI For infrared, visible and multispectral data,

[0118] F IR 、F VIS 、F MULTI is the extracted feature vector,

[0119] represents the convolutional neural network model for infrared data,

[0120] represents the convolutional neural network model for visible light data,

[0121] Represents a convolutional neural network model for multimodal data.

[0122] By designing specialized convolutional neural network models for infrared, visible light, and multispectral data, features are extracted based on the characteristics of various data types, the key information of each modality data is retained, and the accuracy of feature extraction is improved;

[0123] Different modal data carry complementary information, such as infrared bands reflecting temperature characteristics, visible light bands reflecting surface characteristics, and multi-spectral bands providing more detailed spectral information. Multi-input convolutional neural networks can extract unique features of data and provide high-quality input for subsequent feature fusion;

[0124] The multi-input structure avoids the feature loss problem of a single model in processing multi-modal data. The independent extraction and optimization of feature vectors of each modality can accurately express the physical properties and spatial characteristics of different data.

[0125] The multi-input convolutional neural network allows the use of different convolution structures and parameter configurations for data of each modality, so that the model can flexibly adapt to the resolution, scale and physical characteristics of data in different bands, enhancing the generalization ability of the model;

[0126] By extracting feature vectors from different bands, the diversity and independence between modalities are preserved, laying a solid foundation for the subsequent cross-modal self-attention mechanism to achieve efficient fusion;

[0127] The multi-input network structure helps to reduce the information redundancy and noise interference caused by the direct mixing of multimodal data, making feature extraction more targeted and improving the efficiency and effectiveness of subsequent processing tasks;

[0128] Multi-input convolutional neural networks allow the extraction of features from each modality to be performed in parallel, greatly improving computational efficiency and providing technical support for real-time processing of large-scale meteorological satellite data.

[0129] By extracting and optimizing the features of multimodal data, complex meteorological phenomena can be fully and accurately characterized, providing a high-quality feature foundation for subsequent detection, segmentation and prediction tasks.

[0130] In summary, the multi-input convolutional neural network provides an efficient and accurate solution for the comprehensive analysis of complex meteorological systems by extracting multimodal data features in a targeted manner, taking into account the characteristics and complementarity of data of different modalities.

[0131] In step 4, the infrared band features, visible light band features and multi-spectral band features are interactively fused through the cross-modal self-attention mechanism. The fused feature vector formula is:

[0132]

[0133] Among them, Q, K, V are the query matrix, key matrix and value matrix, d is the dimension scaling factor of the feature vector, and F fusion represents the fused feature matrix,

[0134] Softmax acts on the similarity score matrix to represent the importance weight of each feature.

[0135] QK T is the dot product of the query matrix and the key matrix.

[0136] The cross-modal self-attention mechanism captures the complementary relationship between infrared, visible light and multispectral data by calculating the similarity scores between different modal features, ensuring that the information interaction between modalities is more in-depth and efficient, and improving the comprehensive expression ability of fusion features;

[0137] By generating a similarity score matrix and calculating the importance weights of features, the cross-modal self-attention mechanism can dynamically adjust the weights of different modal features in the fusion feature according to their contribution, thus avoiding the problem of information imbalance between modalities and making the fusion result more representative.

[0138] Infrared, visible light and multispectral data have different physical properties and observation advantages. The cross-modal self-attention mechanism can focus on extracting the most useful information of different modalities for complex meteorological phenomena during the fusion process, thereby improving the model's ability to understand meteorological phenomena.

[0139] Through the dot product operation of the query matrix, key matrix and value matrix, the mechanism can effectively remove redundant information and noise in the features of different modalities, making the fused features more refined and efficient, which helps to improve the performance of subsequent tasks;

[0140] The cross-modal self-attention mechanism unifies modal features into a fused feature vector, allowing data from different modalities to be represented in a unified feature space, providing consistent input for subsequent target detection, segmentation, and prediction tasks;

[0141] The synergy between modalities is enhanced through interactive fusion. If the data quality of a certain modality is poor or incomplete, the mechanism can rely on the characteristics of other modalities to compensate, thereby enhancing the robustness of the overall model in an uncertain environment.

[0142] The present invention makes full use of the complementarity and correlation of multimodal data, improves the depth and efficiency of feature fusion, and lays a solid foundation for the comprehensive analysis and accurate prediction of complex meteorological phenomena.

[0143] In step 5, the process of performing target detection of meteorological phenomena on the fused feature vector based on the target detection model includes:

[0144] The YOLO detection model is used to output the target category and bounding box. The input of the target detection is the fused feature vector. The target categories include the typhoon center and the severe convection area. The output of the bounding box is used to locate the detection area.

[0145] In step 5, the process of segmenting the precipitation area and cloud layer of the fused feature vector based on the segmentation model includes:

[0146] The segmentation model is DeepLabv3+, and the segmentation result is optimized by the following segmentation loss function:

[0147] L=L seg +λL phy ,

[0148] Among them, L is the segmentation result, L seg is the cross entropy segmentation loss, L phy is the physical consistency loss, and λ is the weight factor of the physical consistency loss.

[0149] The combination of target detection and segmentation models enables unified analysis of meteorological phenomenon classification, positioning and regional segmentation, and improves data utilization efficiency.

[0150] Target detection locates severe convection areas and typhoon centers, and the segmentation model provides information on dynamic changes in precipitation and clouds. The two work together to provide all-round support for monitoring the evolution of meteorological systems.

[0151] By combining multimodal features and physical consistency loss, the adaptability and robustness of the detection and segmentation model to different weather systems are further enhanced, making it suitable for a variety of meteorological scenarios.

[0152] Both the target detection and segmentation models use mature optimization techniques, which can quickly iterate and optimize during the training process to further improve prediction accuracy.

[0153] Through the above method, the present invention realizes accurate detection, segmentation and dynamic monitoring of complex meteorological phenomena, and provides technical support for meteorological warning and prediction.

[0154] In step 6, the process of analyzing spatiotemporal dynamic characteristics in combination with time series features includes:

[0155] The fused feature vector of the current time step and the hidden state of the previous time step are recursively updated through ConvLSTM. The formula is:

[0156] h t =f ConvLSTM (F fusion，t ,h t-1 ),

[0157] Among them, h t is the hidden state at time step t, h t-1 is the hidden state of the previous time step, F fusion，t is the fused feature vector at time step t, f ConvLSTM Represents the nonlinear transformation function of the convolutional long short-term memory network.

[0158] ConvLSTM effectively integrates the dynamic characteristics and spatial features in the time series by recursively updating the hidden state of the current time step. It can capture the temporal and spatial evolution of meteorological phenomena and improve the ability to describe complex weather systems.

[0159] By combining the fused feature vector of the current time step and the hidden state of the previous time step, ConvLSTM can handle the highly nonlinear dynamic changes of the meteorological system and adapt to the needs of complex weather phenomenon modeling.

[0160] ConvLSTM has memory capabilities and can effectively retain and utilize the information of historical time steps in the recursive process, providing support for feature updates in the current time step, and is suitable for capturing long-term meteorological change trends.

[0161] The recursive update mechanism combines local and global information, enabling the model to identify the dynamic changes of short-term meteorological phenomena and track the overall trend of long-term evolution.

[0162] ConvLSTM uses convolution operations to retain the spatial structure information of features. At the same time, it dynamically updates the hidden state through a recursive mechanism to avoid the feature loss problem in traditional time series methods.

[0163] Combining the above advantages, ConvLSTM is used to analyze spatiotemporal dynamic characteristics, providing strong technical support for modeling complex spatiotemporal changes in meteorological systems and improving the monitoring, analysis and prediction capabilities of meteorological phenomena.

[0164] In step 7, the process of predicting future meteorological evolution through the time series prediction model includes:

[0165] The fused feature vector of the historical time step is input into the time series prediction model, and the prediction result of the future time step is output. The prediction formula is:

[0166]

[0167] in, represents the fused feature vector from historical time steps t to T,

[0168] is the prediction result at time step t+k,

[0169] f Transformer Represents a nonlinear mapping function based on Transformer.

[0170] In step 7, in order to further optimize the computational efficiency of the time series prediction model, the fused feature vector of the historical time step is compressed, and the pruning technology is used to optimize the weight matrix of the time series prediction model. The pruning formula is:

[0171] W pruned =W⊙M mask ,

[0172] Among them, W represents the weight matrix of the time series prediction model, M mask represents the pruning mask matrix, W pruned represents the weight matrix after pruning, and ⊙ is the element-by-element multiplication operation.

[0173] Through the time series prediction model, high-precision prediction of future meteorological evolution can be achieved. At the same time, pruning technology is used to significantly improve computing efficiency, taking into account both performance and real-time requirements.

[0174] The model can capture the dynamic characteristics in long time series, predict the future development trend of complex meteorological systems, and provide a scientific basis for meteorological disaster warning.

[0175] The combination of high-precision prediction and efficient computing enables the model to be widely used in actual meteorological business and is suitable for a variety of scenarios such as real-time weather forecasting, disaster monitoring and meteorological data analysis.

[0176] Through the above method, the present invention improves the prediction ability of future meteorological evolution, reduces the computational burden of the model through pruning technology, and provides strong technical support for real-time and efficient meteorological services.

[0177] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A meteorological satellite data processing method based on deep learning, characterized in that: include: Step 1: Acquire multimodal meteorological satellite data, which includes infrared band data, visible light band data, multispectral band data, auxiliary ground observation data and meteorological radar data; Step 2: De-noise the meteorological satellite data through the convolutional autoencoder to generate denoised data, and then align the denoised data using the geo-reference method to unify the spatial resolution of different data; Step 3: Use the time interpolation method to align the time axis of the multimodal meteorological satellite data to generate the multimodal meteorological satellite data after time and space alignment, and then extract the physical features in the multimodal meteorological satellite data after time and space alignment, the physical features including cloud top temperature, cloud optical thickness and cloud top height; Step 4: extract features from the infrared band data, visible light band data, and multi-spectral band data in step 1, use a multi-input convolutional neural network to extract infrared band features, visible light band features, and multi-spectral band features, and obtain corresponding feature vectors. Then, interactively fuse the infrared band features, visible light band features, and multi-spectral band features through a cross-modal self-attention mechanism to generate a fused feature vector. Step 5: For the fused feature vector generated in step 4, the target detection model is used to perform target detection of meteorological phenomena, and the target category and bounding box information are output. At the same time, the fused feature vector is segmented into precipitation areas and clouds based on the segmentation model to generate a segmentation result mask, wherein the extracted cloud top temperature, cloud optical thickness and cloud top height are used as auxiliary input parameters in the detection and segmentation process; Step 6: The segmentation result mask and target detection result generated in step 5 are combined with the time series characteristics, and the time series modeling method is used to analyze the spatiotemporal dynamic characteristics of the meteorological system. Then, the fused feature vector and the hidden state of the previous time step are updated in a time recursive manner to generate the hidden state of the current time step. Step 7: Based on the spatiotemporal dynamic characteristics analysis results in step 6, the future meteorological evolution is predicted through the time series prediction model, and then the fused feature vector of the historical time step is input into the time series prediction model to output the prediction results of multiple time steps in the future, including typhoon path, precipitation area and precipitation intensity.

2. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 2, the convolutional autoencoder performs denoising processing on the meteorological satellite data, including: Feed the weather satellite data into the convolutional autoencoder model; The convolutional autoencoder model includes an encoder and a decoder structure. The encoder extracts the high-dimensional features of meteorological data through multiple convolutional layers, and the decoder restores and reconstructs the high-dimensional features to generate denoised data. The denoising formula is: Among them, X is the input meteorological satellite data, is the denoised data, and θ is the parameter set of the convolutional autoencoder model.

3. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 3, when the time axis of the multi-modal meteorological satellite data is aligned using the time interpolation method, the time interpolation formula is: Among them, X(t1) and X(t2) represent the data at time points t1 and t2, X aligned (t) represents the alignment data at time T, t2-t1 represents the time interval, is the time scale factor.

4. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 4, the process of extracting features from infrared band data, visible light band data and multispectral band data using a multi-input convolutional neural network includes: The feature extraction formula of infrared band data is: The feature extraction formula for visible light band data is: The feature extraction formula of multispectral band data is: Among them, X IR , X VIS , X MULTI For infrared, visible and multispectral data, F IR 、F VIS 、F MULTI is the extracted feature vector, represents the convolutional neural network model for infrared data, represents the convolutional neural network model for visible light data, Represents a convolutional neural network model for multimodal data.

5. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 4, the infrared band features, visible light band features and multi-spectral band features are interactively fused through the cross-modal self-attention mechanism, and the fused feature vector formula is: Among them, Q, K, V are the query matrix, key matrix and value matrix, d is the dimension scaling factor of the feature vector, and F fusion represents the fused feature matrix, Softmax acts on the similarity score matrix to represent the importance weight of each feature. QK T is the dot product of the query matrix and the key matrix.

6. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 5, the process of performing target detection of meteorological phenomena on the fused feature vector based on the target detection model includes: The YOLO detection model is used to output the target category and bounding box. The input of the target detection is the fused feature vector. The target categories include the typhoon center and the severe convection area. The output of the bounding box is used to locate the detection area.

7. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 5, the process of segmenting the fused feature vector into precipitation areas and clouds based on the segmentation model includes: The segmentation model is DeepLabv3+, and the segmentation result is optimized by the following segmentation loss function: L=L seg +λL phy , Among them, L is the segmentation result, L seg is the cross entropy segmentation loss, L phy is the physical consistency loss, and λ is the weight factor of the physical consistency loss.

8. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 6, the process of analyzing spatiotemporal dynamic characteristics in combination with time series characteristics includes: The fused feature vector of the current time step and the hidden state of the previous time step are recursively updated through ConvLSTM. The formula is: h t =f ConvLSTM (F fusion，t ,h t-1 ), Among them, h t is the hidden state at time step t, h t-1 is the hidden state of the previous time step, F fusion，t is the fused feature vector at time step t, f ConvLSTM Represents the nonlinear transformation function of the convolutional long short-term memory network.

9. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 7, the process of predicting future meteorological evolution by using the time series prediction model includes: The fused feature vector of the historical time step is input into the time series prediction model, and the prediction result of the future time step is output. The prediction formula is: in, represents the fused feature vector from historical time steps t to T, is the prediction result at time step t+k, f Transformer Represents a nonlinear mapping function based on Transformer.

10. The method for processing meteorological satellite data based on deep learning according to claim 1, characterized in that: In step 7, in order to further optimize the computational efficiency of the time series prediction model, the fused feature vector of the historical time step is compressed, and the weight matrix of the time series prediction model is optimized by using pruning technology. The pruning formula is: IN pruned =W⊙M mask , Among them, W represents the weight matrix of the time series prediction model, M mask represents the pruning mask matrix, W pruned represents the weight matrix after pruning, and ⊙ is the element-by-element multiplication operation.

Images