Typhoon intensity prediction and residual error correction method and system based on double-branch convolutional neural network

Abstract

The invention belongs to the technical field of meteorology, and discloses a typhoon intensity prediction and residual error correction method and system based on a double-branch convolutional neural network, and the method comprises the steps: obtaining satellite-typhoon data; data preprocessing and space-time matching: constructing and cleaning a sample set; building a double-branch CNN (Content Node Network)-DenseNet network; model training; and carrying out residual analysis and system error correction. According to the method, key physical information such as severe convection and rapid enhancement of a typhoon core area is fully mined by combining a traditional multi-channel satellite cloud picture and an infrared-water vapor channel difference physical indication factor, and the perception ability of the model to a complex typhoon structure and a development process is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to but is not limited to the field of meteorological technology, and particularly relates to a typhoon intensity prediction and residual correction method and system based on a double-branch convolutional neural network. BACKGROUND

[0002] As a highly destructive meteorological disaster, the monitoring and forecasting of typhoons are of great importance to the safety of coastal areas. Currently, meteorological satellite remote sensing has become the core means for all-weather monitoring of typhoons. Satellites such as FY-4 and S-NPP can provide multi-channel, high-temporal and high-spatial resolution cloud image data, laying a data foundation for typhoon target detection and intensity estimation. However, the field of typhoon forecasting technology still faces significant challenges: on the one hand, the structure of a typhoon is complex, and its intensity change is affected by multiple factors such as sea surface temperature and atmospheric vertical wind shear, showing highly nonlinear dynamic characteristics. On the other hand, traditional forecasting methods have bottlenecks in accuracy and timeliness, and it is urgent to combine artificial intelligence technology to achieve a breakthrough. From the perspective of technological development, typhoon forecasting technology is transforming from traditional methods based on empirical rules to data-driven intelligent algorithms. The breakthroughs of deep learning technology in image recognition and time series prediction provide a new path for typhoon monitoring and forecasting.

[0003] Currently, there are mainly three types of algorithms for typhoon intensity inversion using satellite images: Dvorak technology, directional gradient angle technology (DAVT), and deep learning methods. Traditional typhoon positioning mainly relies on Dvorak technology, which determines the center position by manually analyzing the cloud pattern features such as the typhoon eye and spiral rain bands on satellite cloud images. However, this method has some subjective errors due to the proficiency of forecasters in the method, and has certain limitations, making it difficult to achieve automation and standardization. The subsequently developed DAVT technology is based on the statistical analysis of the directional gradient of satellite infrared image brightness, and finally fits a nonlinear equation to calculate the typhoon intensity. Although this method reduces human intervention, it has limited ability to extract complex textures and dynamic changes in cloud systems, and the positioning error is significant when the cloud structure is fuzzy or there are interfering cloud clusters.

[0004] Both of the above methods mostly rely on converting satellite images into high-level artificial features. However, it is difficult to determine some general features to establish a regression model for typhoons at different development stages and in different regions. Therefore, scholars have attempted to introduce deep learning methods into typhoon research. Early methods mostly used single-input CNN models, i.e., only using a single type of image, such as infrared cloud top brightness temperature images, to infer typhoon intensity Wang et al., 2022. With further research, more and more work has attempted to integrate multi-source information to build a dual-input or multi-input CNN architecture to improve performance. For example, some studies have used multi-spectral satellite channels as different inputs to the CNN, forming a multi-channel feature fusion model. Lee et al. Lee et al., 2020 used multi-spectral satellite channels as different inputs to the CNN, forming a multi-channel feature fusion model, with a root mean square error (RMSE) of about 8.32 knots, which is about 35% lower than the error of a single-channel model. Yin et al. Yin et al., 2024 proposed a regional convolutional fusion network, which uses one CNN to extract global cloud image features of the entire storm and another CNN to extract local details of the typhoon eye region, and uses multi-level feature fusion and spatial / channel attention mechanisms to improve the accuracy of intensity estimation. In addition, the most commonly used input for CNN models is the infrared (IR) cloud image of geostationary meteorological satellites, as infrared brightness temperature can indicate the cloud top structural features of a typhoon and can be monitored continuously day and night. Some studies have combined visible light cloud images (VIS) to utilize more detailed information during the day, and have added other spectral bands such as water vapor channels when feasible to obtain additional features such as cloud height and water vapor distribution. In addition, some studies have integrated visible light, infrared, and water vapor images. Tan et al. selected five cloud parameter products from the Himawari-8 satellite, including cloud optical thickness, cloud top temperature, cloud top height, cloud particle effective radius, and cloud category, and expanded the training data by adding samples at different rotation angles to obtain more rich physical information Tan et al., 2022. Similarly, some scholars have attempted to integrate other meteorological elements as input features, such as combining environmental field variables provided by numerical weather prediction or reanalysis with satellite cloud image features as input to the model. Kim et al. have integrated satellite core region features and large-scale circulation reanalysis data into the CNN framework for typhoon intensity prediction. The adaptability of CNN models to different stages of typhoon development, such as generation, intensification, maturity, and weakening, is another focus of recent research. Some work has found that the model is relatively accurate in estimating the intensity of mature strong typhoons with clear eye and symmetric cloud wall structures, but performs poorly in the early stages of typhoon development or weakening. Some studies have improved the prediction ability of weak typhoons by data augmentation, sample weight adjustment, and the introduction of spatiotemporal coding methods Zhang et al., 2022.

[0005] Through the above analysis, the problems and defects of the prior art are:

[0006] (1). The input feature design lacks pertinence and does not fully exploit key physical indicator factors. Most existing CNN models only use traditional multi-channel satellite cloud images such as infrared brightness temperature, visible light, and water vapor channels as input. Although these models can extract the overall shape features of cloud systems, they lack the design of physical indicator factors that reflect key processes such as strong convection and rapid intensification within a typhoon.

[0007] (2). The output indicators are single, and the prediction results lack physical consistency constraints. Existing deep learning prediction models mainly focus on the direct prediction of single indicators such as maximum wind speed or central pressure, and rarely output joint estimates of multiple physical quantities such as wind speed and pressure. It is difficult to use the physical relationship between them to cross-constrain and correct the prediction results.

[0008] (3). Lack of prediction error correction mechanism based on historical data. Current technology directly outputs prediction results, lacks residual analysis correction methods based on historical best path data wind speed and pressure relationships, and is difficult to correct systematic errors of the model in a timely manner to improve the stability and reliability of the prediction results.

[0009] The difficulty of solving the above problems and defects:

[0010] (1). The design of input feature physical indicator factors is difficult. To extract physical indicator factors from traditional satellite cloud images that can reflect key processes such as strong convection and rapid intensification within a typhoon, not only requires a deep understanding of typhoon structure evolution and related physical mechanisms, but also needs to conduct complex feature construction, preprocessing, and verification work on massive observation data. In addition, how to effectively integrate such new features with deep neural networks, avoiding the introduction of noise or weakening key features, is an important technical difficulty. Especially when the input features include differential channels and other physical indicator factors, to ensure that the network can fully extract deep-level key features and suppress redundant information, a deep network architecture with strong feature extraction and information flow transmission capabilities must be selected. Compared with traditional CNN structures, ResNet effectively alleviates the gradient vanishing and degradation problems in deep network training by introducing residual connections, allowing the model to maintain feature extraction capabilities while increasing the number of layers and accelerating convergence, making it the preferred solution for such complex input feature tasks.

[0011] (2). It is difficult to implement multi-physical quantity joint prediction and physical consistency constraints. Jointly predicting multiple physical indicators such as typhoon center wind speed and pressure and implementing physical consistency constraints during network training requires deep customization of neural network structure and training parameters. This not only increases the complexity of network design and parameter tuning, but also puts higher requirements on the stability, convergence speed, and generalization ability of model training.

[0012] (3) The integration of prediction error correction mechanisms based on historical data is complex. Constructing a residual analysis and correction framework based on historical best path data, such as the relationship between wind speed and air pressure, requires effectively statistically analyzing the prediction error patterns after model output and designing reliable correction methods to achieve dynamic correction. This requires high-quality processing and in-depth statistical analysis of historical data, while ensuring efficient integration of the correction module with the main model in terms of structure and data flow, improving the overall system's robustness and real-time application capabilities, making technical integration quite challenging. Summary of the Invention

[0013] To address the problems existing in the prior art, this invention provides a method for typhoon intensity prediction and residual correction based on a dual-branch convolutional neural network. Specifically, it involves combining traditional multi-channel satellite cloud images and infrared-water vapor channel differential physics indicator factors, using an improved CNN-DenseNet network to jointly predict the typhoon center wind speed and air pressure, and using historical best path data to perform residual analysis to correct the prediction results.

[0014] This invention is implemented as follows: a method for typhoon intensity prediction and residual correction based on a dual-branch convolutional neural network, the method comprising:

[0015] Step 1: Satellite-Typhoon Data Acquisition; Read the best typhoon path files released by the China Meteorological Administration to obtain data such as the occurrence time, center geographical location, central pressure, and maximum wind speed of each typhoon; Batch read satellite radiometric calibration data and supporting geometric positioning data, and organize them into a complete path name list; Traverse each best path record and search for satellite data files of the same time, year-month-day-hour-minute.

[0016] Step 2: Data Preprocessing and Spatiotemporal Matching: Preprocess the read satellite data and obtain the coordinates of the observed satellite pixels; based on the Haversine distance formula, find the coordinates of the observed pixel closest to the center of the typhoon with the optimal path; construct a three-dimensional matrix based on the channel brightness temperature data with this coordinate point as the center, the matrix size is S×S×C, where S is the size of the extracted satellite data and C is the number of channels, and save it as a single sample feature; simultaneously save the corresponding typhoon intensity label, typhoon center pressure and maximum wind speed.

[0017] Step three: sample set construction and cleaning; traverse each best path record to produce a training sample set, which is a four-dimensional satellite data matrix with a size of SxSxCxN, S is the size of the intercepted satellite data, C is the number of channels, and N is the sample size; At the same time, generate the corresponding typhoon intensity label set, which is a four-dimensional typhoon data matrix with a size of 1x1x2xN, 2 is the central pressure and maximum wind speed of the typhoon, and N is the sample size; Delete invalid and abnormal samples, mark wind speed and central pressure outliers with the quartile range method, and simultaneously remove the corresponding cloud images; Randomly shuffle the residual samples and divide them into training, validation and test sets according to the proportion;

[0018] Step four: building a double-branch CNN-DenseNet network; the first branch: directly retaining the original observation channel data; the second branch: generating infrared-water vapor difference physical indicator data by differencing the satellite infrared and water vapor channels two by two; double-branch data standardization; each branch feature extraction is based on Dense-Block construction;

[0019] Step five: model training;

[0020] Step six: residual analysis and system error correction; based on the central pressure and maximum wind speed of each typhoon sample in the best path set, a wind speed-pressure linear model is fitted; input the typhoon data at the time to be predicted into the model to obtain the predicted pressure; calculate the residual error of the linear model and the CNN model predicted central pressure, and then standardize it to determine whether it is abnormal by threshold; adjust the wind speed of the abnormal sample by the proportion of target error ± 10hPa, and correct the pressure output by the wind speed / pressure standard deviation ratio at the same time;

[0021] Further, the step one comprises:

[0022] Step 11: download the FY-4A multi-channel scanning imaging radiometer AGRIL1 radiometric data of Fengyun-4A FY-4A at the time to be corrected, with a resolution of 4KM, GEO data and CMA best path text;

[0023] Step 12: Read the infrared and water vapor channel data in AGRIL1 data and the corresponding scaling table of each channel: 3.72um channel 4KM image data layer NOMChannel08, 3.72um channel scaling table CALChannel08; 6.25um channel 4KM image data layer NOMChannel09, 6.25um channel scaling table CALChannel09; 7.10um channel 4KM image data layer NOMChannel10, 7.10um channel scaling table CALChannel10; 8.50um channel 4KM image data layer NOMChannel11, 8.50um channel scaling table CALChannel11; 10.8um channel 4KM image data layer NOMChannel12, 10.8um channel scaling table CALChannel12; 12um channel 4KM image data layer NOMChannel13, 12um channel scaling table CALChannel13; 13.5um channel 4KM image data layer NOMChannel14, 13.5um channel scaling table CALChannel14; read the row and column numbers in GEO data: LineNumber, ColumnNumber;

[0024] Step 13: Read the data in the CMA best path file: open the text of the best path of typhoon published by China Meteorological Administration, and scan line by line; when the first character of the line is 66666, it is determined as the new record header of a certain typhoon, and the international number international_ID, the path line number traceLine, the tropical cyclone number tcNumber, the typhoon number typhoonID, the termination flag tcEndRecord, the path sampling interval sampleInterval, the typhoon name typhoonName and the data release date datasetDate are extracted and stored in turn; then the specific path data is read according to the traceLine number of the typhoon, and the year year, month month, day day, hour hour, intensity level strength, center latitude Lat, center longitude Long, center pressure pres and maximum wind speed Wnd are parsed line by line, and all are written into the structure array typhoonData, which provides specific numerical values for subsequent satellite-typhoon space-time matching and training set construction;

[0025] Step 14: Read the FY-4A AGRIL1 radiation calibration data and GEO data stored in the local folder in turn, extract the file name, and then splice the path to form a complete path name list; then, traverse the L1 satellite data path name that has been sorted, extract the timestamp in the file name, and compare the year, month, day, and hour of each best path one by one to determine whether they are completely consistent with the best path record time year, month, day, and hour; if the match is successful, it is considered that the file is the supporting satellite observation data of the typhoon path at that time, and the next step of longitude and latitude conversion and brightness temperature data clipping processing step 2 is entered; if the match fails, the data at that time does not participate in the subsequent training data set construction.

[0026] Further, the step two comprises:

[0027] Step 21: Observation pixel coordinate conversion: read the ColumnNumber column number and LineNumber row number variables in the GEO data file, replace the invalid value -1 with NaN first; then, according to the earth ellipsoid parameters and the imaging geometry model of the stationary satellite, first convert each pixel point row and column number to radian, and then calculate the corresponding geographic longitude and latitude through coordinate projection conversion; the specific transformation adopts the nominal projection of the stationary orbit defined by CGMS LRIT / HRIT global specification, and the geographic coordinates are calculated based on the WGS84 reference ellipsoid, and the parameters are as follows: the distance from the center of the earth to the center of the satellite 42164km, the semi-major axis of the earth 6378.137km, the semi-minor axis of the earth 6356.7523km, the longitude of the satellite nadir point 105°, the column offset 1373.5, the column scale factor 10233137, the row offset 1373.5, and the row scale factor 10233137;

[0028] Step 22: Observe pixel brightness-temperature conversion: Two parts of data were read in step 12; the first part is the image data layer "NOMChannelX", where X represents 14 channel numbers, a total of 14 data layers, storing the image data of 14 channels respectively, i.e., digital quantization values ​​DN, which have no physical meaning. This invention selects the infrared and water vapor channel image data with channel numbers 8-14; the data type is integer, the data dimension is 2748×2748, the effective data range of the 7th channel is [0, 65534], and the effective data range of the other channels is [0, 4095], which is invalid. The fill value is 65535; the second part is the calibration table "CALChannelX" (X: 1~14), which stores the calibration tables for 14 channels, used to process digital quantization values ​​into albedo or brightness temperature; the data type is floating point, the data dimension of the 7th channel is 65536×1, and the data dimension of the other channels is 4096×1; the effective value range of the visible / near-infrared band is [0, 1.5], and the DN value is processed into albedo; the effective range of the infrared band is [100, 500], and the DN value is processed into brightness temperature; infrared and water vapor images are used, so X takes values ​​from 8 to 14;

[0029] The calibration table stores the albedo or brightness temperature data corresponding to each valid DN value according to the range of DN values, forming a one-to-one lookup table. The index of the table is the DN value. For example, in the calibration table CALChannel01 of channel 1, the value corresponding to index 200 is 0.0375. Therefore, the albedo corresponding to all pixels with a DN value of 200 in NOMChannel01 is 0.0375. In the calibration table CALChannel07 of channel 7, the value corresponding to index 200 is 493.5. Therefore, the brightness temperature corresponding to all pixels with a DN value of 200 in NOMChannel07 is 493.5K.

[0030] Step 23: In the typhoon path-assisted identification and precise matching method of observed pixels in this invention, the Haversine spherical distance formula is proposed to identify the observed pixel point in remote sensing satellite data that is closest to the typhoon center. First, missing values ​​are processed for the latitude and longitude coordinates of all observed pixels provided by remote sensing satellites, and files containing null values ​​NaN are skipped to ensure the numerical stability of subsequent calculations. Then, the latitude and longitude of the typhoon center position and all observed pixels are uniformly converted from angle to radian to adapt to the input format of the Haversine distance formula. Next, the Haversine spherical distance is calculated based on the following formula:

[0031]

[0032]

[0033] in, and are the latitude of the typhoon center and the observation pixel, respectively; is the latitude difference; Δλ is the longitude difference; R is the radius of the earth, which is 6371 kilometers;

[0034] The distance between the typhoon center point and each observation pixel is batch calculated; finally, the minimum distance value d min is the spherical distance between the current typhoon center and the closest pixel in the remote sensing observation field; if the minimum distance is less than a preset threshold, which is 4 kilometers in the present application, which is the spatial resolution of the observation pixel, then it is determined that the matching is successful, and the corresponding observation pixel row and column number coordinates LineNumber, ColumnNumber are output; otherwise, a default invalid matching flag is returned and the current file is skipped;

[0035] Step 24: The matched observation pixel coordinate point is LineNumber, ColumnNumber. Taking this coordinate as the center, a two-dimensional brightness temperature data sub-region of SxS size is intercepted in the remote sensing satellite channel brightness temperature data in the horizontal and vertical directions respectively, wherein S = 251 in the present application; for each set of remote sensing observation, the brightness temperature data of different channels is extracted and spliced into a three-dimensional brightness temperature feature tensor, whose dimension is SxSxC, wherein SxS is the spatial size of the region, and C is the number of channels used;

[0036] Step 25: After completing the extraction of the three-dimensional brightness temperature observation feature matrix, the sample is saved as an independent file using a unified naming rule containing typhoon number, date, etc. to facilitate batch training; at the same time, the typhoon intensity parameter corresponding to the observation time, the typhoon center minimum pressure value (unit: hPa) and the maximum near-center wind speed (unit: m / s) are extracted and saved as label data at the same time; the label data is used as the regression target in the subsequent supervised learning task, and provides annotation basis for establishing the typhoon intensity estimation model.

[0037] Further, the step three comprises:

[0038] Step 31: Read the historical typhoon hourly best path data records in sequence, including the key meteorological parameters such as the typhoon center longitude and latitude, minimum central pressure, and maximum wind speed at each time; for each time record, call the algorithm in step 2 to determine the observation pixel coordinate point closest to the typhoon center in the remote sensing satellite observation data; take this coordinate point as the center to extract a brightness temperature sub-region with a spatial scale of SxS from the multi-channel satellite observation data, and select the infrared and water vapor channels, so that each typhoon sample at each time corresponds to a three-dimensional matrix with dimensions SxSxC; further, splice all the samples to construct a unified four-dimensional training feature dataset with dimensions SxSxCxNS, where S is the size of the satellite data, C is the number of channels, and N is the sample size; simultaneously extract the central pressure in hPa and maximum wind speed in m / s corresponding to each time path record as intensity labels, and combine to construct a four-dimensional label matrix with dimensions 1x1x2xN, where 2 is the typhoon central pressure and maximum wind speed, and N is the sample size, ensuring strict alignment with the training feature data in the sample dimension;

[0039] Step 32: Extract the wind speed and pressure variables from the typhoon intensity label matrix in the dataset, and use the Interquartile Range (IQR) statistical method to identify outliers; based on the intensity labels of the entire sample set, calculate the 25th percentile Q1 and 75th percentile Q3 values, and then obtain the interquartile range:

[0040] IQR = Q1 - Q3

[0041] Set the outlier threshold range as:

[0042] Lower bound = Q1 - 1.5xIQR

[0043] Upper bound = Q3 + 1.5xIQR

[0044] Any sample with wind speed or central pressure less than the lower bound or greater than the upper bound is considered an outlier in a statistical sense; the detected abnormal samples are not directly changed, but are marked with a Boolean mask for subsequent screening and processing; for the intensity labels identified as outliers, the corresponding multi-channel cloud feature samples are simultaneously removed to ensure that the sample-label structure is strictly matched; after this step, a high-quality training sample set and label set are generated after cleaning, avoiding interference from outliers in model training and improving the prediction accuracy and stability of the model for typhoon intensity;

[0045] Step 33: First, obtain the cleaned multi-channel satellite cloud sample set with M samples, and set a fixed random seed to ensure reproducibility of the data splitting process in multiple experiments; generate a random sequence of length M by calling the random sorting function to shuffle the sample indices;

[0046] Subsequently, the sample index was divided into three subsets according to a ratio of 70%:15%:15%, corresponding to:

[0047] (1) The first 70% of the training set index: used for model parameter fitting;

[0048] (2) The middle 15% of the validation set index: used for hyperparameter tuning and overfitting monitoring during training;

[0049] (3) Last 15% of the test set index: used for independent evaluation of the final model performance;

[0050] Based on the above index, the original multi-channel remote sensing image feature matrix and typhoon intensity label matrix are simultaneously divided into three data subsets, denoted as follows:

[0051] (1)P train ,T train Training samples and labels;

[0052] (2)P vali ,T vali P: Validation samples and labels;

[0053] (3)P test ,T test Test samples and labels;

[0054] By randomly shuffling and stratified sampling in this step, we ensure a balanced data distribution, effectively avoiding modeling biases caused by temporal or distributional deviations of the samples, and providing a reliable data foundation for the subsequent training and generalization evaluation of deep neural network models.

[0055] Furthermore, step four includes:

[0056] Step 41: The model input feature design adopts a dual-branch strategy, which retains the original observation channel data and constructs physical difference features respectively, so as to fully explore the synergistic change law between the infrared and water vapor channels;

[0057] The first input branch will preserve the original channels and directly select the channel sequences in the original multi-channel remote sensing observation data as input features to retain; this part of the data has been divided into training set PtrainP, validation set Pvali and test set Ptest according to step 33;

[0058] The second input branch is constructed for an infrared-water vapor differential factor. Since the infrared-water vapor differential factor is obtained by combining the brightness temperatures of the infrared window channel and the water vapor channel, the high-level through-top strong convective characteristics are highlighted, and the recognition ability of the symmetry and strength of the typhoon core is improved. The factor can effectively suppress thin cirrus and high-level abnormal interference, and is highly related to the height of the minimum central pressure of the typhoon, and has good potential for strength estimation. The specific process is as follows: based on the original data in the first branch, a specific channel combination with strong physical sensitivity is further extracted, and the present application selects channel 9, a water vapor channel with a center wavelength of 6.25 μm, channel 10, a water vapor channel with a center wavelength of 7.1 μm, channel 12, an infrared channel with a center wavelength of 10.7 μm, and channel 13, an infrared channel with a center wavelength of 12.0 μm, to calculate the two-by-two difference value to construct the infrared-water vapor channel differential factor. The specific combinations include the following:

[0059] (1) Channel 9-Channel 12;

[0060] (2) Channel 9-Channel 13;

[0061] (3) Channel 10-Channel 12;

[0062] (4) Channel 10-Channel 13;

[0063] Each set of differential results constitutes an independent channel, and finally generates a differential feature tensor with a dimension of SxSx4xN, which is used for training, verification and testing, and is named Ptrain_diff, Pvali_diff and Ptest_diff, respectively.

[0064] Step 42: First, the original observation data of the first input branch is standardized at the channel level. The mean and standard deviation of each channel in the training set are calculated. When the standard deviation is less than a numerical threshold such as 1e -8 , in order to avoid division by zero error, the standard deviation is set to 1 by default. Then, using the statistical quantity of each channel obtained from the training set, the corresponding channels of the training set, the validation set and the test set are standardized and converted, so that all pixel values conform to the approximate standard normal distribution with a mean of 0 and a variance of 1;

[0065] Subsequently, for the second input branch infrared-water vapor difference factor, four channels perform complete symmetric channel standardization operations, also based on the training set to calculate the mean and standard deviation of each channel, and apply to all subset training, validation, test for normalization processing; for the output end of the regression label, that is, the minimum pressure and maximum wind speed of the typhoon center, the minimum-maximum normalization method is adopted, and the two target variables are scaled to the [0, 1] interval; The specific method is: taking the minimum and maximum values in the training set as the scaling parameters, linear normalization is performed on the training labels and validation labels, and the test labels are denormalized again in the prediction stage; This method keeps the regression label value interval stable, preventing gradient explosion or convergence difficulty caused by uneven target value scale;

[0066] Step 43: Double-branch CNN-DenseNet network structure, the specific design is as follows:

[0067] (1) Input structure: The model designs two input branches: adopts double-branch parallel input, the first branch receives standardized original satellite observation data 7 channels, the input size is S×S×7, which is used to retain complete cloud top temperature and spatial texture information, and the second branch receives the constructed infrared-water vapor difference physical factor 4 channels, the input size is S×S×4, in order to highlight the high-level water vapor disturbance and strong convective top-penetrating signal; The two branches do not interfere with each other before the data enters the network, ensuring that the original observation and the difference feature independently perform deep representation learning;

[0068] (2) Branch feature extraction module DenseBlock structure: Each branch consists of a starting convolution module and four DenseBlock convolution modules, a total of nine convolution layers in each branch; The starting module uses 4×4 convolution with a step of 2, padding of 3, filter of 64, followed by batch normalization BN, ReLU activation and 2×2 maximum pooling, to quickly reduce the resolution and expand the receptive field; Each DenseBlock block contains two layers of convolution: the first layer is a 1×1 convolution with 256 filters to realize cross-channel feature compression, and the second layer is a 3×3 convolution with 128 filters, same padding to extract local texture; Batch normalization BN and ReLU activation are inserted between the two layers inside the DenseBlock block, and the branch output is concatenated with the input at the end of the block through the depthConcatenation function, forming 4 times of feature reuse, so as to balance local detail capture and global information flow;

[0069] (3) After each of the two branches completes four DenseBlock blocks, the highest-level feature maps are concatenated in the channel dimension using the depthConcatenation function to achieve multi-source information fusion. The fusion tensor is then compressed to a 1×1×D channel vector by global average pooling, which significantly reduces the number of parameters and the risk of overfitting while maintaining key semantics. Finally, a 2D regression result is output through 1×1 convolution, which corresponds to the lowest sea level pressure and the maximum wind speed at the center of the typhoon, respectively. This “fusion-compression-regression” design not only fully integrates the complementary information of the original brightness temperature field and the difference factor, but also makes the end-to-end training process more stable and efficient.

[0070] Furthermore, step five includes:

[0071] Step 51: Configure the optimizer and learning rate plan: The optimizer uses momentum stochastic gradient descent with a momentum coefficient of 0.9; the initial learning rate is 0.1, and it is implemented in a piecewise decay cycle of 5 epochs with a decay coefficient of 0.2; to suppress overfitting, L2 regularization of 1×10 is applied in the weight update. -4 ;

[0072] Step 52: Implement gradient clipping and mini-batch training: During the training phase, enable absolute gradient clipping with a threshold of 0.005 for all trainable parameters to prevent gradient overflow; use a batch size of 64 for forward and backward propagation in each round; the total training cycle is 40 epochs; calculate the metric on the validation set every 20 iterations; the training process runs on the GPU, and the sample order is reshuffled at the end of each epoch to improve convergence stability;

[0073] Step 53: Model Saving: After complete training, export the parameter weights for subsequent deployment.

[0074] Furthermore, to establish the first-order regression baseline: the maximum wind speed x and minimum central pressure y of typhoons are extracted from historical samples of the typhoon's best path set. A one-dimensional linear model y = ax + b is established using first-order polynomial fitting. Subsequently, based on this model, the air pressure is extrapolated using the test set wind speed x1 to obtain the test set air pressure estimation baseline y. base This baseline serves as a reference curve for subsequent residual analysis and adaptive correction.

[0075] Step 62: Residual Standardization and Outlier Identification: Calculate the original air pressure prediction y of the network. pred Compared with baseline y base The residual vector r = y pred y base Decentralize r and divide by the standard deviation to obtain the standardized residual r0. * Set absolute threshold |r * Samples with a value greater than 2 are identified as abnormal and an abnormal sample index is generated for subsequent error correction processes.

[0076] Step 62: Target error driven adaptive correction: define target error E tar ==10hPa, calculate correction ratio alpha = E for each abnormal sample tar / |r|; fuse the original prediction with the true value by (1-alpha) weight to generate the corrected air pressure y pred '; simultaneously map the air pressure change to the wind speed according to the power-frequency standard deviation ratio to obtain the corrected wind speed x pred '; finally output the typhoon intensity estimation result x that meets the application accuracy requirement pred ' and y pred '.

[0077] In combination with the above technical solutions and the technical problems solved, the technical solutions to be protected by the present application have the following advantages and positive effects:

[0078] First, the significance of solving the above problems and defects is that the improved scheme has important significance for the problems of insufficient input features, single model architecture, lack of physical consistency constraints, and insufficient error correction ability in the typhoon intensity prediction of the prior art. First, fully mining the differential features of infrared and water vapor channels can more accurately identify the supercell structure of the typhoon core area, improve the sensitivity of the model to the key process of typhoon rapid intensification, and effectively improve the scientificity and accuracy of typhoon intensity prediction. Second, by designing a double-input CNN-DenseNet architecture, combining conventional multi-channel cloud images and physical feature differential indexes can help to fully utilize the complementary advantages of information from different data sources, enhance the expression and discrimination ability of the model for complex typhoon structures. Third, simultaneously outputting the typhoon center wind speed and air pressure and introducing residual error analysis correction based on historical best path data in the post-processing stage can help to improve the physical consistency, robustness and credibility of the prediction results, and enhance the application value of the model in actual forecasting business. Finally, constructing an end-to-end prediction framework integrating deep feature extraction, physical factor utilization and error correction can help to promote the intelligent prediction technology of typhoon intensity to high precision, diversification and physical consistency, and has important engineering application value and social significance for typhoon disaster prevention, meteorological forecasting and disaster reduction.

[0079] To solve the three types of bottlenecks in the objective estimation of typhoon intensity for a long time: ① Only direct modeling of traditional multi-channel cloud images, lack of physical indicators for strong convective / rapid enhancement, ② Single index output (only typhoon maximum wind speed or only typhoon center minimum pressure) leads to the lack of wind pressure relationship constraints and poor physical consistency, ③ Lack of systematic error correction based on historical best path, and these problems correspond to "feature physical design difficulty, joint prediction and consistency constraint difficulty, and correction framework integration difficulty" in engineering, respectively, the invention proposes a technical solution of "double branch + differential physical factor + residual adaptive correction" to solve the above problems. At the same time, from the observation of physics, FY-4A / AGRI 6.25 mu m and 7.1 mu m two water vapor channels respectively indicate the middle and upper tropospheric water vapor, and form a contrast sensitive pair with 10.8 / 12.0 mu m infrared window channel, which provides a physical indicator for depicting deep convection and structure reorganization from "cloud top cooling-upper layer water vapor disturbance".

[0080] Based on the above understanding, the invention adopts a double-branch design of "original multi-channel observation (IR / WV) + infrared-water vapor differential physical indicator" on the input side: branch A retains 7 original observation channels; Branch B makes a difference between AGRI 10.8 / 12.0 mu m and 6.25 / 7.1 mu m (9-12, 9-13, 10-12, 10-13), which enhances the contrast clues of "strong convective penetration-upper layer dry-wet gradient", so as to explicitly inject the key information of "whether there is deep convection / whether it enters the rapid enhancement precursor" into the machine learning space. The above construction is not a simple channel stacking, but decoupling of "original spectral information" and "physical differential clues" and then deep fusion after recoupling; DenseNet as the backbone of each branch, with the help of "cross-layer feature reuse and gradient short circuit" capability, ensures efficient information flow transmission after the introduction of differential channels, suppresses the risk of redundancy and overfitting, and further improves the detectability of weak signals and training stability. In research and development, we complete the spatio-temporal registration at the pixel level according to the best path matching, construct an SxSxC (S=251) sample cube with the center pixel as the window, and use the maximum wind speed and center pressure of the CMA best path as the paired labels; Before training, the intensity label is identified for IQR anomaly and the corresponding cloud image sample is removed, and then the training / validation / test set is divided according to 70%, 15%, 15%; The two branch channels are standardized respectively, the output end adopts double target regression (maximum wind speed and center pressure joint), and after the fusion layer, the global average pooling is used for dimensionality reduction regression, and the momentum SGD, L2 regularization and gradient clipping are introduced on the training side to ensure convergence stability and reproducibility.

[0081] In the output layer and post-editing, the invention does not follow the traditional path of "first estimating the maximum wind speed and then converting the central pressure (or vice versa)", but simultaneously outputs both end-to-end, and implements "consistency conservation" with the historical wind pressure relationship as the physical baseline: first, fit the linear wind pressure relationship of historical samples as the baseline, then calculate the standardized residual r* of the prediction and the baseline, and judge the abnormality when |r*|>2; for abnormal samples, perform proportional correction with a target error of ±10 hPa, and map the change between the maximum wind speed and the central pressure according to the ratio of wind speed / pressure standard deviation, so as to integrate the classical wind pressure relationship into the network output in the form of "soft constraint", making the final result physically self-consistent and statistically robust.

[0082] Thus, the invention brings about the following creative technical effects:

[0083] (a) The perception of typhoon complex structure and rapid intensification is significantly enhanced. The IR-WV difference amplifies the physical contrast of "cold top deep convection - upper layer water vapor anomaly", making the network more sensitive to key processes such as rapid intensification and eye wall replacement, which is superior to the method of simply stacking channels side by side; the FY-4A WV channel weighting function is located in the middle and upper troposphere, providing a physical basis for difference sensitivity.

[0084] (b) Dual-target consistency is improved. Through the joint prediction and post-editing consistency conservation of the historical wind pressure baseline, the output of physical contradictions such as "high wind pressure / low wind pressure" is systematically reduced.

[0085] (c) The robustness and engineering implementability are improved. Pixel-level registration and abnormal sample IQR cleaning reduce the label-feature mislocation; DenseNet branch and standardization process improve training stability; residual threshold (|r*|>2) and target error ±10 hPa adaptive post-editing are explainable and adjustable on the business side, which is convenient for direct integration into the meteorological business platform.

[0086] Secondly, as the creative auxiliary evidence of the invention's claim, it is also reflected in the following important aspects:

[0087] 1. The expected income and commercial value of the technical solution of the invention after transformation are:

[0088] Facing the coastal industrial pattern of China, the dual-branch CNN-DenseNet typhoon intensity prediction and residual correction method can be directly embedded into the provincial and municipal emergency warning and industry platform, directly serving the new mechanism of warning to people-disposal to post, and the task book of "meteorological high-quality development". On the one hand, about 8 tropical cyclones cause social and economic losses in China every year, and the economic risk center of typhoon is moving northward at a speed of about 19.7 km per year, and the demand for early identification of intensity level and rapid enhancement in high-density economic zones such as the Yangtze River Delta and the Bohai Rim continues to grow; on the other hand, only in the first half of 2025, the direct economic losses caused by natural disasters in China reached 541.1 billion yuan, and 461,000 people were relocated in China during the "Mars" process in 2024, showing the real demand for strengthening the intelligent estimation of intensity. The application can be connected with the grassroots emergency mechanism of "warning direct grid responsibility person, call for response", drive the threshold triggering of linkage strategies such as closure of ports, reduction of railway / flights, evacuation of oil and gas and wind power-stop-machine-reinforcement, reduce the implicit cost and downtime caused by "excessive disposal / disposal lag" (such as the typical situation of the Yangshan deepwater port stopping operation during typhoon period, nearly 30% of flights being reduced or canceled at Shanghai airport during typhoon period), and provide more stable "wind-pressure pair" elements for insurance / reinsurance pricing and post-damage determination, reducing model uncertainty. At the policy level, it is highly consistent with the "more accurate forecast, more refined service, and'meteorology +'industry empowerment" proposed in the "Meteorological High-Quality Development Outline (2022-2035)", and has a comprehensive commercial prospect of government disaster reduction value, port energy supply chain value, and insurance financial value.

[0089] 2, The technical scheme of the application fills the domestic and foreign technical blank:

[0090] (1) Input level: At present, the typhoon intensity estimation method based on deep learning is mostly multi-channel stacking or attention fusion, and less "infrared-water vapor difference" is constructed as an explicit physical indicator to amplify the contrast of "cold top convection-upper water vapor". The application adopts a dual-branch design of "original multi-channel branch + infrared-water vapor difference branch", which systematically highlights the typhoon core convection information.

[0091] (2) Model and output level: The existing objective estimation algorithm (such as ADT, SATCON) and a large number of DL researches are mostly single indicators or post-conversion; the application jointly regresses the maximum wind speed and central minimum pressure of typhoon, introduces the wind-pressure historical relationship as a posterior "soft constraint", and then uses the standardized residual threshold (|r*|>2) and the adaptive mechanism of target error ±10hPa to close loop correction, forming an integrated process that takes into account accuracy and physical consistency.

[0092] (3) Observation utilization layer: two water vapor channels (6.25 mu m, 7.1 mu m) and infrared window (10.8 mu m, 12.0 mu m) of FY-4A / AGRI are fully engineered, and the peak value of the weight function represents the upper troposphere / middle troposphere sensitive layer, which is suitable for difference to enhance the physical interpretability.

[0093] 3. The technical scheme of the present application overcomes technical bias:

[0094] Bias one: the more channels, the better, and let the network learn by itself

[0095] The present application does not throw all satellite data into the network for "black box experiment", but constructs an infrared-water vapor difference factor and fuses in the double branch, embeds physical priori from the input side into the network, and improves the interpretability of the network.

[0096] Bias two: predict one first, and then use empirical relationship to predict the other

[0097] The present application avoids the path dependence of using "single index" for empirical conversion by end-to-end joint regression and soft constraint and residual correction based on historical baseline. BRIEF DESCRIPTION OF DRAWINGS

[0098] Figure 1 is a flowchart of a typhoon intensity prediction and residual correction method based on a double-branch convolutional neural network provided by an embodiment of the present application;

[0099] Figure 2 is a double-branch CNN-DenseNet network structure provided by an embodiment of the present application;

[0100] Figure 3 Scatter plot of intensity estimate value and optimal path set;

[0101] Figure 4 Absolute error curve of intensity estimate value;

[0102] Figure 52024 tropical cyclone sample optimal path set and model estimated intensity deviation curve. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0104] As shown in Figure 1 , an embodiment of the present application provides a typhoon intensity prediction and residual correction method based on a double-branch convolutional neural network, which comprises:

[0105] Step one: satellite-tropical cyclone data acquisition; read the tropical cyclone best track file released by China Meteorological Administration, obtain the occurrence time, center geographical position information, central pressure, maximum wind speed and other data of each tropical cyclone; batch read satellite radiation calibration data and supporting geometric positioning data, and arrange into a complete path name list; traverse each best track record, search for satellite data files at the same time, same year-month-day-hour-minute;

[0106] Step two: data preprocessing and space-time matching: preprocess the read satellite data and obtain the coordinates of the satellite observation pixels; based on the Haversine distance formula, find the observation pixel coordinate point closest to the center of the best track tropical cyclone; take the coordinate point as the center, construct a three-dimensional matrix based on the channel brightness temperature data, the matrix size is SxSxC, S is the size of the intercepted satellite data, C is the channel number, and save it as a single sample feature; save the corresponding tropical cyclone intensity label, tropical cyclone central pressure and maximum wind speed at the same time;

[0107] Step three: sample set construction and cleaning; traverse each best track record, produce training sample set, which is a four-dimensional satellite data matrix, the matrix size is SxSxCxN, S is the size of the intercepted satellite data, C is the channel number, and N is the sample size; At the same time, generate the corresponding tropical cyclone intensity label set, which is a four-dimensional tropical cyclone data matrix, the matrix size is 1x1x2xN, 2 is the central pressure and maximum wind speed of the tropical cyclone, and N is the sample size; delete invalid and abnormal samples, mark wind speed and central pressure outliers with quartile range method, and simultaneously remove the corresponding cloud image; randomly shuffle the residual samples, and cut the training, validation and test sets in proportion;

[0108] Step four: double-branch CNN-DenseNet (Convolutional Neural Network-Densely Connected Convolutional Networks) network building; first branch: directly retain the original observation channel data; second branch: difference between satellite infrared and water vapor channels, generate infrared-water vapor difference physical indicator factor data; double-branch data standardization; each branch feature extraction is based on Dense-Block building;

[0109] Step five: model training;

[0110] Step six: residual analysis and system error correction; based on the central pressure and maximum wind speed of each tropical cyclone sample in the best track set, a wind speed-pressure linear model is fitted; input the tropical cyclone data at the time to be predicted into the model to get the predicted pressure; calculate the residual error of the linear model and the CNN model predicted central pressure, and adjust whether it is abnormal after standardization by threshold; adjust the wind speed by the proportion of target error ±10hPa for abnormal samples, and correct the pressure output by the wind speed / pressure standard deviation ratio.

[0111] The step one comprises:

[0112] Step 11: download the FY-4A Advanced Geosynchronous Radiation Imager (AGRIL1) radiometric calibration data, resolution 4KM, GEO data and CMA optimal path text to be revised;

[0113] Step 12: read the infrared and water vapor channel data in the AGRIL1 data and the calibration table corresponding to each channel: 3.72um channel 4KM image data layer NOMChannel08, 3.72um channel calibration table CALChannel08; 6.25um channel 4KM image data layer NOMChannel09, 6.25um channel calibration table CALChannel09; 7.10um channel 4KM image data layer NOMChannel10, 7.10um channel calibration table CALChannel10; 8.50um channel 4KM image data layer NOMChannel11, 8.50um channel calibration table CALChannel11; 10.8um channel 4KM image data layer NOMChannel12, 10.8um channel calibration table CALChannel12; 12um channel 4KM image data layer NOMChannel13, 12um channel calibration table CALChannel13; 13.5um channel 4KM image data layer NOMChannel14, 13.5um channel calibration table CALChannel14; read the row and column numbers in the GEO data: LineNumber row number, ColumnNumber column number;

[0114] Step 13: Read data in CMA best path file: open the text of the best path of typhoon released by China Meteorological Administration, and scan line by line; when the first character of the line is 66666, it is determined to be the new record header of a certain typhoon, and the international ID international_ID, the path line number traceLine, the tropical cyclone number tcNumber, the typhoon number typhoonID, the termination flag tcEndRecord, the path sampling interval sampleInterval, the typhoon name typhoonName and the data release date datasetDate are extracted and stored in turn; then the specific path data of the typhoon is read according to the traceLine number of the typhoon, and the year year, the month month, the day day, the hour hour, the intensity level strength, the center latitude Lat, the center longitude Long, the center pressure pres and the maximum wind speed Wnd are parsed line by line, and all are written into the structure array typhoonData, to provide specific numerical values for subsequent satellite-typhoon space-time matching and training set construction;

[0115] Step 14: Read the FY-4A AGRIL1 radiation calibration data and GEO data stored in the local folder in turn, extract the file name, and then form a complete path name list by path splicing; then, the path name of the L1 satellite data is traversed, the timestamp in the file name is extracted, and the year, month, day and hour of each best path are compared one by one to determine whether they are consistent with the best path record time year, month, day and hour; if the matching is successful, it is considered that the file is the matching satellite observation data of the typhoon path at that time, and the next step of longitude and latitude conversion and brightness temperature data cutting processing step 2 is entered; if the matching fails, the data at that time does not participate in the subsequent training data set construction.

[0116] The step two comprises:

[0117] Step 21: Observation pixel coordinate conversion: read ColumnNumber column number and LineNumber row number variables in GEO data file, first replace the invalid value -1 with NaN; then according to the earth ellipsoid parameters and the static satellite imaging geometry model, first convert each pixel point row and column number into radian, and then calculate the corresponding geographic longitude and latitude by coordinate projection conversion; the specific transformation adopts the static orbit nominal projection defined by CGMS LRIT / HRIT global specification, and the geographic coordinates are calculated based on the WGS84 reference ellipsoid, and the parameters are as follows: the distance from the center of the earth to the satellite center is 42164km, the semi-major axis of the earth is 6378.137km, the semi-minor axis of the earth is 6356.7523km, the satellite nadir point is at longitude 105°, the column offset is 1373.5, the column scale factor is 10233137, the row offset is 1373.5, and the row scale factor is 10233137;

[0118] Step 22: Observation pixel brightness temperature conversion: two parts of data are read in step 12; the first part is the image data layer “NOMChannelX”, X represents 14 channel numbers, a total of 14 data layers, which store the image data of 14 channels, i.e. digital quantization value DN, which has no physical meaning, and the present application selects the infrared and water vapor channel image data with channel numbers 8-14; the data type is integer, the data dimension is 2748x2748, the effective data range of the 7th channel is [0, 65534], and the effective data range of the remaining channels is [0, 4095], and the invalid filling value is 65535; the second part is the calibration table “CALChannelX” (X: 1~14), which stores the calibration table of 14 channels, which is used to process the digital quantization value into albedo or brightness temperature; the data type is floating point, the data dimension of the 7th channel is 65536x1, and the data dimension of the remaining channels is 4096x1; the effective value range of visible light / near infrared band is [0, 1.5], and the DN value after processing is albedo, and the effective range of infrared band is [100, 500], and the DN value after processing is brightness temperature; X takes values from 8 to 14 since infrared and water vapor images are used;

[0119] The calibration table stores the albedo or brightness temperature data corresponding to each effective DN value according to the value range of the DN value, forming a one-to-one lookup table, and the index of the table is the DN value; for example, in the channel 1 calibration table CALChannel01, the corresponding value is 0.0375 according to the index 200, so the albedo of all pixel points with DN value 200 in NOMChannel01 is 0.0375; in the calibration table CALChannel07 of channel 7, the corresponding value is 493.5 when the index is 200, so the brightness temperature value of all pixel points with DN value 200 in NOMChannel07 is 493.5K.

[0120] Step 23: In a typhoon path auxiliary identification and observation pixel accurate matching method in the application, the Haversine spherical distance formula is used to identify the observation pixel point closest to the typhoon center in the remote sensing satellite data. First, the latitude and longitude coordinates of all observation pixels provided by the remote sensing satellite are processed for missing values, and the file containing the null value NaN is skipped to ensure the numerical stability of the subsequent calculation. Then, the latitude and longitude of the typhoon center and all observation pixels are converted from angle system to radian system to adapt to the input format of the Haversine distance formula. Next, based on the following Haversine spherical distance calculation formula:

[0121]

[0122]

[0123] wherein, and are the latitudes of the typhoon center and the observation pixel, respectively; is the difference between the two latitudes; Δλ is the difference in longitude; R is the radius of the earth, which is 6371 kilometers;

[0124] The distance between the typhoon center point and each observation pixel is calculated in batches. Finally, the minimum distance value d min from the calculated distance matrix d is extracted, which is the spherical distance between the current typhoon center and the closest pixel in the remote sensing observation field. If the minimum distance is less than a predetermined threshold value, which is 4 kilometers in the application, which is the spatial resolution of the observation pixel, it is determined that the matching is successful, and the row and column number coordinates LineNumber, ColumnNumber of the corresponding observation pixel are output. Otherwise, the default invalid matching flag is returned and the file is skipped;

[0125] Step 24: The matched observation pixel coordinate point is LineNumber, ColumnNumber. Taking this coordinate as the center, a two-dimensional brightness temperature data sub-region of size SxS is extracted in the remote sensing satellite channel brightness temperature data in the horizontal and vertical directions respectively, wherein S = 251 in the application. For each set of remote sensing observations, the brightness temperature data of different channels is extracted and spliced into a three-dimensional brightness temperature feature tensor with dimensions SxSxC, where SxS is the spatial size of the region, and C is the number of channels used.

[0126] Step 25: After completing the extraction of the three-dimensional brightness temperature observation feature matrix, the sample is saved as an independent file containing the typhoon number, date, etc. using a unified naming convention for batch training. At the same time, the typhoon intensity parameters corresponding to the observation time, the typhoon center minimum pressure value (unit: hPa) and the maximum near-center wind speed (unit: m / s) are extracted and saved as label data. This label data is used as the regression target in subsequent supervised learning tasks, providing a basis for labeling the typhoon intensity estimation model.

[0127] The step three includes:

[0128] Step 31: Read the historical typhoon hourly best path data records one by one, including the typhoon center longitude and latitude, minimum center pressure, and maximum wind speed at each time. For each time record, call the algorithm in step 2 to determine the nearest observation pixel coordinate point in the remote sensing satellite observation data to the typhoon center. Take this coordinate point as the center, extract a brightness temperature sub-region with a spatial scale of SxS from the multi-channel satellite observation data, and select the infrared and water vapor channels. Thus, each typhoon sample at each time corresponds to a three-dimensional matrix with dimensions SxSxC. Further, all samples at all times are spliced to construct a unified four-dimensional training feature dataset with dimensions SxSxCxNS, where NS is the size of the satellite data, C is the number of channels, and N is the sample size. The center pressure (unit: hPa) and maximum wind speed (unit: m / s) corresponding to each path record at each time are extracted as intensity labels, and a four-dimensional label matrix with dimensions 1x1x2xN is constructed, where 2 is the typhoon center pressure and maximum wind speed, and N is the sample size, ensuring strict alignment with the training feature data in the sample dimension.

[0129] Step 32: Extract the wind speed and pressure variables in the typhoon intensity label matrix in the dataset, and use the Interquartile Range (IQR) statistical method to identify outliers. Based on the intensity labels of the entire sample set, calculate the 25th percentile Q1 and the 75th percentile Q3 values, and then obtain the interquartile range:

[0130] IQR = Q1 - Q3

[0131] Set the outlier threshold range as:

[0132] Lower limit = Q1 - 1.5xIQR

[0133] Upper limit = Q3 + 1.5xIQR

[0134] Any sample with wind speed or central pressure less than the lower limit or greater than the upper limit is considered an abnormal value in a statistical sense; the abnormal sample detected is not directly changed in the original data, but is marked by a Boolean mask for subsequent screening and processing; for the intensity label identified as abnormal, the corresponding multi-channel cloud image feature sample will be synchronously excluded to ensure that the sample-label structure is strictly matched; after this step, the cleaned high-quality training sample set and label set are generated, avoiding interference of abnormal values on model training and improving the prediction accuracy and stability of the model for typhoon intensity;

[0135] Step 33: First, obtain the cleaned multi-channel satellite cloud image sample set, with a sample number of M, and set a fixed random seed to ensure the reproducibility of the data splitting process in multiple experiments; generate a random sequence of length M by calling the random sorting function to shuffle the sample index;

[0136] Subsequently, the sample index is divided into three subsets in the ratio of 70%:15%:15%, corresponding to:

[0137] (1) The first 70% of the training set index: used for model parameter fitting;

[0138] (2) The middle 15% of the validation set index: used for hyperparameter tuning and overfitting monitoring during training;

[0139] (3) The last 15% of the test set index: used for independent evaluation of the final model performance;

[0140] According to the above index, the original multi-channel remote sensing image feature matrix and the typhoon intensity label matrix are synchronously divided into three data subsets, denoted as:

[0141] (1) P train ,T train : training sample and label;

[0142] (2) P vali ,T vali P: validation sample and label;

[0143] (3) P test ,T test : test sample and label;

[0144] Through the random shuffling and stratified sampling division of this step, the data distribution is balanced, effectively avoiding modeling bias caused by sample time series or distribution deviation, and providing a reliable data foundation for subsequent deep neural network model training and generalization evaluation.

[0145] The fourth step includes:

[0146] Step 41: The model input feature design adopts a double-branch strategy, respectively retaining the original observation channel data and constructing physical difference features to fully tap the collaborative change law between the infrared and water vapor channels;

[0147] The first input branch will keep the original channel, directly selecting the channel sequence in the original multi-channel remote sensing observation data as the input feature retention; this part of data has been divided into the training set PtrainP, the validation set Pvali and the test set Ptest according to step 33;

[0148] The second input branch is the construction of infrared-water vapor difference factors; since the infrared-water vapor difference factor highlights the high-level through-top strong convective features by combining the infrared window channel and the water vapor channel brightness temperature, it improves the recognition ability of the symmetry and intensity of the typhoon core; this factor can effectively suppress thin cirrus clouds and high-level abnormal interference, and is highly related to the height of the lowest central pressure of the typhoon, and has good potential for intensity estimation; specifically as follows: based on the original data in the first branch, further extract specific channel combinations with strong physical sensitivity, the present application selects channel 9 water vapor channel, center wavelength 6.25 μm, channel 10 water vapor channel, center wavelength 7.1 μm, channel 12 infrared channel, center wavelength 10.7 μm, channel 13 infrared channel, center wavelength 12.0 μm, and calculates the difference value of each two to construct the infrared-water vapor channel difference factor; specifically including the following combinations:

[0149] (1) Channel 9-Channel 12;

[0150] (2) Channel 9-Channel 13;

[0151] (3) Channel 10-Channel 12;

[0152] (4) Channel 10-Channel 13;

[0153] Each difference result constitutes an independent channel, and finally generates a difference feature tensor with a dimension of SxSx4xN, which is used for training, verification and testing, and is named as Ptrain_diff, Pvali_diff, Ptest_diff;

[0154] Step 42: First, the first input branch of the original observation data is 7 channel row channel level standardization; traverse each channel in the training set, calculate the mean and standard deviation of the whole pixel; when the standard deviation is less than a numerical threshold such as 1e -8 In order to avoid division by zero error, the standard deviation is set to 1 by default; then, using the per-channel statistics obtained from the training set, the corresponding channels of the training set, the validation set and the test set are standardized and converted, so that all pixel values conform to the approximate standard normal distribution with a mean of 0 and a variance of 1;

[0155] Subsequently, for the second input branch infrared-water vapor difference factor, four channels perform complete symmetric channel standardization operations, also based on the training set to calculate the mean and standard deviation of each channel, and applied to all subset training, validation, and testing for normalization processing; for the output end regression label, i.e., the minimum pressure and maximum wind speed of the typhoon center, the minimum-maximum normalization method is adopted to scale the two target variables to the [0, 1] interval; the specific method is: taking the minimum and maximum values in the training set as the scaling parameters, linear normalization is performed on the training labels and validation labels, and the test labels are denormalized again in the prediction stage; this method keeps the regression label value interval stable, preventing gradient explosion or convergence difficulty caused by uneven target value scale;

[0156] Step 43: Double-branch CNN-DenseNet network structure, as shown in Figure 2 , the specific design is as follows:

[0157] (1) Input structure: the model is designed with two input branches: a double-branch parallel input is adopted, the first branch receives standardized original satellite observation data 7 channels, with an input size of SxSx7, used to retain complete cloud top temperature and spatial texture information, and the second branch receives constructed infrared-water vapor difference physical factors 4 channels, with an input size of SxSx4, to highlight high-level water vapor disturbance and strong convective penetration signals; the two branches do not interfere with each other before the data enters the network, ensuring that the original observation and difference features independently perform deep representation learning;

[0158] (2) Branch feature extraction module DenseBlock structure: each branch consists of a starting convolution module and four DenseBlock convolution modules, a total of nine convolution layers in each branch; the starting module uses 4x4 convolution with a stride of 2, padding of 3, filter of 64, followed by batch normalization BN, ReLU activation, and 2x2 max pooling to quickly reduce the resolution and expand the receptive field; each DenseBlock block contains two layers of convolution: the first layer is a 1x1 convolution with 256 filters to realize cross-channel feature compression, and the second layer is a 3x3 convolution with 128 filters, same padding to extract local texture; batch normalization BN and ReLU activation are inserted between the two layers inside the DenseBlock block, and the branch output is concatenated with the input at the end of the block through the depthConcatenation function, forming 4 times of feature reuse, thus balancing local detail capture and global information flow;

[0159] (3) After each of the two branches completes four DenseBlock blocks, the highest-level feature maps are concatenated in the channel dimension using the depthConcatenation function to achieve multi-source information fusion. The fusion tensor is then compressed to a 1×1×D channel vector by global average pooling, which significantly reduces the number of parameters and the risk of overfitting while maintaining key semantics. Finally, a 2D regression result is output through 1×1 convolution, which corresponds to the lowest sea level pressure and the maximum wind speed at the center of the typhoon, respectively. This “fusion-compression-regression” design not only fully integrates the complementary information of the original brightness temperature field and the difference factor, but also makes the end-to-end training process more stable and efficient.

[0160] Step five includes:

[0161] Step 51: Configure the optimizer and learning rate plan: The optimizer uses momentum stochastic gradient descent with a momentum coefficient of 0.9; the initial learning rate is 0.1, and it is implemented in a piecewise decay cycle of 5 epochs with a decay coefficient of 0.2; to suppress overfitting, L2 regularization of 1×10 is applied in the weight update. -4 ;

[0162] Step 52: Implement gradient clipping and mini-batch training: During the training phase, enable absolute gradient clipping with a threshold of 0.005 for all trainable parameters to prevent gradient overflow; use a batch size of 64 for forward and backward propagation in each round; the total training cycle is 40 epochs; calculate the metric on the validation set every 20 iterations; the training process runs on the GPU, and the sample order is reshuffled at the end of each epoch to improve convergence stability;

[0163] Step 53: Model Saving: After complete training, export the parameter weights for subsequent deployment.

[0164] Furthermore, to establish the first-order regression baseline: the maximum wind speed x and minimum central pressure y of typhoons are extracted from historical samples of the typhoon's best path set. A one-dimensional linear model y = ax + b is established using first-order polynomial fitting. Subsequently, based on this model, the air pressure is extrapolated using the test set wind speed x1 to obtain the test set air pressure estimation baseline y. base This baseline serves as a reference curve for subsequent residual analysis and adaptive correction.

[0165] Step 62: Residual Standardization and Outlier Identification: Calculate the original air pressure prediction y of the network. pred Compared with baseline y base The residual vector r = y pred y base Decentralize r and divide by the standard deviation to obtain the standardized residual r0. * Set absolute threshold |r * Samples with a value greater than 2 are identified as abnormal and an abnormal sample index is generated for subsequent error correction processes.

[0166] Step 62: Target error driven adaptive correction: define target error E tar == 10 hPa, calculate correction ratio a = E for each abnormal sample tar / |r|; fuse the original prediction with the true value by (1-a) weight to generate the corrected air pressure y pred '; map the air pressure change to the wind speed according to the power-frequency standard deviation ratio to obtain the corrected wind speed x pred '; finally output the typhoon intensity estimation result x that meets the application accuracy requirement pred ' and y pred '.

[0167] I. Related evidence of the technical effects obtained by the embodiments of the present application.

[0168] Table 1 Confusion matrix of CNN-DenseNeT classification results based on the test set

[0169]

[0170] The confusion matrix is an important index for evaluating the model results, and can be used to analyze the classification reliability of different typhoon grades. Table 1 is the confusion matrix of CNN-DenseNeT for identifying different typhoon intensity grades. It can be seen that CNN-DenseNeT has good classification effect on tropical storm, strong tropical storm, typhoon and strong typhoon, and the classification accuracy is more than 95%. In comparison, the identification effect of tropical depression and super typhoon is poor, and the classification accuracy is about 80%. The possible reason is that there is different degree of confusion between tropical storms, and the intensity of tropical depression and tropical storm is weak, the structure is loose, and the shape is diverse, so the possibility of misjudgment between them is large, and the main reason for the error in this test is that the training of super typhoon is less.

[0171] Figure 3 The scatter plot of the intensity estimation value and the best path set; from 2021 to 2023, a total of 105 tropical cyclones were used to train the CNN-DenseNet model, and 21 tropical cyclones in 2024, that is, the test set, were used to evaluate the model. Figure 1 The relationship between the intensity estimation value and the best path set is shown. The red line represents the linear fitting of the intensity estimation value to the true wind speed, and the correlation coefficient is 0.80. The intensity estimation value varies from 8.57 to 60.43 m / s.

[0172] Figure 4The absolute error curve of the intensity estimates; it reflects the probability distribution of the absolute error. The horizontal axis is the proportion of the absolute error being less than a given value, and the vertical axis represents the absolute error between the ensemble of best tracks and the intensity estimates. It can also be seen from the figure that 25% of the absolute error is less than 2.47 m / s, 50% of the absolute error is less than 5.36 m / s, 75% of the absolute error is less than 9.58 m / s, and 90% of the absolute error is less than 12.68 m / s.

[0173] Figure 5 The bias curve of the intensity estimates versus the ensemble of best tracks for the 2024 tropical cyclone sample is shown.

[0174] Figure 5 The bias distribution on the test set is shown. Overall, the absolute degree of overestimation and underestimation is similar, and the bias oscillates around 0 m / s.

[0175] It should be noted that the embodiments of the present application can be realized by hardware, software, or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control codes, such as carrier media, such as magnetic disk, CD or DVD-ROM, programmable memory, such as read-only memory (firmware), or data carrier, such as optical or electronic signal carrier. The devices of the present application and their modules can be realized by hardware circuits, such as very large scale integrated circuits or gate arrays, semiconductors, such as logic chips, transistors, or programmable hardware devices, such as field programmable gate arrays, programmable logic devices, etc. They can also be realized by software executed by various types of processors, or by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0176] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any modification, equivalent replacement, and improvement made by those skilled in the art within the technical range disclosed by the present application, as long as it is within the spirit and principle of the present application, should be covered by the protection scope of the present application.

Claims

1. A typhoon intensity prediction and residual correction method based on a double-branch convolutional neural network, characterized in that, The method comprises the following steps: Obtaining satellite radiation calibration data and geometric positioning data, combining the typhoon best path file issued by the China Meteorological Administration to extract the typhoon center longitude and latitude, pressure and wind speed information; Performing coordinate conversion and brightness temperature calibration on the satellite observation pixels, matching the pixel point closest to the typhoon center based on the Haversine distance formula, and intercepting a fixed-size multi-channel brightness temperature sub-region to construct a three-dimensional feature matrix; Constructing a sample set and cleaning it to generate a multi-channel brightness temperature data feature set and a corresponding typhoon center pressure and maximum wind speed label set, and removing abnormal samples; Building a double-branch CNN-DenseNet network, wherein the first branch inputs the original multi-channel observation data, the second branch inputs the infrared-water vapor difference physical factor data, and the two branches are fused after extracting features through the DenseBlock convolution module respectively; Training the network model to obtain the typhoon intensity prediction output; Based on the first-order linear regression model of wind speed-pressure, residual analysis is performed on the CNN prediction results for abnormal identification and system error correction, and the corrected typhoon center pressure and maximum wind speed are output.

2. The method of claim 1, wherein, The satellite data is FY-4A AGRIL1 radiation calibration data with a resolution of 4km, and the channels include channels 8 to 14.

3. The method of claim 1, wherein, The observation pixel coordinate conversion adopts the stationary satellite projection model of CGMS LRIT / HRIT specification, calculates the longitude and latitude based on the WGS84 reference ellipsoid, and the parameters include: the distance from the center of the earth to the center of the satellite 42164km, the length of the earth's semi-major axis 6378.137km, the length of the earth's semi-minor axis 6356.7523km, and the sub-point longitude 105°.

4. The method of claim 1, wherein, The abnormal sample removal adopts the interquartile range method, calculates the 25th percentile value Q1 and the 75th percentile value Q3 of the wind speed and pressure in the typhoon intensity label, determines the abnormal judgment interval as $Q1-1.5×IQR,Q3+1.5×IQR$, and removes samples outside the range.

5. The method of claim 1, wherein, The second branch infrared-water vapor difference physical factor includes the brightness temperature difference values of the following channel combinations: (1) Channel 9 and channel 12; (2) Channel 9 and channel 13; (3) Channel 10 and channel 12; (4) Channel 10 and channel 13.

6. The method of claim 1, wherein, Each branch of the double-branch CNN-DenseNet network includes a starting convolution module, four DenseBlock modules and a pooling operation, and finally the outputs of the two branches are fused through channel splicing, and the typhoon intensity prediction value is output through global average pooling and convolution regression.

7. The method of claim 1, wherein, The training employed a momentum stochastic gradient descent optimizer with a momentum coefficient of 0.9, an initial learning rate of 0.1, and a piecewise decay performed once every 5 epochs with a decay coefficient of 0.2, and an L2 regularization parameter of 1 x 10 -4 .

8. The method of claim 1, wherein, The residual correction includes: Establishing a first-order linear model y=ax+b of wind speed x and pressure y; Calculating the residual r of the CNN prediction result and the linear model output, and normalizing it; When ∣r∣ is greater than the target threshold 10hPa, adjust the wind speed and pressure output in proportion.

9. The method of claim 8, wherein, The wind speed adjustment ratio a=Etar / ∣r∣, where Etar=10hPa, the corrected predicted pressure is ypred', and the corrected predicted wind speed is xpred'.

10. The method of claim 1, wherein, The final output result includes the corrected typhoon center minimum pressure and maximum wind speed, which is suitable for real-time typhoon monitoring and path prediction business.

Images