A Method and Equipment for Reconstructing Three-Dimensional Temperature and Humidity Fields of Tropical Cyclones Based on Deep Learning and Sparse Observations
By fusing radiosonde and satellite remote sensing data using deep learning methods, and utilizing an encoder-decoder architecture and a gated cyclic fusion module, the problem of reconstructing the three-dimensional temperature and humidity field of tropical cyclones under sparse observations was solved, achieving high-precision and physically reasonable reconstruction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SECOND INST OF OCEANOGRAPHY MNR
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies struggle to effectively utilize sparse, high-precision observation data and large-scale remote sensing information to reconstruct a physically consistent and structurally reasonable three-dimensional temperature and humidity field within tropical cyclones. Traditional interpolation methods cannot generate complex structures that conform to atmospheric physical laws, and traditional assimilation methods have limited effectiveness in sparse regions.
A deep learning-based approach is adopted, utilizing radiosonde observation data and satellite remote sensing data. Data fusion and completion are performed through a deep neural network model to generate a complete three-dimensional temperature and humidity field. An encoder-decoder architecture and a gated loop fusion module are used for information propagation and refinement, and adversarial training is combined to improve reconstruction accuracy.
It achieves high-precision three-dimensional temperature and humidity field reconstruction, generating a more refined and physically consistent structure with high computational efficiency, and has the potential for quasi-commercial applications.
Smart Images
Figure CN121616767B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of meteorological detection and data processing technology, and relates to a method and device for reconstructing the three-dimensional temperature and humidity field of tropical cyclones, and more particularly to a method and device for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observation. Background Technology
[0002] The detailed three-dimensional temperature and humidity structure inside tropical cyclones (typhoons) is crucial for predicting their intensity and track. Currently, the main technical methods for obtaining this information are as follows:
[0003] 1. Aircraft Dropsonde: It can provide high-precision, high-vertical-resolution profiles of parameters such as temperature and humidity along the drop trajectory and is considered a "true value" observation.
[0004] 2. Satellite remote sensing: It can provide cloud images and cloud top brightness temperature products with large-scale and continuous coverage, but these data are difficult to invert the three-dimensional temperature and humidity structure inside the cyclone.
[0005] 3. Numerical weather prediction models: These models simulate atmospheric conditions using physical equations and parameterization schemes, providing a complete three-dimensional field. However, their ability to characterize the fine structure of the core regions of extreme weather systems is limited, and they are affected by errors in the initial field and boundary conditions.
[0006] 4. Traditional data assimilation methods, such as optimal interpolation and three-dimensional / four-dimensional variational assimilation, aim to integrate various observation data into numerical models. These methods require a large amount of computational resources and are extremely costly.
[0007] Despite the existence of the above-mentioned technical means, the following core shortcomings still exist in practical applications:
[0008] 1. Extreme spatial sparsity of observational data: Drop-in radiosonde observations are like "point sampling" on a vast ocean, unable to directly provide a continuous three-dimensional structure. How to perform scientific and reasonable interpolation between these sparse observation points is the core challenge facing current technology.
[0009] 2. Difficulty in multi-source data fusion: Satellite remote sensing data and radiosonde data differ greatly in physical meaning and spatiotemporal resolution. Traditional methods lack effective mechanisms to deeply integrate the advantages of these two types of data, making it difficult to achieve complementarity.
[0010] 3. Limitations of traditional interpolation and assimilation methods: Simple spatial interpolation methods (such as inverse distance weighting) cannot generate complex structures that conform to atmospheric physical laws (such as typhoon warm cores and eyewalls). Furthermore, variational assimilation methods, in regions with extremely sparse observational data, rely heavily on the quality of the background field. If the background field itself has significant errors, the assimilation effect is limited, making it difficult to effectively recover the true structural features caused by small- to medium-scale processes.
[0011] Therefore, there is an urgent need for a method that can effectively utilize sparse high-precision observations and combine them with large-scale remote sensing information to intelligently "fill" data gaps in order to reconstruct a physically consistent and structurally reasonable three-dimensional temperature and humidity field. Summary of the Invention
[0012] To address the aforementioned issues, this invention proposes a method and apparatus for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations.
[0013] The technical solution adopted in this invention is as follows:
[0014] A method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations includes the following steps:
[0015] Acquire radiosonde observation data and satellite remote sensing data for the target tropical cyclone region;
[0016] The radiosonde observation data is mapped onto a two-dimensional grid on multiple preset pressure layers to generate an initial temperature and humidity field that includes regions with missing data.
[0017] Based on the initial temperature and humidity field, a mask is generated to identify the locations where data is missing;
[0018] The satellite remote sensing data is interpolated and aligned to a two-dimensional grid that is spatially aligned with the initial temperature and humidity field to form an auxiliary remote sensing field;
[0019] For each of the aforementioned pressure layers, the initial temperature and humidity field, mask, and auxiliary remote sensing field of the current layer are combined to form a multi-channel input tensor, which is then input into a pre-trained deep neural network model.
[0020] The deep neural network model infers and fills in the missing data regions based on the multi-channel input tensor, and outputs the complete two-dimensional temperature and humidity field of the current layer.
[0021] The complete two-dimensional temperature and humidity fields corresponding to all pressure layers are vertically synthesized to obtain the three-dimensional temperature and humidity field of the target tropical cyclone.
[0022] Furthermore, the radiosonde observation data is from a drop-in radiosonde.
[0023] Furthermore, the preset multiple pressure layers are standard isobaric surfaces.
[0024] Furthermore, the radiosonde observation data is mapped onto a two-dimensional grid on multiple preset pressure layers, specifically including:
[0025] At each pressure layer, the longitude, latitude, temperature, and humidity values of the target tropical cyclone are extracted at discrete points based on radiosonde observation data. The data of the discrete points are then mapped onto a regular two-dimensional grid covering the region of the target tropical cyclone to form an initial temperature and humidity field including areas with missing data.
[0026] Furthermore, based on the initial temperature and humidity field, a mask is generated to identify the locations of missing data, specifically including:
[0027] Based on the spatial distribution of data points in the initial temperature and humidity field, a binary mask with the same size as the two-dimensional grid is generated, wherein the mask value corresponding to the known data location is the first value, and the mask value corresponding to the missing data location is the second value.
[0028] Furthermore, the deep neural network model is a deep convolutional network based on an encoder-decoder architecture, and a gated recurrent fusion module is provided between the encoder and the decoder; the encoder is used to downsample the multi-channel input tensor to extract abstract feature maps; the gated recurrent fusion module is used to fuse and complete the abstract feature maps; the decoder is used to upsample the completed feature maps to reconstruct a complete two-dimensional temperature and humidity field.
[0029] Furthermore, the gated recurrent fusion module includes a gated convolutional unit and a recurrent refinement unit connected in sequence. The gated convolutional unit processes the abstract feature map output by the encoder, and distinguishes and regulates the feature flow from known data regions and missing data regions by generating adaptive gate values for the feature map. The recurrent refinement unit adopts a recurrent neural network structure to iteratively update the feature map processed by the gated convolutional unit multiple times, so that information propagates and merges from the known region to the inside of the missing data region, thereby refining and completing the feature representation of the missing data region.
[0030] Furthermore, the deep neural network model is obtained through adversarial training; the adversarial training adopts a generator-discriminator framework, wherein the deep neural network model serves as the generator and a discriminator network is provided; the discriminator is used to distinguish between the complete two-dimensional temperature and humidity field output by the generator and the real complete two-dimensional temperature and humidity field, and outputs the discrimination result;
[0031] The training objective of the generator is constrained by a composite loss function, which includes an adversarial loss based on the discrimination result and a pixel-level reconstruction loss between the generator output field and the real field.
[0032] A computer device, the computer device comprising:
[0033] One or more processors;
[0034] Memory, used to store one or more programs;
[0035] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations.
[0036] A computer-readable storage medium storing computer instructions that, when executed by one or more processors, cause the one or more processors to perform the steps in the method described above.
[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0038] 1. High reconstruction accuracy and reasonable physical structure: This invention utilizes the powerful nonlinear modeling capabilities of deep learning models to learn prior knowledge of the complex physical structure inside tropical cyclones. By fusing high-precision radiosonde data as "anchor points," the reconstructed temperature and humidity field is not only accurate at known points, but the structure of its filled regions (such as the warm core, eyewall, and spiral rainbands) is also more refined and physically consistent, outperforming traditional interpolation methods.
[0039] 2. Strong data fusion capability: This invention unifies sparse radiosonde data and satellite remote sensing data from vastly different sources and with distinct characteristics under a single deep learning framework, achieving effective information fusion. Satellite data provides valuable, large-scale background constraints for the model, significantly improving the reliability of reconstruction in vast data gaps.
[0040] 3. High automation and computational efficiency: Compared with traditional assimilation methods that require complex background field error settings, this invention only involves forward computation of the neural network in the application stage (inference), which is fast and has low computational cost. The entire reconstruction process can be automated end-to-end, and has the potential for quasi-business applications. Attached Figure Description
[0041] Figure 1 This is a flowchart of the method in an embodiment of the present invention.
[0042] Figure 2 This is a structural diagram of a deep neural network model in an embodiment of the present invention.
[0043] Figure 3This is a schematic diagram comparing sparse observation samples constructed based on HWRF model data with the actual temperature and humidity field in an embodiment of the present invention. Detailed Implementation
[0044] The technical solution of the present invention will be further explained clearly and in detail below with reference to the accompanying drawings and specific examples.
[0045] A method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations includes the following steps:
[0046] Acquire radiosonde observation data and satellite remote sensing data for the target tropical cyclone region;
[0047] The radiosonde observation data is mapped onto a two-dimensional grid on multiple preset pressure layers to generate an initial temperature and humidity field that includes regions with missing data.
[0048] Based on the initial temperature and humidity field, a mask is generated to identify the locations where data is missing;
[0049] The satellite remote sensing data is interpolated and aligned to a two-dimensional grid that is spatially aligned with the initial temperature and humidity field to form an auxiliary remote sensing field;
[0050] For each of the aforementioned pressure layers, the initial temperature and humidity field, mask, and auxiliary remote sensing field of the current layer are combined to form a multi-channel input tensor, which is then input into a pre-trained deep neural network model.
[0051] The deep neural network model infers and fills in the missing data regions based on the multi-channel input tensor, and outputs the complete two-dimensional temperature and humidity field of the current layer.
[0052] The complete two-dimensional temperature and humidity fields corresponding to all pressure layers are vertically synthesized to obtain the three-dimensional temperature and humidity field of the target tropical cyclone.
[0053] Furthermore, the radiosonde observation data is from a drop-in radiosonde.
[0054] Furthermore, the preset multiple pressure layers are standard isobaric surfaces.
[0055] Furthermore, the radiosonde observation data is mapped onto a two-dimensional grid on multiple preset pressure layers, specifically including:
[0056] At each pressure layer, the longitude, latitude, temperature, and humidity values of the target tropical cyclone are extracted at discrete points based on radiosonde observation data. The data of the discrete points are then mapped onto a regular two-dimensional grid covering the region of the target tropical cyclone to form an initial temperature and humidity field including areas with missing data.
[0057] Furthermore, based on the initial temperature and humidity field, a mask is generated to identify the locations of missing data, specifically including:
[0058] Based on the spatial distribution of data points in the initial temperature and humidity field, a binary mask with the same size as the two-dimensional grid is generated, wherein the mask value corresponding to the known data location is the first value, and the mask value corresponding to the missing data location is the second value.
[0059] Furthermore, the deep neural network model is a deep convolutional network based on an encoder-decoder architecture, and a gated recurrent fusion module is provided between the encoder and the decoder; the encoder is used to downsample the multi-channel input tensor to extract abstract feature maps; the gated recurrent fusion module is used to fuse and complete the abstract feature maps; the decoder is used to upsample the completed feature maps to reconstruct a complete two-dimensional temperature and humidity field.
[0060] Furthermore, the gated recurrent fusion module includes a gated convolutional unit and a recurrent refinement unit connected in sequence. The gated convolutional unit processes the abstract feature map output by the encoder, and distinguishes and regulates the feature flow from known data regions and missing data regions by generating adaptive gate values for the feature map. The recurrent refinement unit adopts a recurrent neural network structure to iteratively update the feature map processed by the gated convolutional unit multiple times, so that information propagates and merges from the known region to the inside of the missing data region, thereby refining and completing the feature representation of the missing data region.
[0061] Furthermore, the deep neural network model is obtained through adversarial training; the adversarial training adopts a generator-discriminator framework, wherein the deep neural network model serves as the generator and a discriminator network is provided; the discriminator is used to distinguish between the complete two-dimensional temperature and humidity field output by the generator and the real complete two-dimensional temperature and humidity field, and outputs the discrimination result;
[0062] The training objective of the generator is constrained by a composite loss function, which includes an adversarial loss based on the discrimination result and a pixel-level reconstruction loss between the generator output field and the real field.
[0063] Example
[0064] like Figure 1 As shown, in one specific embodiment of the present invention, a method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observation is provided. Its core is to use a deep neural network model trained adversarially to transform the problem of reconstructing the temperature and humidity field into a computer image inpainting problem.
[0065] This method mainly includes three stages: data preprocessing, model layer-by-layer repair, and three-dimensional field synthesis.
[0066] Step 1: Multi-source data acquisition and preprocessing
[0067] This step is intended to prepare input data for subsequent model processing.
[0068] 1.1 Acquiring and Gridding Sparse Radiosonde Data: Collect data from all drop-sondes used in the target tropical cyclone observation mission, and extract their longitude, latitude, temperature, and humidity values on preset standard isobaric surfaces (e.g., 1000 hPa, 925 hPa, 850 hPa...). Map these discrete points to a regular two-dimensional grid covering the entire tropical cyclone region through preliminary spatial matching, forming an initial temperature and humidity field with large data voids.
[0069] 1.2 Mask Generation: Based on the spatial distribution of the radiosonde data points, a binary mask of the same size as the grid is generated. Within the grid, locations with valid radiosonde data have a mask value of 0 (data known areas); other data gap areas have a mask value of 1 (data missing areas).
[0070] 1.3 Acquire and align auxiliary remote sensing data: Acquire satellite remote sensing products (such as cloud top brightness temperature) that are closest to the radiosonde observation time, and interpolate them onto the same two-dimensional grid as the initial temperature and humidity field to obtain the auxiliary remote sensing field as physical background information.
[0071] Step 2: Layer-by-layer repair based on deep generative networks
[0072] This step is the core of the invention: the isobaric surface is processed cyclically on a computer processor.
[0073] 2.1 Constructing the input tensor: For the isobaric surface being processed, multiple two-dimensional data fields, such as the temperature field, humidity field, satellite remote sensing field, and mask field with "holes", are stacked along the channel dimension to form a multi-channel input tensor.
[0074] 2.2 Model Inference: The multi-channel input tensor is fed into a pre-trained deep neural network model. The deep neural network model uses its learned prior knowledge of the internal structure of tropical cyclones, combined with known information in the input tensor (sonographic data and satellite remote sensing background), to infer and fill in unknown regions with a mask value of 1.
[0075] 2.3 Output a complete single-layer field: The deep neural network model outputs multiple complete two-dimensional temperature and humidity fields with the same size as the input and without voids.
[0076] Step 3: 3D Field Synthesis and Output
[0077] 3.1 Stacking and Synthesis: The complete two-dimensional temperature and humidity fields obtained from repairing all isobaric surfaces are stacked in a vertical order.
[0078] 3.2 Output Results: The final result is a complete, continuous, and high-resolution three-dimensional temperature and humidity field data cube of tropical cyclones, which is stored in a standard format (such as NetCDF) for subsequent analysis and application.
[0079] In this embodiment, the architecture and training method of the deep neural network model are as follows:
[0080] The deep neural network model used in this invention is not a simple stack of convolutions, but a deep convolutional network that employs an encoder-decoder structure and embeds a gated recurrent fusion module (e.g., Figure 2 (As shown).
[0081] The encoder consists of a series of strided convolutional layers. It receives multi-channel input tensors and extracts multi-scale feature maps from fine to coarse by downsampling layer by layer, effectively capturing hierarchical information of tropical cyclone structure at different spatial scales.
[0082] Gated Cyclic Fusion Module: This is the core innovative module of the network, located at the end (bottleneck) of the encoder. It receives the most abstract features extracted by the encoder and performs the following operations:
[0083] (1) Gated Convolution: This module does not use standard convolution internally, but instead uses gated convolution. Gated convolution learns an additional "gating value" for each feature channel, which can dynamically adjust the degree to which features are passed through. This allows the network to automatically distinguish between valid pixels (from regions where the data is known) and invalid pixels (from regions where the data is missing) in the feature map, and to take different processing strategies for them to prevent the spread of contaminating information from regions where the data is missing.
[0084] (2) Iterative Refinement Mechanism: This module contains a structure similar to a Recurrent Neural Network (RNN), such as a Convolutional Long Short-Term Memory (ConvLSTM) unit. It iteratively updates the features processed by gated convolution multiple times. In each iteration, information propagates and permeates from the known region boundaries of the feature map to the center of the missing data region, gradually refining and filling the feature representation of the missing data region. This design simulates the process of information diffusion, making the final filling result more globally coordinated and physically consistent.
[0085] The decoder consists of a series of transposed convolutional layers, connected to the outputs of corresponding encoder layers via skip connections. The decoder is responsible for progressively upsampling the abstract features refined by the gated recurrent fusion module, restoring them to the original resolution. Skip connections directly pass high-frequency detail information from the encoder to the decoder, helping the network reconstruct sharp edges and fine structural textures.
[0086] In the preferred configuration of this embodiment, taking the processing of 256×256 resolution grid data as an example, the specific parameter settings of the deep neural network model are as follows:
[0087] The encoder consists of four stages: the first stage is used for shallow feature extraction, which uses a 5×5 gated convolution with 48 output channels, keeping the feature map size unchanged; the following three stages are downsampling layers, which all use a 3×3 gated convolution with a stride of 2, and the number of output channels is doubled to 96, 192 and 384 respectively, and the feature map size is halved layer by layer, finally outputting a 32×32 deep abstract feature.
[0088] The gated cyclic fusion module is located at the end of the encoder. Its core ConvLSTM unit uses a 3×3 convolution kernel, and the number of channels remains at 384. This module is set to perform 8 iterations T=8, that is, to perform 8 iterations of inference in the low-resolution feature space of 32×32×32, in order to gradually repair feature holes.
[0089] The decoder consists of three upsampling stages, symmetrical to the encoder's downsampling stages. Each stage first enlarges the feature map size by a factor of 2 using bilinear interpolation, then reduces the number of channels using 3×3 gated convolutions (to 192, 96, and 48 respectively). To preserve high-frequency details, each stage of the decoder concatenates channels with the corresponding resolution feature layers of the encoder via skip connections.
[0090] Finally, the output layer uses a 3×3 ordinary convolution to map the number of channels to 2 (corresponding to the temperature field and humidity field respectively), and outputs the final reconstruction result through the Tanh activation function.
[0091] In this embodiment, the capability of the deep neural network model is obtained by adversarially training the deep neural network model as a generator G with a discriminator network D.
[0092] Training Process: During training, the discriminator D's task is to distinguish between the "real" temperature and humidity field and the "fake" temperature and humidity field generated by the generator G. The generator G's goal is to generate a temperature and humidity field that is as realistic as possible to "deceive" the discriminator D. The discriminator D distinguishes between the complete two-dimensional temperature and humidity field output by the generator G and the real complete two-dimensional temperature and humidity field, and outputs the discrimination result.
[0093] The training of generator G is driven not only by an adversarial loss based on the discriminator D's discrimination results, but also by a reconstruction loss. The reconstruction loss directly calculates the pixel-level difference (such as L1 distance) between the generator's output field and the real field, ensuring the numerical fidelity of the reconstruction results. The combination of these two factors drives the generator to generate temperature and humidity fields that are more structurally realistic and conform more closely to physical laws, while ensuring the accuracy of physical numerical values.
[0094] Figure 3 This document presents an example of data samples used for model training and testing in an embodiment of the present invention. This embodiment uses high-resolution reanalysis data from the Hurricane Weather Research and Forecasting Model (HWRF) as the base dataset to construct a "sparse observation-complete ground truth" data pair. The figure contains four subplots: The bottom two subplots (temperature / humidity) show the true complete temperature field (bottom left) and true complete humidity field (bottom right) at a specific pressure layer at a given moment in the HWRF model output. These high-resolution grid data contain the fine physical structure within tropical cyclones (such as the warm core structure of the typhoon eye and the distribution of the outer spiral rainbands), serving as "ground truth" labels during training. The black "×" marks in the figure indicate the simulated sampling location of the dropsonde. The top two subplots (sparse temperature / sparse humidity) show the sparse observation data obtained through simulated sampling based on the aforementioned HWRF field data. The top left shows sparse temperature observations, and the top right shows sparse humidity observations. This data is extracted from the complete HWRF field by simulating the trajectory of a dropsonde deployed by an aircraft passing through a typhoon. It can be seen that the observation data are only distributed radially or spirally, with large areas remaining data-free. Through... Figure 3 The comparison shows that this invention uses HWRF data to simulate the real-world dilemma of "sparse observations." With only a limited number of discrete observation points in the upper row, it is difficult to reconstruct the complex nonlinear flow field characteristics in the true HWRF data of the lower row using traditional interpolation. This invention aims to train a deep learning model to recover the complete HWRF temperature and humidity field of the lower row from the sparse HWRF sampling points in the upper row.
[0095] This invention innovatively models the sparse reconstruction problem of physical fields as an image inpainting problem in computer vision, and designs a complete data preprocessing, model inference, and post-processing workflow for this purpose. This invention utilizes an encoder-decoder network structure with an embedded "gated recurrent fusion module." This structure effectively handles large-area holes in the input data through a gated convolution mechanism, and achieves progressive, physically consistent information filling from known to unknown regions through a recurrent refinement mechanism, significantly improving reconstruction quality. This invention employs an adversarial training framework, combined with reconstruction loss, enabling the model to pursue not only numerical similarity but also the realism of the generated field in terms of physical structure and statistical characteristics, thereby generating high-fidelity results that are difficult to achieve with traditional methods.
[0096] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0100] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.
Claims
1. A method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations, characterized in that, Includes the following steps: Acquire radiosonde observation data and satellite remote sensing data for the target tropical cyclone region; The radiosonde observation data is mapped onto a two-dimensional grid on multiple preset pressure layers to generate an initial temperature and humidity field that includes regions with missing data. Based on the initial temperature and humidity field, a mask is generated to identify the locations where data is missing; The satellite remote sensing data is interpolated and aligned to a two-dimensional grid that is spatially aligned with the initial temperature and humidity field to form an auxiliary remote sensing field; For each of the aforementioned pressure layers, the initial temperature and humidity field, mask, and auxiliary remote sensing field of the current layer are combined to form a multi-channel input tensor, which is then input into a pre-trained deep neural network model. The deep neural network model infers and fills in the missing data regions based on the multi-channel input tensor, and outputs the complete two-dimensional temperature and humidity field of the current layer. The complete two-dimensional temperature and humidity fields corresponding to all pressure layers are vertically synthesized to obtain the three-dimensional temperature and humidity field of the target tropical cyclone.
2. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 1, characterized in that, The radiosonde observation data refers to the observation data from a drop-in radiosonde.
3. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 1, characterized in that, The preset multiple pressure layers are standard isobaric surfaces.
4. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 1, characterized in that, The radiosonde observation data is mapped onto a two-dimensional grid on multiple preset pressure layers, specifically including: At each pressure layer, the longitude, latitude, temperature, and humidity values of the target tropical cyclone are extracted at discrete points based on radiosonde observation data. The data of the discrete points are then mapped onto a regular two-dimensional grid covering the region of the target tropical cyclone to form an initial temperature and humidity field including areas with missing data.
5. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 1, characterized in that, Based on the initial temperature and humidity field, a mask is generated to identify the locations of missing data, specifically including: Based on the spatial distribution of data points in the initial temperature and humidity field, a binary mask with the same size as the two-dimensional grid is generated, wherein the mask value corresponding to the known data location is the first value, and the mask value corresponding to the missing data location is the second value.
6. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 1, characterized in that, The deep neural network model is a deep convolutional network based on an encoder-decoder architecture, and a gated recurrent fusion module is provided between the encoder and the decoder; the encoder is used to downsample the multi-channel input tensor to extract abstract feature maps; The gated loop fusion module is used to fuse and complete the abstract feature map; The decoder is used to upsample the completed feature map to reconstruct a complete two-dimensional temperature and humidity field.
7. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 6, characterized in that, The gated cyclic fusion module includes a gated convolutional unit and a cyclic thinning unit connected in sequence; the gated convolutional unit is used to process the abstract feature map output by the encoder, and by generating an adaptive gate value for the feature map, it distinguishes and regulates the feature flow from the known data region and the missing data region; The recurrent refinement unit adopts a recurrent neural network structure to iteratively update the feature map after it has been processed by the gated convolution unit, so that information can be propagated and fused from the known region to the inside of the data missing region, thereby refining and completing the feature representation of the data missing region.
8. The method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations according to claim 1, characterized in that, The deep neural network model is obtained through adversarial training; the adversarial training adopts a generator-discriminator framework, wherein the deep neural network model serves as the generator and a discriminator network is provided; the discriminator is used to distinguish between the complete two-dimensional temperature and humidity field output by the generator and the real complete two-dimensional temperature and humidity field, and outputs the discrimination result. The training objective of the generator is constrained by a composite loss function, which includes an adversarial loss based on the discrimination result and a pixel-level reconstruction loss between the generator output field and the real field.
9. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for reconstructing the three-dimensional temperature and humidity field of tropical cyclones based on deep learning and sparse observations as described in any one of claims 1-8.
10. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the method according to any one of claims 1-8.