A short-term precipitation prediction method based on the fusion of radar imagery and reanalysis data

Abstract

The present invention discloses a short-term precipitation prediction method based on the fusion of radar images and reanalysis data, comprising the following steps: collecting radar images and reanalysis data at consecutive times, generating input data in a unified grid format through spatial interpolation, time alignment and normalization processing; using a dual-channel encoder to extract the spatiotemporal features of the radar images and reanalysis data respectively, and performing weighted fusion through a channel attention mechanism to generate a fused feature tensor; inputting the fused features into a convolutional long short-term memory (ConvLSTM) neural network to model the spatiotemporal evolution process of the precipitation system and output a preliminary precipitation forecast image for the next 0-3 hours; constructing a residual learning network to perform bias correction on the preliminary forecast result based on historical residuals and observation information; when the radar image input is missing, maintaining the input structure integrity through a replacement feature generation module; generating a precipitation intensity image or probability map for the next 0-3 hours; and the present invention supports visual output.

Description

Technical Field

[0001] The present invention relates to the technical field of precipitation forecasting, and in particular to a short-term precipitation prediction method based on the fusion of radar images and reanalysis data. Background Art

[0002] In recent years, with the advancement of artificial intelligence (AI), particularly deep learning, a growing number of researchers have explored the application of deep neural networks to nowcast precipitation forecasting. Architectures such as convolutional neural networks (CNNs), convolutional long short-term memory (ConvLSTMs), generative adversarial networks (GANs), and the Transformer have been widely applied to radar echo sequence modeling. Some of these results have surpassed the performance of traditional algorithms such as TREC in terms of prediction accuracy and structure preservation.

[0003] For example, models such as PredRNN and NowcastNet achieve multi-step extrapolation of echo images by introducing spatiotemporal memory modules, showing good continuity and robustness; structures such as CIUnet combine channel attention mechanisms to probabilistically identify convection triggering signals in satellite cloud images; NNDA-VAE, Neural-Koopman and other studies introduce the idea of ​​data assimilation, model the state field through the encoder network, optimize the model residual in the latent space, and improve prediction stability.

[0004] Although deep learning technology provides new possibilities, existing research still faces the following major challenges:

[0005] (1) Multi-source data is difficult to fuse: radar images have high spatial resolution but strong timeliness, while reanalysis data such as ERA5 and FNL have low temporal resolution and coarse spatial granularity. Direct splicing will introduce physical inconsistencies or scale aliasing problems;

[0006] (2) Model errors are difficult to control: Neural networks have the problem of error accumulation during step-by-step prediction, especially in terms of echo intensity or boundary structure prediction, which easily deviates from actual observations and is difficult to work stably in the long term;

[0007] (3) Incomplete input is a common problem: radar sites are affected by terrain obstruction, equipment maintenance, signal loss, and other issues, and image sequences are prone to frame loss or holes, which affects model operation;

[0008] (4) Lack of explainability and credibility: Most AI models are “black box structures”, making it difficult for business personnel to track the basis of their predictions, which limits their use and promotion in actual weather warning systems.

[0009] Currently, there is still a lack of a unified modeling method with the following characteristics. Summary of the Invention

[0010] Purpose of the invention: The purpose of the present invention is to provide a short-term precipitation prediction method based on the fusion of radar images and reanalysis data. Based on a dual-channel coding structure and a convolutional long short-term memory neural network, the method can fuse data of different scales within a unified structure, effectively control the prediction residual through a dynamic error correction network, and have a modal missing completion function to ensure continuous operation when radar data is unavailable. The method is suitable for the refined intelligent forecasting needs of short-term heavy precipitation processes.

[0011] Technical solution: The present invention provides a short-term precipitation prediction method based on the fusion of radar images and reanalysis data, comprising the following steps:

[0012] (1) Collect radar images and reanalysis data at consecutive times, and generate input data in a unified grid format through spatial interpolation, time alignment, and normalization;

[0013] (2) A dual-channel encoder is used to extract the spatiotemporal features of radar images and reanalysis data respectively, and then weighted fusion is performed through a channel attention mechanism to generate a fused feature tensor;

[0014] (3) Input the fused features into the Convolutional Long Short-Term Memory (ConvLSTM) network to model the spatiotemporal evolution of the precipitation system and output a preliminary precipitation forecast image for the next 0-3 hours;

[0015] (4) Construct a residual learning network to correct the bias of the initial prediction results based on historical residuals and observation information;

[0016] (5) When radar image input is missing, the input structure integrity is maintained through the alternative feature generation module;

[0017] (6) Generate precipitation intensity images or probability maps for the next 0-3 hours and adapt them to the business system interface output.

[0018] Furthermore, in step (1), the radar image includes reflectivity Z, differential reflectivity , phase ratio The temporal resolution is 6-10 minutes, and the spatial resolution is 0.5 km × 0.5 km or 1 km × 1 km. The reanalysis data include temperature T, relative humidity RH, horizontal wind speed U, V and geopotential height. , with a temporal resolution of 1 hour and a spatial resolution of 0.25°×0.25°, aligned to the radar image grid by bilinear interpolation.

[0019] Furthermore, in step (2), the dual-channel encoder includes: a radar image encoder and a reanalysis variable encoder; the radar image encoder is a three-layer convolutional neural network, the number of channels is 32→64→64, and the activation function is ReLU; the reanalysis variable encoder adopts a 1×1 convolution and full connection structure to maintain the independence of variable channels, and the number of output channels is 64; the channel attention weight is a learnable parameter, and the fusion formula is:

[0020] ;

[0021] in, It is a learnable channel attention weight that adaptively allocates the fusion ratio according to the importance of the input features.

[0022] Furthermore, the reanalysis variable encoder independently encodes the variables of each pressure layer at 850 hPa, 700 hPa, and 500 hPa, and then concatenates them and upsamples them to the radar image resolution.

[0023] Furthermore, in step (3), ConvLSTM is a two-layer stacked structure with 64 input channels, 64 output channels, a convolution kernel size of 3×3, a step size of 1, and activation functions of Tanh and Sigmoid. The temporal modeling module adopts a parallel multi-step decoding mechanism to generate the next 18 frames of precipitation images through a two-layer convolutional network, namely a 3×3 convolution kernel and a 1×1 convolution kernel, with a time interval of 10 minutes and an output size of H×W.

[0024] Furthermore, the ConvLSTM unit enables the Dropout mechanism during the training phase and turns it off during the inference phase.

[0025] Furthermore, in step (3), the model training adopts an end-to-end approach, and simulated missing samples are constructed in the training set; the loss function is:

[0026] ;

[0027] in, is the mean square error loss; M is the number of time steps of the predicted image; is the spatial dimension; is the model's predicted value of precipitation intensity at location (i, j) at time t.

[0028] Furthermore, in step (4), the input of the dynamic error correction network includes the preliminary prediction image, the historical residual and reanalysis background variables;

[0029] The correction formula is:

[0030] ;

[0031] in, Represents the preliminary precipitation image output by the ConvLSTM network; The residual image predicted by the error correction network; is the final precipitation forecast result after revision;

[0032] The loss function includes the mean square error term and the L2 regularization term: ;in, is the predicted value With the observed value The mean square error between is the correction term amplitude control term; where λ is the adjustment weight.

[0033] Furthermore, in step (5), the modality missing completion includes: radar image validity detection, marking ;when When , a historical mean template or conditional generative network is used to generate alternative radar features ; Fusion weight Adjust to 0.5 or adaptively adjusted by the attention module.

[0034] Furthermore, in step (6), the forecast output includes a real-valued precipitation intensity map, a probability map converted by Sigmoid or Softmax, and a cumulative precipitation map; the output format is NetCDF or GeoTIFF, with attached geographic coordinate metadata, and supports sliding window inference, that is, updating the forecast results for the next 3 hours every 10 minutes.

[0035] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention can fuse meteorological data from different sources and at different scales (radar image + reanalysis background); while maintaining the accuracy of image details, it introduces a dynamic error correction mechanism to control output stability; it can automatically switch to "background-driven mode" to continue completing the precipitation forecast task when radar data is unavailable; it supports explanatory visual output to assist forecasters in understanding the model prediction logic. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the present invention;

[0037] Figure 2 It is a flow chart of the model of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, an embodiment of the present invention provides a short-term precipitation prediction method based on the fusion of radar images and reanalysis data, comprising the following steps:

[0040] (1) Acquisition and preprocessing of source meteorological data: Acquire the multi-source meteorological data required for model input and perform standardization on the original data in terms of space, time, and physical dimensions to construct a data sample set with a unified structure and dimension suitable for deep learning model input. Input data includes the following two categories:

[0041] Radar image data: reflectivity Z, differential reflectivity obtained by ground-based Doppler weather radar in the target area , phase ratio Image sequences of variables such as the radial velocity V, with high temporal resolution (e.g., once every 6 minutes or 10 minutes) and high spatial resolution (e.g., 0.5 km × 0.5 km or 1 km × 1 km);

[0042] Numerical reanalysis data: reanalysis field data from global or regional numerical weather models (such as ERA5, NCEPFNL, CMADS, etc.), with a temporal resolution of generally 1 hour and a spatial resolution of 0.25°× 0.25°. The main variables include surface and upper air temperature T, relative humidity RH, horizontal wind speed U, V and geopotential height etc., covering multiple standard air pressure layers (such as 850hPa, 700hPa, 500hPa).

[0043] The two types of data mentioned above have different temporal resolutions, spatial distribution ranges, and physical meanings. To ensure consistency between the data, collaborative training can be performed in the neural network, and the data can be preprocessed as follows:

[0044] (11) Spatial range clipping and resampling: Determine a unified grid range based on the target prediction area (e.g., 64 km × 64 km), and clip the corresponding spatial area from the radar image and reanalysis data with this area as the center:

[0045] For radar image data, directly select the original resolution image blocks within the target area;

[0046] The reanalysis data are resampled from the original 0.25° × 0.25° resolution to a grid structure consistent with the radar image (e.g. 128 × 128 or 64 × 64 pixels) using bilinear interpolation and aligned with the radar grid.

[0047] (12) Time series alignment and interpolation: Since radar image data are usually at intervals of 6–10 minutes, and reanalysis data are at intervals of 1 hour, in order to synchronize and form a time-consistent input sequence, the present invention adopts the following time alignment method:

[0048] The reanalysis variables are interpolated on the time axis using linear interpolation to match the radar image time series;

[0049] If there are missing reanalysis data, use forward filling or time mean filling to handle them;

[0050] Construct a unified sliding time window. For example, take the current time t as the benchmark and select the past N time periods (such as 6 frames, a total of 60 minutes) as the input sequence.

[0051] (13) Physical variable standardization and channel fusion: Normalize the input variables to enhance the stability of neural network training and cross-variable compatibility. The standardization methods used include: maximum and minimum normalization or logarithmic transformation of radar variables (Z, Kdp, etc.); and z-score standardization of reanalysis variables, that is:

[0052] ;

[0053] Among them, μ is the historical mean and σ is the standard deviation, both of which are pre-statistically obtained in the training set;

[0054] The final single time frame input tensor is a multi-channel image structure containing multiple radar and reanalysis variable channels;

[0055] (14) Construct input and output sample pairs: Construct a training sample set in a time-series sliding window manner. Each sample consists of an input sequence X and an output sequence Y: the input sequence is the N frames of images before the current moment and the reanalysis variable ; The output sequence is the precipitation image of the next M target time periods The label data can come from interpolated maps of surface rainfall station observations, fused precipitation analysis products (such as IMERG, CaPA), or numerical model forecast precipitation data. All samples are constructed in batches to form a four-dimensional tensor structure, which is input into the model for training. The data sample construction form can be expressed as:

[0056] ;

[0057] in represents the radar image variable, represents the reanalysis variable, Represents an interpolation operation that has been uniformly processed in space and time.

[0058] The radar image data has a temporal resolution of 6 minutes and a spatial resolution of 0.5 km × 0.5 km. The reanalysis data has a temporal resolution of 1 hour and a spatial resolution of 0.25° × 0.25°. After spatial bilinear interpolation, they are aligned with the radar image grid and have an output size of 64 × 64. Radar variables include reflectivity Z, relative phase Kdp, radial velocity V, etc. Reanalysis variables include temperature T, relative humidity RH, wind speed U / V, and geopotential height Zg at the 850 hPa, 700 hPa, and 500 hPa altitude levels.

[0059] The multi-source data preprocessing mechanism of the present invention ensures the consistency of different data sources in physical structure and spatial alignment, providing a stable and reliable input basis for subsequent feature extraction modules and time series modeling modules. It is a key prerequisite for the present invention to achieve end-to-end intelligent precipitation prediction.

[0060] (2) Construction of multi-source feature extraction and fusion module: Feature extraction is performed on the pre-processed multi-source meteorological input data, and structural fusion is achieved in the intermediate latent space. The core goal of this step is to extract the spatiotemporal features of the radar image and the reanalysis background field, compress redundancy and enhance semantic expression capabilities while maintaining the original physical information, and provide a unified format of fused feature input for the subsequent time series modeling module. The feature extraction module adopts a parallel dual-channel neural network structure, corresponding to the radar observation image and the reanalysis meteorological variable respectively. The outputs of the two are upsampled, dimensionally aligned and weightedly fused to form a fused feature tensor with a unified spatial size and channel depth. The feature extraction module adopts a dual-channel neural network structure, including:

[0061] The radar image encoder uses a three-layer convolutional network with the number of channels increasing from 32 to 64 and then to 64, and uses the ReLU activation function. The reanalysis variable encoder uses a 1×1 convolutional and fully connected structure to maintain the independence of the variable channels, and the final output channel number is 64. The two-way encoding results are weightedly fused through the channel attention mechanism, with the weight α being a learnable parameter. The following steps are included:

[0062] (21) Radar image encoder construction: The present invention adopts a convolutional neural network (CNN) structure to construct a radar image encoder module for extracting spatial structural features and echo morphological information from radar sequence images. This module can effectively capture precipitation boundaries, strong echo centers and organizational structures at different scales, and provide high-quality representation for subsequent feature fusion and time series modeling. The radar image encoder consists of multiple convolutional layers, batch normalization layers (BatchNorm) and nonlinear activation functions (ReLU), forming a shallow encoding network. Each convolution unit structure includes: 1. A two-dimensional convolution layer (Conv2D) with a convolution kernel size of 3×3 and a step size of 1, which is used to extract local spatial features; 2. A batch normalization layer (BatchNormalization) that standardizes the activation output in a small batch of samples, effectively alleviating internal covariate offsets, improving training stability, and accelerating convergence; 3. A rectified linear unit (ReLU) activation function, expressed as , used to introduce nonlinear modeling capabilities, enhance the response to echo gradient changes, and prevent the gradient vanishing problem; the above structure is repeated several times, with the number of channels increasing layer by layer, to gradually extract semantic features from weak to strong and from shallow to deep in the image. The input of the radar image encoder is a four-dimensional tensor formed by stacking multiple frames of images, with dimensions: Where T represents the number of time frames, H and W represent the image height and width, and C represents the number of channels (such as variables such as Z and Kdp). The output is the intermediate feature tensor after fusion. .

[0063] (22) Construction of reanalysis variable encoder: The reanalysis data is a structured grid tensor containing three-dimensional tensor information of multiple meteorological variables and multiple pressure layers. A lightweight CNN or a 1×1 convolution + full connection combination is required to construct the encoder. The reanalysis encoder designed in this invention performs convolution processing on each variable channel, without sharing between channels, maintaining the independence of variables, and finally outputs the feature tensor , whose size is consistent with the radar encoder output.

[0064] When the reanalysis input is of low resolution (e.g., 32 × 32), upsampling (bilinear or deconvolution) is used to interpolate it to the same resolution as the radar image (e.g., 64 × 64 or 128 × 128) to ensure matching of spatial dimensions.

[0065] (23) Feature fusion strategy: This paper adopts a feature fusion method based on the channel attention mechanism to perform a weighted combination of radar and reanalysis coding results. The fusion expression is: ,in It is a learnable channel attention weight that adaptively allocates the fusion ratio according to the importance of the input features.

[0066] (24) Fusion result output and structure matching: fused feature tensor The output is a uniform dimensional structure with the following dimensions: , serving as the input interface for the subsequent time series modeling module (ConvLSTM) and error correction network.

[0067] The feature extraction and fusion structure of the present invention supports flexible expansion and is applicable to different radar data structures (such as PPI, CAPPI, and RHI) and various reanalysis data formats (such as ERA5, FNL, and regional self-built modes). It has good versatility and engineering deployment capabilities.

[0068] This step achieves unified spatial resolution, feature channel matching, information structure retention, and trainable weight fusion, providing a high-quality input foundation for dynamic modeling and error control of subsequent models in the time dimension. It is a crucial step in the multi-source information fusion prediction system.

[0069] (3) The time series modeling module is constructed to simulate the evolution of the precipitation system over time. This module takes the spatiotemporal feature tensor generated by the fusion in the second step as input and completes the precipitation intensity prediction of the target time in the future through the recursive propagation of information in the time dimension. Figure 2 As shown in Figure 2, this paper uses a convolutional long short-term memory (ConvLSTM) network as the primary temporal modeling unit. Compared to the traditional LSTM structure, the ConvLSTM uses convolution operations instead of fully connected operations within the unit, which can simultaneously capture dynamic correlations in time and space. It is particularly suitable for temporal modeling tasks such as radar image sequences or spatial tensors.

[0070] The time series modeling module includes two stacked layers of convolutional long short-term memory (ConvLSTM) networks, which are used to model the spatiotemporal evolution characteristics of the precipitation system.

[0071] (31) ConvLSTM structure design and parameter setting: The ConvLSTM network structure consists of multiple layers of stacked ConvLSTM units, each of which is responsible for modeling the spatiotemporal variation trends at different semantic levels. The network structure is shown below:

[0072] ConvLSTM Layer 1: Input channel D, output channel D′, convolution kernel size 3×3, stride 1;

[0073] ConvLSTM Layer 2: further abstracts spatial semantics and increases the receptive field;

[0074] LayerNorm is used for regularization and normalization control.

[0075] At each time step t, ConvLSTM receives the fused feature input , and in the hidden state and memory status Basic update:

[0076] ;

[0077] Finally, a hidden state sequence of a time series is output , used to generate image forecasts of future precipitation.

[0078] (32) Multi-step output mechanism design: The present invention supports continuous output prediction of multiple time steps, that is, the model can output the precipitation field distribution at several future moments at one time, such as the target time of the next 30 minutes, 1 hour, 2 hours, 3 hours, etc.

[0079] To achieve multi-step prediction, the present invention designs the following two strategies:

[0080] a Direct multi-step output method: After ConvLSTM outputs the hidden sequence, it is connected to the decoder network to directly output the precipitation forecast images for all future target time periods (parallel prediction);

[0081] b Autoregressive prediction method: The current output is used as the input for the next moment, and the prediction process is gradually advanced (recursive prediction).

[0082] The present invention prefers direct multi-step output because it has higher computational efficiency and less error accumulation. The corresponding formula is as follows:

[0083] ;

[0084] in For the Predicted precipitation image at time, It is a set of convolutional decoding layers that can be shared or independent.

[0085] (33) Decoder decoding structure design: The decoding module of the present invention adopts a shallow convolutional neural network structure, which consists of 1 to 2 layers of convolution inverse transformation, and is used to transform the hidden state H t Mapped to a precipitation image of target size. The output image dimensions are the same as the original radar Figure 1 To, such as:

[0086] ;

[0087] Where M is the number of output time steps, H and W are the height and width of the image.

[0088] Decoder output types: Real-valued precipitation intensity map, in millimeters per hour (mm / h), used to represent the precipitation intensity per pixel in the prediction area; probabilistic precipitation map, which generates pixel-level precipitation probability by applying a Sigmoid or Softmax function transformation to the decoder output result, and can be used to determine whether precipitation exceeds a set threshold (such as 5mm / h, 10mm / h). The output method can be flexibly configured according to different business needs to support various application scenarios such as quantitative forecasting or graded early warning.

[0089] (34) Training method and loss function selection: The training objective of the model of the present invention is to minimize the error between the predicted image and the real precipitation image. The model training adopts an end-to-end approach, and simulates missing samples in the training set. The optional loss function is: square error loss (MSE):

[0090] ;

[0091] in, is the mean square error loss; M is the number of time steps of the predicted image; is the spatial dimension; is the model's predicted value of precipitation intensity at location (i, j) at time t.

[0092] This step realizes the deep modeling of time series characteristics, captures the nonlinear evolution trend of the precipitation system over time, and has the ability to model processes such as movement, generation, enhancement, and attenuation. It is a key step in the transformation of the present invention from static perception to dynamic prediction.

[0093] (4) Construction of dynamic error correction module: Secondary adjustment is performed on the output preliminary precipitation forecast results to reduce the forecast deviation caused by model nonlinear deviation, input uncertainty or multi-source error transmission, and improve the accuracy and credibility of precipitation images.

[0094] Unlike traditional methods that rely on a single output from the main model, this paper proposes a dynamic error correction network based on a residual learning structure. It learns the error term between the predicted value and the true value and uses this error term to correct the initial output, achieving "fine-tuning enhancement" of the predicted image. The following steps are involved:

[0095] (41) Residual modeling structure design: The dynamic error correction network adopts a lightweight convolutional neural network structure, whose input includes the primary model's preliminary prediction image , historical residual image (such as the residual at the previous moment ) and optional background variables, such as humidity gradients and wind vector changes in the reanalysis field. The network structure may include the following components: Convolutional layer Conv1: input channels are N, output channels are D; BatchNorm + ReLU activation; Convolutional layer Conv2: output channel is 1, mapped to the residual map; residual connection or skip connection mechanism is set in the convolutional structure to improve feature extraction capabilities and network stability. The residual prediction module outputs a correction term image , added to the original predicted image to form the final output:

[0096] ;

[0097] in, Represents the preliminary precipitation image output by the ConvLSTM network; The residual image predicted by the error correction network; This is the final precipitation forecast result after correction.

[0098] (42) Input composition and feature encoding method: In order to enhance the generalization ability of the error correction module, the present invention combines and expands the input of the correction network, specifically including: the current main model output ; Difference between historical predictions and observations (i.e. residual plot): Optional reanalysis background amount, such as , wind field convergence, etc.; all inputs are stacked to form a multi-channel image, which is jointly encoded by the feature extraction module (shallow convolutional network) and then enters the residual prediction subnet to complete the correction term calculation.

[0099] (43) Correction strategy and constraint mechanism: To avoid structural damage caused by excessive correction, the present invention introduces the following mechanisms to control the correction intensity and direction: setting a maximum amplitude limit for the correction term (e.g., ±10 mm / h); adding a regularization term to the loss function to penalize excessive corrections; using a sliding window to smooth the prediction trend and enhance temporal consistency;

[0100] (44) Training method and loss combination: The present invention performs cascade training or stage training on the error correction network and the main model. A combined loss function is used during the training process: ;in, is the predicted value With the observed value The mean square error between is the correction term amplitude control term; where λ is the adjustment weight.

[0101] The correction network can be deployed using standalone training, end-to-end joint training, or online incremental learning, with flexible selection based on actual business system requirements. This step, by constructing a lightweight and stable dynamic error correction network, enables the model to automatically adjust its output based on historical deviations and current status, thereby reducing the effect of error accumulation, improving the accuracy, consistency, and credibility of the predicted image, and enhancing the model's deployability and fault tolerance in actual meteorological business systems. This is one of the key technical innovations of this invention.

[0102] (5) The modality missing completion mechanism is constructed to solve the problem of stable operation of the model when some input data are unavailable (such as radar image missing, obscured, quality abnormal, etc.).

[0103] In actual meteorological business systems, radar observation images may be temporarily missing due to equipment maintenance, communication failures, regional blind spots, or heavy rainfall obstruction. The neural network model is highly sensitive to the input structure. If missing data is forcibly input, it is easy to cause model performance degradation or operation interruption. Therefore, in order to deal with the lack of key modes, the present invention introduces input perception and dynamic completion mechanism, automatically switching to completion mode to ensure the stability of the reasoning process. It includes the following steps:

[0104] (51) Input modality detection mechanism: The present invention embeds modality validity detection logic in the data reading module and performs integrity and quality assessment operations on each frame of radar image: if the radar image data is complete and has a reasonable structure, it is marked as "valid modality";

[0105] If the image is empty, has severe structural defects, or the noise ratio exceeds the threshold, it is marked as "modality missing". The detection module dynamically evaluates the current input at each time step and generates a modality state mark. ;in, : Indicates that radar imagery is available; : Indicates that the radar image is missing or unavailable.

[0106] (52) Alternative feature generation strategy: When When the radar signal is detected, the present invention starts the "modal completion module" to generate an alternative radar signature. This module is designed with two typical implementation methods: static template filling: using historical statistical mean, radar image mean template or blank tensor as alternative input, suitable for lightweight and fast deployment scenarios;

[0107] Dynamic generation and filling: Introduce small auxiliary networks (such as conditional generation networks and mapping networks) to generate dynamic filling according to the characteristics of the current reanalysis variables. Predict estimated radar signatures , which is input into the main model as the completion value. The replacement features must be consistent with the normal radar encoder output in terms of spatial size and channel dimension to ensure the normal operation of the subsequent feature fusion module.

[0108] (53) Fusion module adaptive processing: The feature fusion module receives After replacing the radar features at time t, the channel weighted fusion operation is still performed:

[0109] ;

[0110] in The default weight (0.5) can be used for rapid scenario deployment, or adaptively learned by the attention module. In the absence of radar imagery, the fusion mechanism can further increase the relative weight of the reanalysis variable channels, thereby enhancing the model's ability to rely on the background meteorological field and ensuring the stability and integrity of the forecast results.

[0111] (54) Modality Completion Compatibility Structure: The overall model structure of the present invention supports modality sensing operation. That is, if the radar modality is available, the system will fully run the fusion process. If the radar modality is missing, the input integrity is maintained through a substitution mechanism, and the output result is slightly worse but the structure is stable. The modality completion mechanism is embedded between the model entry and the encoder. It has the characteristics of structural decoupling, clear logic, and easy deployment. It is suitable for a variety of practical business scenarios, including: remote mountainous areas with unstable radar signals, sudden device offline scenarios, and data channel frame loss.

[0112] (55) Joint training strategy: The present invention supports simultaneous training of data samples in both complete and missing modes, that is, constructing some “simulated missing” samples in the training set so that the model can work normally in both modes, thereby improving generalization and robustness.

[0113] This step achieves the goal of ensuring the stable operation of the model even when key observation modalities are unavailable by introducing input modality detection, alternative feature generation, and fusion adaptation mechanisms. It solves the problem of traditional deep learning models being highly dependent on data and prone to failure, and is one of the important technical designs of the present invention to improve engineering practicality and system continuity.

[0114] (6) Construction of prediction output and visualization mechanism: Format the precipitation prediction results output by the model and provide various types of business output interfaces and visualization functions to meet different meteorological business needs, forecast modes, and practical application scenarios of explanatory analysis. The core goal of this step is to convert the intermediate features output by the model (precipitation images at multiple time steps) into standard products that can be directly used in forecasting systems, visualization platforms, or decision support systems. It includes the following steps:

[0115] (61) Real-valued precipitation intensity map generation: After the model output tensor is processed by the decoder, a pixel-by-pixel precipitation intensity image at several future moments is obtained in millimeters per hour (mm / h). The structure is as follows:

[0116] ;

[0117] Where H and W represent the spatial dimensions of the image, and M is the number of output time steps (e.g., 6 steps correspond to the next 3 hours). It can be used to generate cumulative precipitation maps with any time step (10 min / 30 min / 60 min).

[0118] The present invention can optionally configure a cumulative precipitation output mode, that is, superimposing continuous outputs to generate products such as "total precipitation in the next 1 hour" and "total precipitation in the next 3 hours", the format of which is consistent with the requirements of the actual business system.

[0119] (62) Generation of precipitation probability maps and threshold products: To enhance the business interpretability of the model, the present invention supports converting the real-value precipitation output into a probability image via a Sigmoid function or a Softmax classifier, which is used to generate the following types of products: pixel-level precipitation probability maps (such as "the probability of precipitation > 5 mm / h at this point in the next hour"); pixel-by-pixel category maps (such as "no precipitation / light rain / moderate rain / heavy rain" multi-classification maps); zoning statistical maps (such as the percentage of area in a certain area where the probability of precipitation exceeds a certain threshold); the above image structures are highly compatible with actual meteorological service platforms (such as early warning visualization systems and meteorological release platforms).

[0120] (63) Multi-time-step output interface design: To support sliding window continuous forecasting, the model of the present invention supports rolling input of the latest data at fixed time intervals (e.g., 10 minutes), outputs a precipitation forecast image for the next three hours, and automatically updates. Each round of forecast output can be packaged in standard NetCDF format with attached geographic coordinate metadata, making it easy to directly embed into GIS platforms, business systems, or disaster warning interfaces.

[0121] (64) The output results can be connected to existing meteorological business systems, such as forecast duty stations, urban rainstorm warning systems, airport operation and dispatch systems, water conservancy emergency management systems, etc., and can be configured as multi-resolution, multi-variable parallel output to meet the accuracy requirements of different business systems.

[0122] The present invention provides a complete prediction result output mechanism, supports the generation of real-valued, probabilistic and multi-category precipitation images, and has good interpretability and business compatibility, forming an effective bridge from model inference results to meteorological business services, ensuring the goals of practical deployment, rapid integration and continuous optimization of the present invention.

Claims

1. A short-term precipitation prediction method based on the fusion of radar images and reanalysis data, characterized in that: The following steps are involved: (1) Collect radar images and reanalysis data at consecutive times, and generate input data in a unified grid format through spatial interpolation, time alignment and standardization. The radar images include reflectivity Z, differential reflectivity, , phase ratio The temporal resolution is 6-10 minutes, and the spatial resolution is 0.5 km × 0.5 km or 1 km × 1 km. The reanalysis data include temperature T, relative humidity RH, horizontal wind speed u, v and geopotential height. , the temporal resolution is 1 hour, the spatial resolution is 0.25° × 0.25°, and it is aligned to the radar image grid by bilinear interpolation; (2) Validity detection of radar images. When the radar image input is missing, the input structure integrity is maintained by replacing the feature generation module. Then, the dual-channel encoder is used to extract the spatiotemporal features of the radar image and reanalysis data respectively, and the channel attention mechanism is used to perform weighted fusion to generate a fused feature tensor. Validity detection is to dynamically evaluate the current input situation at each time step and generate a modal state marker. ;in, Indicates that radar imagery is available; Indicates that the radar image is missing or unavailable. When , a historical mean template or conditional generation network is used to generate alternative radar features; The dual-channel encoder includes a radar image encoder and a reanalysis variable encoder. The radar image encoder is a three-layer convolutional neural network with the number of channels in the order of 32 → 64 → 64, and the activation function is ReLU. The reanalysis variable encoder uses a 1×1 convolution and fully connected structure to maintain the independence of variable channels, and the number of output channels is 64. The channel attention weight is a learnable parameter, and the fusion formula is: ； in, The channel attention weights can be learned, and the fusion ratio is adaptively allocated according to the importance of the input features. The reanalysis variable encoder independently encodes the variables of each pressure layer 850hPa, 700hPa, and 500hPa, and then splices them and upsamples them to the radar image resolution. (3) The fused features are input into the convolutional long short-term memory neural network ConvLSTM to model the spatiotemporal evolution of the precipitation system and output a preliminary precipitation forecast image for the next 0-3 hours. ConvLSTM is a two-layer stacked structure with 64 input channels, 64 output channels, a convolution kernel size of 3×3, a step size of 1, and activation functions of Tanh and Sigmoid. The time series modeling module adopts a parallel multi-step decoding mechanism to generate 18 frames of precipitation images in the future through a two-layer convolutional network, namely a 3×3 convolution kernel and a 1×1 convolution kernel, with a time interval of 10 minutes and an output size of H×W. (4) Construct a residual learning network to correct the deviation of the initial prediction results based on historical residuals and observation information; the input of the dynamic error correction network includes the initial prediction image, historical residuals, and and reanalysis background variables; The correction formula is: ； in, Represents the preliminary precipitation image output by the ConvLSTM network; The residual image predicted by the error correction network; is the final precipitation forecast result after revision; The loss function includes the mean square error term and the L2 regularization term: ;in, is the predicted value With the observed value The mean square error between is the correction term amplitude control term; where λ is the adjustment weight; (5) Generate precipitation intensity images or probability maps for the next 0-3 hours and adapt them to the business system interface output.

2. The short-term precipitation prediction method based on radar image and reanalysis data fusion according to claim 1, characterized in that: The ConvLSTM unit enables the Dropout mechanism during the training phase and disables it during the inference phase.

3. The short-term precipitation prediction method based on radar image and reanalysis data fusion according to claim 1, characterized in that: In step (3), the model training adopts an end-to-end approach, and simulated missing samples are constructed in the training set; the loss function is: ； in, is the mean square error loss; M is the number of time steps of the predicted image; is the spatial dimension; is the model's predicted value of precipitation intensity at location (i, j) at time t.

4. The short-term precipitation prediction method based on radar image and reanalysis data fusion according to claim 1, characterized in that: In step (5), the forecast output includes a real-valued precipitation intensity map, a probability map converted by Sigmoid or Softmax, and a cumulative precipitation map; the output format is NetCDF or GeoTIFF, with geographic coordinate metadata, and supports sliding window inference, that is, updating the forecast results for the next 3 hours every 10 minutes.

Images