Low-altitude torrent and typhoon refined monitoring method and system

Abstract

The embodiment of the invention provides a low-altitude torrent and typhoon refined monitoring method and system, and is applied to the technical field of meteorological monitoring. The method comprises the following steps: acquiring cloud air guide data; time downscaling processing is carried out on the cloud guide wind data based on an RIFE model, and a high-resolution wind field sequence is generated; inputting the high-resolution wind field sequence into a pre-trained E3D-LSTM model, and outputting forecast wind field data; and respectively inputting the high-resolution wind field sequence and the forecast wind field data into a pre-trained Deeplabv3 + semantic segmentation model to obtain a low-altitude jet flow core area and a typhoon key area at corresponding moments so as to obtain an overall structure and an evolution trend of the wind field. In this way, an efficient and intelligent new normal form can be provided for refined monitoring and early warning of disastrous weather.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of meteorological monitoring, and in particular to a low-level jet and typhoon fine monitoring method and system. BACKGROUND

[0002] Typhoon and low-level jet are two types of wind field phenomena with important meteorological influence, and their occurrence and development are often accompanied by serious natural disasters and social and economic impact. Typhoon, as a strong tropical cyclone, not only brings devastating gale and heavy rain, but also often causes secondary disasters such as storm surge, flood and landslide. Low-level jet is a kind of high-speed airflow occurring in the boundary layer, and its core wind speed can usually reach 12 m / s or more, which has important influence on aviation safety, heavy rainfall distribution and wind energy utilization. The wind shear caused by low-level jet is an important dangerous factor in the take-off and landing stage of the aircraft, and its water vapor transport in the heavy rain system often aggravates the regional extreme precipitation event.

[0003] At present, the monitoring of typhoon and low-level jet mainly relies on meteorological radar, meteorological satellite, and numerical weather prediction (NWP) systems. However, the traditional observation means is limited by the spatial and temporal resolution, and it is difficult to capture its rapid evolution characteristics. For example, satellite remote sensing has a wide coverage, but there is still a lot of uncertainty in the vertical wind field detection; numerical weather prediction model can provide three-dimensional structure of wind field, but due to the limitation of computing resources, its time resolution is usually hourly, and the simulation capability of small-scale wind field characteristics is limited. The lack of monitoring and prediction ability leads to the possibility of missing or misjudging the key processes such as sudden change of typhoon path or enhancement of low-level jet, which directly affects the timeliness and accuracy of disaster prevention and reduction decision-making. SUMMARY

[0004] The present disclosure provides a low-level jet and typhoon fine monitoring method and system, which solves the technical problems of insufficient spatial and temporal resolution, weak small-scale wind field feature capturing ability, and low computing efficiency of traditional meteorological monitoring and prediction methods, and realizes real-time fine monitoring of low-level jet and typhoon and automatic target region segmentation and identification.

[0005] According to a first aspect of the present disclosure, a low-level jet and typhoon fine monitoring method is provided. The method comprises: obtaining cloud wind data; performing time downscaling processing on the cloud wind data based on a RIFE model to generate a high-resolution wind field sequence; inputting the high-resolution wind field sequence into a pre-trained E3D-LSTM model to output predicted wind field data; The high-resolution wind field sequence and the forecast wind field data are respectively input into the pre-trained Deeplabv3+ semantic segmentation model to obtain the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time, thereby obtaining the overall structure and evolution trend of the wind field.

[0006] As described above and in any possible implementation, a further implementation is provided in which the acquisition of cloud wind guidance data further includes: The cloud-guided wind data is encoded, including: calculating the zonal U component and the meridional V component based on the wind speed and wind direction data in the cloud-guided wind data, and then encoding the zonal U component, the meridional V component and the full wind speed channel together into a three-channel image-like data format.

[0007] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the RIFE model adopts an IFNet stacked hourglass structure, contains multiple cascaded IFBlock units, and progressively refines the intermediate optical flow estimation in conjunction with a privileged distillation strategy.

[0008] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the cooperative privileged distillation strategy progressively refines the intermediate optical flow estimation, including: performing a preliminary estimation of the intermediate optical flow on the cloud wind-guiding data through the first IFBlock unit; The preliminary estimated optical flow result is concatenated with the previous stage features and input into the subsequent cascaded IFBlock unit to obtain the optical flow result estimated by the RIFE model; wherein, each stage of IFBlock unit corrects the optical flow result of the previous stage. Using a teacher model that can directly access real intermediate frame wind field data, calculate the real optical flow result corresponding to the real intermediate frame; The deviation between the optical flow estimated by the RIFE model and the actual optical flow is calculated, and the deviation is backpropagated to the RIFE model through a loss function to adjust the RIFE model parameters.

[0009] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the E3D-LSTM model includes a 3D convolutional neural network and a long short-term memory network, and introduces a memory attention mechanism; wherein, The pre-trained E3D-LSTM model is obtained through the following steps: The high-resolution wind field sequence is reorganized in terms of time, space, and feature dimensions to construct three-dimensional wind field data that is adapted to the model input format; The E3D-LSTM model was trained using the reconstructed 3D wind field data. We use 3D convolution operations to capture the three-dimensional structure and local spatiotemporal dynamics of the wind field, use the gating mechanism of LSTM to learn the long-term dependencies of the wind field time series, and use the memory attention mechanism to simultaneously perceive the local dynamic changes and long-term evolution patterns of the wind field.

[0010] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the Deeplabv3+ semantic segmentation model includes an encoder and decoder, dilated convolution and spatial pyramid pooling modules.

[0011] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein inputting the high-resolution wind field sequence and the forecast wind field data into a pre-trained Deeplabv3+ semantic segmentation model to obtain the low-level jet core region and the typhoon key region at the corresponding time includes: The encoder extracts features from the input wind field data and expands the receptive field by combining dilated convolution. Multi-scale features are fused using spatial pyramid pooling modules; The decoder upsamples the fused features and then splices them with the shallow features from the encoder to refine them. Each pixel is then classified, and the core area of ​​the low-level jet stream and the key area of ​​the typhoon are automatically segmented.

[0012] According to a second aspect of this disclosure, a refined monitoring system for low-level jet streams and typhoons is provided. The system includes: The acquisition module is used to acquire cloud-guided wind data; The processing module is used to perform time downscaling processing on the cloud wind guidance data based on the RIFE model to generate a high-resolution wind field sequence. The prediction module is used to input the high-resolution wind field sequence into a pre-trained E3D-LSTM model and output forecast wind field data. The generation module is used to input the high-resolution wind field sequence and the forecast wind field data into the pre-trained Deeplabv3+ semantic segmentation model to obtain the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time, thereby obtaining the overall structure and evolution trend of the wind field.

[0013] According to a third aspect of this disclosure, an electronic device is provided. The electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.

[0014] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the methods according to the first and / or second aspects of this disclosure.

[0015] This disclosure acquires cloud-guided wind data and encodes it into a three-channel image format. It then uses the RIFE model to downscale the wind field over time to generate a high-resolution wind field sequence. The E3D-LSTM model, which integrates 3D convolution, LSTM, and memory attention mechanisms, is used to accurately extrapolate and forecast the wind field. Finally, the Deeplabv3+ semantic segmentation model is used to perform pixel-level category determination on the high-resolution wind field sequence and the forecast wind field, thereby automatically segmenting the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time. This effectively solves the problems of insufficient spatiotemporal resolution, weak ability to capture small-scale wind field features, and low degree of automation in traditional monitoring methods. It achieves refined and efficient monitoring of the overall structure and evolution trend of the wind field, providing reliable technical support for accurate early warning of low-level jet streams and typhoon-related hazardous weather.

[0016] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0017] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart of a method for refined monitoring of low-level jet streams and typhoons according to an embodiment of the present disclosure is shown; Figure 2 A block diagram of a low-level jet stream and typhoon fine monitoring system according to an embodiment of the present disclosure is shown; Figure 3 A block diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0019] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0020] This disclosure acquires cloud-guided wind data and encodes it into a three-channel image format. It then uses the RIFE model to downscale the wind field over time to generate a high-resolution wind field sequence. The E3D-LSTM model, which integrates 3D convolution, LSTM, and memory attention mechanisms, is used to accurately extrapolate and forecast the wind field. Finally, the Deeplabv3+ semantic segmentation model is used to perform pixel-level category determination on the high-resolution wind field sequence and the forecast wind field, thereby automatically segmenting the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time. This effectively solves the problems of insufficient spatiotemporal resolution, weak ability to capture small-scale wind field features, and low degree of automation in traditional monitoring methods. It achieves refined and efficient monitoring of the overall structure and evolution trend of the wind field, providing reliable technical support for accurate early warning of low-level jet streams and typhoon-related hazardous weather.

[0021] Figure 1 A flowchart of a method 100 for refined monitoring of low-level jet streams and typhoons according to an embodiment of this disclosure is shown. Figure 1 As shown, method 100 includes: S101, obtain cloud wind guidance data.

[0022] In some embodiments, the cloud-guided wind data is encoded, including: calculating the zonal U component and the meridional V component based on the wind speed and wind direction data in the cloud-guided wind data, and then encoding the zonal U component, the meridional V component and the full wind speed channel together into a three-channel image-like data format.

[0023] Specifically, cloud-guided wind (AMV) data provided by China's second-generation geostationary meteorological satellite FY-4B was selected. The dataset spans from January 2023 to January 2024, with a raw data time resolution of 15 minutes, stored in NetCDF format. It includes key elements such as wind speed and direction, and can basically reflect the basic state of the atmospheric wind field. Since the original two-dimensional wind speed and direction element format cannot be directly adapted to the training and inference requirements of the later model, it needs to be encoded. The specific process is as follows: Component calculation: Based on wind speed and direction from cloud-guided wind data, the zonal U-component and meridional V-component are calculated using trigonometric functions. It is used to characterize the movement of air currents in the east-west direction, such as east wind as positive and west wind as negative. It is used to characterize the airflow movement in the north-south direction, such as south is positive and north is negative. The directional information of wind direction can be completely preserved by combining the U and V components. Channel Supplement: A full wind speed channel is introduced, which directly quantifies the actual wind speed of each pixel (in m / s). Format conversion: The latitudinal U-component and meridional V-component obtained above are integrated with the full wind speed channel and encoded into a three-channel image-like data format. This format simulates the RGB three-channel structure of an image and can be adapted to the extraction logic of spatial and numerical features by deep learning models.

[0024] S102, based on the RIFE model, performs time downscaling on cloud-guided wind data to generate a high-resolution wind field sequence.

[0025] In some embodiments, the RIFE model employs an IFNet stacked hourglass structure, containing multiple cascaded IFBlock units, and uses a privileged distillation strategy to progressively refine intermediate optical flow estimates.

[0026] In some embodiments, the intermediate optical flow is initially estimated from the cloud wind-guiding data using the first IFBlock unit; The preliminary estimated optical flow result is concatenated with the previous stage features and input into the subsequent cascaded IFBlock unit to obtain the optical flow result estimated by the RIFE model; wherein, each stage of IFBlock unit corrects the optical flow result of the previous stage. Using a teacher model that can directly access real intermediate frame wind field data, calculate the real optical flow results corresponding to the real intermediate frames; The deviation between the optical flow estimated by the RIFE model and the actual optical flow is calculated, and the deviation is backpropagated to the RIFE model through a loss function to adjust the RIFE model parameters.

[0027] Specifically, IFNet, as the core network unit of the RIFE model, is a lightweight convolutional neural network that adopts a stacked hourglass structure design, containing multiple cascaded IFBlock units. It achieves coarse-to-fine optimization of the intermediate optical flow through an iterative logic of initial estimation and round-by-round correction. Initial optical flow estimation: Two frames of adjacent cloud-guided wind data after encoding, such as at 15-minute intervals, are input into the first IFBlock unit. This unit is based on the idea of ​​optical flow and uses convolution operation to initially extract the spatial features and spatiotemporal correlation information of the wind field, and then estimates the airflow motion vector at the intermediate moment between the two frames (i.e., the initial intermediate optical flow). The core is to capture the overall evolution trend of the wind field and the large-scale airflow motion pattern. Multi-round iterative correction: The preliminary optical flow result output by the first IFBlock is concatenated with the wind field features extracted by the unit and input into the subsequent cascaded IFBlock units. Each IFBlock unit takes the optical flow result and features of the previous level as input and makes targeted corrections for the deviations in the optical flow of the previous level. Through multiple rounds of cascaded iteration, the fineness of the optical flow estimation is gradually improved, and the optical flow result estimated by the RIFE model is obtained.

[0028] In some embodiments, to further reduce the systematic error in optical flow estimation, the RIFE model introduces a privileged distillation strategy to achieve external calibration of the optical flow results through the collaboration of teacher-student models. The specific process is as follows: Real optical flow calculation: Using the teacher model that can directly access real intermediate frame wind field data, calculate the real optical flow results corresponding to the real intermediate frames; Bias Calculation and Parameter Optimization: The deviation between the optical flow estimated by the RIFE model (student model) and the actual optical flow is calculated. This deviation is then backpropagated to all IFBlock units of the IFNet network through a loss function. Key configurations such as convolution kernel parameters and activation function thresholds of each unit are dynamically adjusted to correct the systematic errors accumulated during the iteration process, ultimately achieving accurate optimization of the intermediate optical flow.

[0029] In some embodiments, based on the accurate intermediate optical flow obtained by the aforementioned internal iterative refinement and external benchmark calibration, the RIFE model performs frame interpolation processing on two adjacent frames of original coded wind field data to generate three or more intermediate timeframes of wind field data. For example, for two frames of data with a 15-minute interval, three intermediate frames are generated, achieving a minute-level time resolution of 3 minutes / frame. The generated intermediate frame wind field data is then integrated with the original two frames of wind field data in chronological order to finally form a continuous, smooth minute-level high-resolution wind field sequence.

[0030] S103 inputs high-resolution wind field sequences into a pre-trained E3D-LSTM model and outputs forecast wind field data.

[0031] In some embodiments, the E3D-LSTM model includes a 3D convolutional neural network and a long short-term memory network, and incorporates a memory attention mechanism; wherein, The pre-trained E3D-LSTM model is obtained through the following steps: The high-resolution wind field sequence is reorganized in terms of time, space, and feature dimensions to construct three-dimensional wind field data that is adapted to the model input format; The E3D-LSTM model was trained using the reconstructed 3D wind field data. We use 3D convolution operations to capture the three-dimensional structure and local spatiotemporal dynamics of the wind field, use the gating mechanism of LSTM to learn the long-term dependencies of the wind field time series, and use the memory attention mechanism to simultaneously perceive the local dynamic changes and long-term evolution patterns of the wind field.

[0032] In some embodiments, the generated minute-level high-resolution wind field sequence is used as a training sample, and the data is reorganized according to the three dimensions of time, space, and features. The reorganized three-dimensional wind field data is then input into the E3D-LSTM model for iterative training, and each core module collaborates to complete feature learning according to its division of labor. 3D convolution operation: Convolution calculation is performed on three-dimensional wind field data in spatial (width, height) and temporal (or height layer) dimensions to effectively capture the three-dimensional spatial structure of the wind field (such as the ring distribution of the eyewall of a typhoon, the banded extension of the low-level jet stream) and local spatiotemporal dynamics (such as the instantaneous changes of small-scale turbulence, and the short-period fluctuations of airflow). LSTM gating mechanism: Through the synergistic effect of input gate, forget gate and output gate, the local features extracted by 3D convolution are filtered and memorized in the time dimension, focusing on learning the long-term dependence of wind field (such as the continuous trend of typhoon movement and the intensity increase and decrease law of low-level jet stream) to avoid the interference of short-term noise on prediction results. Memory Attention Mechanism: This mechanism dynamically allocates attention weights to link and fuse the local dynamic features captured by 3D convolution with the long-term evolution patterns of LSTM memory. It assigns higher weights to key changes such as sudden changes in typhoon paths and enhancement of low-level jet streams, simultaneously perceiving local details and global trends, and significantly improving the model's predictive sensitivity to complex meteorological processes.

[0033] During training, the loss function optimization targets are the mean absolute error (MAE), root mean square error (RMSE), and structural similarity (SSIM). The 3D convolution kernel parameters, LSTM gating threshold, and attention weight allocation rules are continuously adjusted through backpropagation until the prediction error of the model on the validation set tends to stabilize, thus obtaining the pre-trained model with the best performance.

[0034] In some embodiments, the latest eight consecutive minute-level high-resolution wind field samples are selected and adapted according to the three-dimensional reconstruction format of the training phase to ensure that the dimensions and feature channels of the input data are consistent with those during model training. The adapted input data is processed by the model's 3D convolutional layer to extract the three-dimensional structure and local dynamic features of the current wind field. Then, the LSTM layer calls the long-term evolution patterns remembered during the historical training process, and the memory attention mechanism strengthens the association and weight allocation of key features. Finally, the prediction result is output through the 3D deconvolutional decoding layer. After the prediction result is format-decoded, the forecast wind field data at three time steps (a total of 45 minutes) with 15-minute intervals are obtained. Its spatial resolution and feature channels (U component, V component, and full wind speed channel) are completely consistent with the input high-resolution wind field sequence and can be directly used as input for subsequent semantic segmentation models.

[0035] S104. The high-resolution wind field sequence and the forecast wind field data are respectively input into the pre-trained Deeplabv3+ semantic segmentation model to obtain the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time, thereby obtaining the overall structure and evolution trend of the wind field.

[0036] In some embodiments, the Deeplabv3+ semantic segmentation model includes an encoder and decoder, dilated convolution, and spatial pyramid pooling modules.

[0037] In some embodiments, the high-resolution wind field sequence and the forecast wind field data are respectively input into a pre-trained Deeplabv3+ semantic segmentation model to obtain the low-level jet core area and the key typhoon area at the corresponding time, including: The encoder extracts features from the input wind field data and expands the receptive field by combining dilated convolution. Multi-scale features are fused using spatial pyramid pooling modules; The decoder upsamples the fused features and then splices them with the shallow features from the encoder to refine them. Each pixel is then classified, and the core area of ​​the low-level jet stream and the key area of ​​the typhoon are automatically segmented.

[0038] In some embodiments, the specific training process of the pre-trained Deeplabv3+ semantic segmentation model is as follows: Training data preparation: Wind field samples labeled with the core area of ​​the low-level jet, the key structural area of ​​the typhoon, and the background area were selected as the training set. Among them, there were 547 training samples of the low-level jet, specifically data from 8:00 on June 17, 2024 to 8:00 on June 20, 2024 and from 8:00 on July 9, 2024 to 8:00 on July 12, 2024. There were 790 training samples of the typhoon, specifically data from 8:00 on July 26, 2023 to 8:00 on July 29, 2023 and from 8:00 on August 28, 2023 to 8:00 on September 3, 2023. All samples were wind field data in three-channel image format. Training objectives and optimization: The core objective is to achieve pixel-level class classification accuracy. The cross-entropy loss function is used to measure the difference between the model's prediction results and the sample labels. The model parameters are adjusted through backpropagation. During the training process, the average intersection-union ratio (MIoU) is used as the core evaluation metric until the MIoU of the model on the validation set tends to stabilize. Finally, a pre-trained model adapted for low-level jet stream and typhoon identification is obtained.

[0039] In some embodiments, the process of identifying high-resolution real-time wind fields and forecast wind fields follows the logic of feature extraction, multi-scale fusion, and precise segmentation, with the specific steps as follows: Feature extraction and receptive field expansion: High-resolution real-time wind field sequences and forecast wind field data are input into the encoder of the Deeplabv3+ model. The encoder performs hierarchical feature extraction on the wind field data through multi-layer convolution operations: shallow convolution extracts low-level spatial details of the wind field (such as local wind speed gradients and areas of abrupt changes in airflow direction), and deep convolution extracts high-level semantic features (such as the banded distribution features of low-level jet streams and the ring-shaped eyewall features of typhoons). At the same time, dilated convolution is introduced into the key convolutional layers of the encoder. This convolution expands the receptive field of the convolution kernel without increasing the model parameters and computational cost, accurately capturing large-scale contextual information of the wind field, and avoiding the omission of small-scale targets or incomplete segmentation of large-scale targets due to insufficient receptive field. Multi-scale feature fusion: The deep semantic features extracted by the encoder are input into the Spatial Pyramid Pooling (ASPP) module. This module simultaneously extracts and fuses wind field features at different scales by calling dilated convolutions with dilation rates of 6, 12, and 18 in parallel, combined with 1×1 convolution and image pooling operations. The dilated convolution with dilation rate of 6 captures the core features of the small-scale low-level jet stream, while the dilated convolutions with dilation rates of 12 and 18 capture the key structural features of the large-scale typhoon. Finally, the fused features covering multi-scale information are obtained by feature concatenation, ensuring that the model can accurately identify two types of weather systems with significant differences in shape and size at the same time. Feature refinement and pixel-level classification: The fused features output from the ASPP module are input into the decoder, which gradually restores the spatial resolution of the feature map through a 4x upsampling operation. At the same time, the upsampled features are concatenated with the low-level spatial detail features extracted from the shallow layer of the encoder to achieve complementary fusion of high-level semantic information and low-level spatial details, accurately correcting the segmentation boundary of the target region. Finally, the fused features are refined through a 3×3 convolutional layer, and each pixel is classified into three categories: low-level jet stream core area, typhoon key structure, and background region, automatically segmenting the target region at the corresponding time.

[0040] In some embodiments, the model outputs two types of recognition results: pixel-level segmentation maps of the current low-level jet core region and key typhoon structures based on high-resolution real-time wind fields, clearly presenting the current location, extent, and intensity distribution of the target area; and a segmentation map of the target area within the next 45 minutes (3 time steps) based on the forecast wind field, reflecting the expected movement direction and extent change trend of the target area. By comparing the regional overlap, magnitude of change, and positional offset of the two types of results, the overall spatial structure and short-term evolution patterns of the wind field can be directly obtained.

[0041] In some embodiments, the evaluation metrics are specifically: (1) Histogram test Histogram testing is a statistical method used to test whether data conforms to a specific distribution or to compare whether the distributions of two sets of data are consistent. It typically involves binning the data and then comparing the observed frequencies in each bin with the expected frequencies of the theoretical distribution.

[0042] (2) MAE (Mean Absolute Error) MAE is the average of the absolute values ​​of the differences between predicted and actual values. It is obtained by summing the absolute values ​​of the differences between each predicted value and its corresponding actual value, and then dividing by the sample size. The MAE value represents the average degree of difference between the model's predictions and the actual values. The calculation formula is as follows.

[0043] (3) RMSE (Root Mean Square Error) RMSE is the average of the squared differences between the predicted and actual values. Its calculation represents the difference in units between the original data, making it easier to understand and interpret the model's error. A smaller RMSE indicates that the model's prediction is closer to the actual value. The calculation formula is as follows.

[0044] (4) SSIM (Structural Similarity Measure) SSIM is a metric used to measure the similarity between two images. It is used to compare structural similarities and is suitable for error calculation in the current scene. The SSIM calculation range is -1 to 1, where 1 indicates that the two images are completely identical. The calculation formula is as follows.

[0045] (5) MIoU (Mutual Intersection and Union Ratio) MIoU is a commonly used evaluation metric in semantic segmentation tasks. Its calculation formula is as follows: In this system, TP (True Positive) indicates a correct prediction, where the predicted class is positive and the actual class is also positive; FP (False Positive) indicates a wrong prediction, where the predicted class is positive but the actual class is negative; and FN (False Negative) indicates a wrong prediction, where the predicted class is negative but the actual class is positive. A higher MIoU value indicates better model performance on the wind field recognition task, and this metric effectively measures the model's prediction accuracy for each category.

[0046] The following detailed description, using a specific embodiment, illustrates a method 100 for refined monitoring of low-level jet streams and typhoons provided in this disclosure: The data used was Atmospheric Motion Vector (AMV) product data provided by China's second-generation geostationary meteorological satellite FY-4B. The dataset spanned from January 2023 to January 2024, with the original data having a time resolution of 15 minutes. It was stored in NetCDF format and included key elements such as wind speed and direction. Before being input into the model, the data was encoded into a three-channel image-like data format.

[0047] Encoded data T0 and T1 were input into the RIFE model to generate three frames of wind field data at intermediate time points. These frames were then integrated with the original T0 and T1 data to form a minute-level high-resolution wind field sequence. The downscaling inference effect of the RIFE model was quantitatively evaluated. Wind field data from May 1, 2023 to July 27, 2023, totaling 7983 time points, were selected. Using a sliding frame method, five consecutive data points were taken as a group for each window, resulting in 1596 data groups. Randomized independent experiments were conducted, and based on the histogram test method, 60 data groups were selected for testing and statistical analysis. Based on the first and last time point wind field data of each group, three frames of wind field data were generated based on the model. The U and V winds were statistically analyzed separately. The histogram fitting results showed a good fit, indicating a high consistency between the frequency of the interpolated frames and the expected frequency of the actual distribution. This means that the data distribution obtained through frame interpolation closely approximates the actual data distribution, with most data points falling within the expected range. However, in lower-frequency regions, the interpolated results exceed the actual values. This phenomenon may be due to the limitations of frame interpolation methods when processing low-frequency data, such as improper parameter settings in the interpolation algorithm or insufficient handling of noise and outliers. In regions with lower data frequencies, the frame interpolation method may overestimate the frequency of the data, resulting in a fitted result that is higher than the expected frequency of the true distribution.

[0048] Based on wind field data from 2390 window samples from 2023 to the end of 2024, the extrapolated results of the U-wind and V-wind components were compared with the actual wind fields, and the mean absolute error (MAE), root mean square error (RMSE), and structural similarity index (SSIM) were calculated. The results are shown in Tables 1 and 2. Regarding the U-wind component, the error gradually increases with the increase of the prediction time step. The MAE for the first time step was 2.523, the RMSE was 3.191, and the SSIM was 0.739. By the third time step, the MAE increased to 3.766, the RMSE rose to 4.890, and the SSIM decreased to 0.609, indicating that the prediction accuracy decreases with the increase of the prediction time, but the model can still capture the overall structure and evolution trend of the wind field relatively well. The error trend of the V-wind component is basically consistent with that of the U-wind component, but the overall error level is slightly lower than that of the U-wind component. The MAE at the first time step is 2.041, RMSE is 2.764, and SSIM is 0.756. By the third time step, the MAE increases to 2.948, the RMSE increases to 4.04, and the SSIM decreases to 0.657, indicating that the model's prediction performance for the V-wind component is slightly better than that for the U-wind component, and it maintains relatively high structural similarity across multiple time steps. Overall, the model demonstrates good ability in wind field prediction, effectively capturing the spatiotemporal evolution characteristics of the wind field. However, as the prediction time increases, all error indicators increase, which is consistent with the general laws of meteorological forecasting. A total of 547 data points were selected from 8:00 on June 17, 2024 to 8:00 on June 20, 2024 and from 8:00 on July 9, 2024 to 8:00 on July 12, 2024 as low-level jet stream identification verification data; and a total of 790 data points were selected from 8:00 on July 26, 2023 to 8:00 on July 29, 2024 and from 8:00 on August 28, 2023 to 8:00 on September 3, 2023 as typhoon identification verification data. Low-level jet stream and typhoon identification were performed using the Deeplab V3+ model. The performance of the identification model was evaluated using the mean intersection-over-union ratio (MIoU), which measures the overlap between the predicted segmented region and the actual region. The evaluation results are shown in Table 3. The identification performance of both types of weather phenomena reached a good level. In the low-level jet stream identification task, the model achieved an MIoU of 66.2% on 547 test samples; while in the typhoon identification task, the model achieved an MIoU of 75.6% on 790 test samples. The results show that the model's ability to identify typhoons is better than that of low-level jet streams, and it demonstrates good identification accuracy and consistency in region segmentation. According to the embodiments of this disclosure, the limitations of traditional monitoring methods, such as insufficient spatiotemporal resolution, weak ability to capture small-scale wind field features, and low level of automation, are effectively overcome. A refined monitoring scheme with a full-process process of time downscaling, short-term extrapolation, and intelligent identification is formed, providing a new, efficient, and intelligent paradigm for real-time monitoring and short-term early warning of severe weather.

[0049] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0050] The above is an introduction to the method embodiments. The following describes the solution described in this disclosure further through device embodiments.

[0051] Figure 2 A block diagram of a low-level jet stream and typhoon fine-grained monitoring system 200 according to an embodiment of the present disclosure is shown. Figure 2 As shown, the device 200 includes: Module 201 is used to acquire cloud wind guidance data; Processing module 202 is used to perform time downscaling processing on cloud-guided wind data based on the RIFE model to generate high-resolution wind field sequences; The prediction module 203 is used to input high-resolution wind field sequences into a pre-trained E3D-LSTM model and output forecast wind field data. The generation module 204 is used to input the high-resolution wind field sequence and the forecast wind field data into the pre-trained Deeplabv3+ semantic segmentation model to obtain the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time, thereby obtaining the overall structure and evolution trend of the wind field.

[0052] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0053] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0054] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0055] Figure 3A schematic block diagram of an electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0056] Electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in ROM 302 or a computer program loaded into RAM 303 from storage unit 308. RAM 303 can also store various programs and data required for the operation of electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. I / O interface 305 is also connected to bus 304.

[0057] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0058] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as method 100. For example, in some embodiments, method 100 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of method 100 described above may be performed. Alternatively, in other embodiments, the computing unit 301 may be configured to perform method 100 by any other suitable means (e.g., by means of firmware).

[0059] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0060] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0061] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0062] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including voice input, speech input, or tactile input).

[0063] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0064] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0065] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0066] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for refined monitoring of low-level jet streams and typhoons, characterized in that, include: Obtain cloud-guided wind data; The cloud-guided wind data is downscaled over time based on the RIFE model to generate a high-resolution wind field sequence. The high-resolution wind field sequence is input into a pre-trained E3D-LSTM model, which outputs forecast wind field data. The high-resolution wind field sequence and the forecast wind field data are respectively input into the pre-trained Deeplabv3+ semantic segmentation model to obtain the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time, thereby obtaining the overall structure and evolution trend of the wind field.

2. The method according to claim 1, characterized in that, The acquisition of cloud wind guidance data also includes: The cloud-guided wind data is encoded, including: calculating the zonal U component and the meridional V component based on the wind speed and wind direction data in the cloud-guided wind data, and then encoding the zonal U component, the meridional V component and the full wind speed channel together into a three-channel image-like data format.

3. The method according to claim 1, characterized in that, The RIFE model adopts an IFNet stacked hourglass structure, which contains multiple cascaded IFBlock units, and uses a privileged distillation strategy to progressively refine the intermediate optical flow estimation.

4. The method according to claim 3, characterized in that, The collaborative privileged distillation strategy for progressively refining intermediate optical flow estimation includes: using the first IFBlock unit to initially estimate the intermediate optical flow from the cloud wind-guiding data; The preliminary estimated optical flow result is concatenated with the previous stage features and input into the subsequent cascaded IFBlock unit to obtain the optical flow result estimated by the RIFE model; wherein, each stage of IFBlock unit corrects the optical flow result of the previous stage. Using a teacher model that can directly access real intermediate frame wind field data, calculate the real optical flow result corresponding to the real intermediate frame; The deviation between the optical flow estimated by the RIFE model and the actual optical flow is calculated, and the deviation is backpropagated to the RIFE model through a loss function to adjust the RIFE model parameters.

5. The method according to claim 1, characterized in that, The E3D-LSTM model includes a 3D convolutional neural network and a long short-term memory network, and incorporates a memory attention mechanism; among which, The pre-trained E3D-LSTM model is obtained through the following steps: The high-resolution wind field sequence is reorganized in terms of time, space, and feature dimensions to construct three-dimensional wind field data that is adapted to the model input format; The E3D-LSTM model was trained using the reconstructed 3D wind field data. We use 3D convolution operations to capture the three-dimensional structure and local spatiotemporal dynamics of the wind field, use the gating mechanism of LSTM to learn the long-term dependencies of the wind field time series, and use the memory attention mechanism to simultaneously perceive the local dynamic changes and long-term evolution patterns of the wind field.

6. The method according to claim 1, characterized in that, The Deeplabv3+ semantic segmentation model includes encoders and decoders, dilated convolutions, and spatial pyramid pooling modules.

7. The method according to claim 6, characterized in that, The step of inputting the high-resolution wind field sequence and the predicted wind field data into a pre-trained Deeplabv3+ semantic segmentation model to obtain the low-level jet core region and the key typhoon region at the corresponding time includes: The encoder extracts features from the input wind field data and expands the receptive field by combining dilated convolution. Multi-scale features are fused using spatial pyramid pooling modules; The decoder upsamples the fused features and then splices them with the shallow features from the encoder to refine them. Each pixel is then classified, and the core area of ​​the low-level jet stream and the key area of ​​the typhoon are automatically segmented.

8. A refined monitoring system for low-level jet streams and typhoons, characterized in that, include: The acquisition module is used to acquire cloud-guided wind data; The processing module is used to perform time downscaling processing on the cloud wind guidance data based on the RIFE model to generate a high-resolution wind field sequence. The prediction module is used to input the high-resolution wind field sequence into a pre-trained E3D-LSTM model and output forecast wind field data. The generation module is used to input the high-resolution wind field sequence and the forecast wind field data into the pre-trained Deeplabv3+ semantic segmentation model to obtain the core area of ​​the low-level jet stream and the key area of ​​the typhoon at the corresponding time, thereby obtaining the overall structure and evolution trend of the wind field.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method described in any one of claims 1-7.

Images