Medium-and-small-scale extreme rainfall forecasting method and device

Abstract

The invention discloses a medium and small scale extreme rainfall forecasting method and device, and the method comprises the steps: S1, obtaining an atmospheric precipitable water GNSS PWV image sequence and a meteorological radar reflectivity image sequence according to the multi-mode meteorological observation data of a to-be-measured region; s2, according to the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence, a water vapor feature map and a radar feature map containing long-range dependence are obtained; s3, deformable cross attention fusion is carried out on the water vapor feature map and the radar feature map, and fusion features after space-time alignment are generated; s4, performing decoding and detail enhancement on the fused features to obtain multi-scale features; and S5, mapping the multi-scale features into a rainfall intensity distribution diagram of a future continuous time period. By adopting the technical scheme of the invention, the accuracy of extreme rainfall forecasting and the image fineness are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of meteorological disaster forecasting technology, specifically relating to a method and apparatus for forecasting small- and medium-scale extreme precipitation. Background Technology

[0002] In recent years, global warming has led to frequent extreme weather events, especially urban flooding and landslides caused by short-duration heavy rainfall, posing a severe challenge to public safety. Traditional extreme precipitation forecasting mainly relies on numerical weather prediction (NWP) models. Although they perform well on a macroscopic scale, they are limited by the approximation of physical parameterization schemes, the uncertainty of the initial field, and the extremely high computational resource consumption. When facing localized, sudden, and rapidly evolving severe convective weather, NWP models often suffer from long "onset-off" times, insufficient spatiotemporal resolution, and low accuracy in fixed-point and quantitative forecasts, making it difficult to meet the timeliness requirements of short-term forecasts (0-6 hours).

[0003] With the improvement of modern meteorological observation systems, multi-source remote sensing data provides rich perspectives for capturing precipitation processes. Global Navigation Satellite Systems (GNSS), by retrieving precipitable water volume (PWV), can monitor the transport and accumulation of water vapor around the clock and is considered a sensitive early warning signal for heavy precipitation. Meanwhile, weather radar, through echo reflectivity, can accurately characterize the three-dimensional spatial structure and intensity of precipitation clouds. However, these two types of data exhibit significant multimodal heterogeneity: GNSS data is typically sparse station data or interpolated low-frequency fields, focusing on the "water vapor environment"; radar data, on the other hand, is a high-resolution grid image, focusing on the "precipitation outcome." Efficiently fusing these two data sources with different physical meanings and vastly different spatial resolutions is key to improving forecast accuracy.

[0004] In recent years, deep learning has made significant progress in short-term weather forecasting. Early convolutional neural networks (CNNs) struggled to capture temporal evolution features, while recurrent neural networks (such as LSTM and GRU), although adept at processing time series, employed a serial computation mode, resulting in low training efficiency and difficulty in capturing historical dependencies of extremely long sequences. The subsequent Transformer architecture, while introducing a global attention mechanism and improving long-range modeling capabilities, suffers from significant quadratic computational complexity, leading to substantial memory consumption when processing high-resolution meteorological image sequences, making efficient inference difficult with limited resources.

[0005] Furthermore, in multimodal fusion, existing methods often employ simple channel concat or element-wise addition. However, atmospheric water vapor fields and precipitation clouds often exhibit spatial misalignment (e.g., wind shear causes water vapor centers and precipitation centers to not coincide). Traditional rigid fusion methods cannot adaptively correct for this geometric deformation, resulting in insufficient feature fusion and limiting the accurate location of extreme precipitation areas. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method and apparatus for forecasting small- and medium-scale extreme precipitation events. By deeply fusing GNSS atmospheric water vapor field and meteorological radar echo data, and utilizing a state-space model with linear complexity and an adaptive feature alignment mechanism, it solves the problems of low efficiency in long-sequence modeling and spatial heterogeneity of multi-source data in traditional methods, thereby achieving efficient and high-precision short-term forecasting of extreme precipitation events.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for forecasting small- and medium-scale extreme precipitation, including Step S1: Based on the multimodal meteorological observation data of the area to be measured, obtain the GNSS PWV image sequence of atmospheric precipitable water and the meteorological radar reflectivity image sequence; Step S2: Based on the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence, obtain water vapor feature map and radar feature map containing long-range dependence. Step S3: Perform deformable cross-attention fusion on the water vapor feature map and the radar feature map to generate spatiotemporally aligned fused features; Step S4: Decode and enhance the details of the fused features to obtain multi-scale features; Step S5: Map the multi-scale features into a precipitation intensity distribution map for future continuous time periods.

[0008] Preferably, in step S2, the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence are input into two independent Bi-Mamba branches, and the spatiotemporal feature encoding with linear complexity is performed using the bidirectional state model SSM to obtain water vapor feature maps and radar feature maps containing long-range dependencies.

[0009] Preferably, the Bi-Mamba branch includes stacked patch embedding layers and several Bi-Mamba modules, which capture global spatiotemporal dependencies in long-sequence meteorological data through forward and backward scanning mechanisms.

[0010] Preferably, in step S3, PWV features are used as the query, radar features are used as the key and value, and a sampling offset is generated by an offset prediction network; the radar features are dynamically sampled based on the sampling offset, and the interaction weights of multimodal features are calculated by a deformable attention mechanism to generate spatiotemporally aligned fused features.

[0011] Preferably, in step S3, the offset prediction network simultaneously receives PWV features and radar features, learns the spatial misalignment information between features through convolutional layers, and outputs a two-dimensional spatial offset for each sampling point.

[0012] Preferably, in step S3, non-grid aligned dynamic interpolation sampling is performed on the radar feature map using the generated spatial offset, so that the sparsely distributed GNSS PWV features can be adaptively aggregated to obtain effective contextual information from the high-resolution radar echo.

[0013] Preferably, in step S4, the fused features are reshaped and upsampled, while local detail features of different receptive fields are captured by a multi-scale convolutional neural network (CNN).

[0014] The present invention also provides a device for forecasting small- and medium-scale extreme precipitation, comprising: The first processing module is used to obtain the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence based on the multimodal meteorological observation data of the area to be measured. The second processing module is used to obtain water vapor feature maps and radar feature maps containing long-range dependence based on the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence. The third processing module is used to perform deformable cross-attention fusion on water vapor feature map and radar feature map to generate spatiotemporally aligned fused features. The fourth processing module is used to decode and enhance the details of the fused features to obtain multi-scale features; The fifth processing module is used to map multi-scale features into precipitation intensity distribution maps for future continuous time periods.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Bi-Mamba (bidirectional state-space model) is used to replace traditional RNN (such as LSTM) or Transformer. Mamba has a linear computational complexity and can process longer historical meteorological sequences without significantly increasing memory consumption, thereby capturing longer-term weather evolution patterns. Its inference speed is significantly better than traditional attention mechanism models.

[0016] (2) A deformable cross-attention mechanism is introduced to break the rigid fusion limitation of traditional "pixel-by-pixel correspondence". By adaptively learning the spatial offset between the water vapor field and the precipitation cloud, the model can "correct" the spatial mismatch caused by wind shear or sensor geometric differences, which significantly improves the physical consistency of fused features and forecast accuracy.

[0017] (3) By using a dual-stream architecture to process GNSS water vapor (macro-environment) and radar echo (micro-structure) respectively, and combining it with a multi-scale CNN decoder, the model can not only grasp the large-scale precipitation trend, but also finely depict the landing area and intensity of the heavy precipitation center, effectively solving the problems of "missed reports" and "false reports" in extreme precipitation forecasts.

[0018] (4) Thanks to the efficient inference capabilities of the Mamba architecture, this invention can complete precipitation prediction for several hours in a very short time, meeting the extremely high timeliness requirements of meteorological operations for early warning of extreme weather disasters. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the small-scale extreme precipitation forecasting method in an embodiment of the present invention; Figure 2 This is a visualization diagram of the multimodal input data constructed in the example of the present invention, wherein (a) shows the spatiotemporal distribution map of GNSS PWV (atmospheric precipitable water) generated by interpolation, reflecting the macroscopic water vapor background; and (b) shows the combined reflectivity image of meteorological radar at the same time, reflecting the fine structure of precipitation cloud clusters. Figure 3 This is a schematic diagram illustrating the principle of the deformable cross-attention fusion mechanism of the present invention; Figure 4 This is a comparison chart of the predicted spatial distribution of future precipitation intensity output by the model of this invention and the actual values. Figure 5 This is a schematic diagram comparing the prediction performance of the model of this invention with that of existing mainstream models; Figure 6 The bar chart shows the POD index of the model of this invention and the traditional model under different precipitation thresholds. Figure 7 The graph shows the FAR index curves of the model of this invention and the traditional model under different precipitation thresholds. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 like Figure 1 As shown, this invention provides a method for forecasting small- and medium-scale extreme precipitation, comprising: S1: Acquire multimodal meteorological observation data of the area to be measured and construct a spatiotemporal sequence input. This step aims to acquire high-quality GNSS images of precipitable water (PWV) and images of weather radar reflectivity.

[0024] S1-1: Construction of GNSS PWV image sequences First, raw observation data (carrier phase and pseudorange) from GNSS satellites are acquired. The GNSS observation data is then processed based on a dual-frequency ionospherically free precision positioning (PPP) model.

[0025] The dual-frequency observation equation for the epoch of a GNSS station is: ; ; The parameter definitions are consistent with those of the conventional PPP model. After eliminating the first-order ionospheric delay through a dual-frequency ionospheric-free combination, a random walk model is used to estimate the total zenithal-tropospheric delay (ZTD).

[0026] Next, the dry zenith delay (ZHD) was calculated using the Saastamoinen model, and the wet zenith delay (ZWD) was separated: ; The final inversion yields the GNSS PWV: ; Since GNSS stations are typically unevenly distributed, this embodiment uses Kriging interpolation to interpolate the PWV data from discrete stations into a regular grid image, generating a time-continuous PWV image sequence (e.g., ...). Figure 2 (as shown in (a)). This figure reflects the macroscopic water vapor transport and accumulation processes before and after extreme precipitation.

[0027] S1-2: Construction of Weather Radar Reflectivity Image Sequence Volume scan data from multiple Doppler weather radars were collected, quality control and spatial stitching were performed, and regional composite reflectivity (CR) was generated. ; Based on the combined reflectance, precipitation intensity can be further estimated using the Zr relationship, such as... Figure 2 As shown in (b), this data reflects the fine spatial structure of precipitation clouds.

[0028] S1-3: Data Alignment and Normalization The GNSS and radar data are aligned to a uniform time step (one frame every 6 minutes) using a cubic spline interpolation algorithm, and the data is normalized to the [0,1] interval to construct the final model input tensor: and ; S2: Construct a two-stream Bi-Mamba feature extraction network to extract spatiotemporal features. To efficiently process long-series meteorological data, this example abandons the traditional RNN and Transformer architectures and adopts a dual-stream Bi-Mamba network based on the State Space Model (SSM).

[0029] Will and Two independent Bi-Mamba encoder branches are input separately. Each branch first maps the image blocks into a one-dimensional vector sequence through a Patch Embedding layer, and then enters a stacked Bi-Mamba module.

[0030] Within each Mamba module, feature mapping is performed using the discretized state-space equations: ; ; in, and These are the discretized system parameters. Enter the current time. This is in a hidden state. To capture the anisotropy of weather systems, the Bi-Mamba module performs bidirectional scans, forward and backward. The forward scan captures the evolutionary trends of weather systems, while the backward scan enhances the contextual relevance of historical information.

[0031] After step S2, water vapor feature maps containing long-range dependence are obtained respectively. and radar feature map .

[0032] S3: Multimodal fusion based on deformable cross-attention Due to wind field effects, there is often a spatial misalignment between the atmospheric water vapor center (high PWV area) and the precipitation area (high radar echo area). Traditional direct concat stitching cannot solve this problem. Figure 3 As shown, the deformable cross-attention fusion process in this embodiment includes: S31: Offset Prediction PWV features with low resolution but containing macroscopic trends As a query, with high-resolution radar features Serving as both key and value, the concatenation of these two values ​​is fed into a lightweight convolutional network (OffsetPredictor) to predict the sampling offset Δp for each spatial location. ; The offset Δp implies the cloud's motion vector and deformation information.

[0033] S32: Dynamic Sampling & Fusion Using the generated offset Δp, non-mesh bilinear interpolation sampling is performed on the radar feature map. For position p on the feature map, the model no longer focuses solely on point p itself, but rather on the features at position p+Δp: ; in, The attention weights are K, where K is the number of sampling points. Through this mechanism, the model can adaptively align PWV features with the "drifted" radar echo features, thereby generating fused features with stronger physical consistency.

[0034] S4: Multi-scale Decoding and Detail Enhancement The fused features Input decoder.

[0035] First, the spatial dimensions of the feature map are recovered through deconvolution or upsampling operations. Then, a multi-scale CNN module is designed, containing parallel 3×3 standard convolutions (extracting local texture), 5×5 convolutions (extracting mesoscale structure), and dilated convolutions with a dilation rate of d=2 (extracting large-scale background). Features from different branches are concatenated to simultaneously recover both macroscopic rainbands and microscopic centers of intense precipitation in extreme precipitation events.

[0036] S5: Generating Prediction Results and Model Training Using a 1×1 convolutional prediction head, multi-scale features are mapped to precipitation intensity distribution maps for future continuous time periods (e.g., the next 0-2 hours). .

[0037] During the model training phase, ground-based rain gauge measured data or radar quantitative precipitation estimation (QPE) products calibrated by rain gauges are used as ground truth labels to construct a loss function for end-to-end supervised learning optimization of grid parameters.

[0038] Implementation Result Analysis: To verify the effectiveness of this invention, tests were conducted using historical extreme precipitation examples, and the results are as follows: Figure 4 , Figure 5 and Figure 6 As shown.

[0039] S5-1: Qualitative Analysis ( Figure 4 ): Figure 4 The image shows a comparison between the model's predicted precipitation intensity for the next hour (right) and the actual measured values ​​from ground automatic weather stations (left). It can be seen that the model accurately predicts the spatial characteristics of the heavy precipitation center distribution, and the direction of the rainband closely matches the actual situation, demonstrating the effectiveness of the deformable attention mechanism in correcting spatial misalignment.

[0040] S5-2: Comparative Analysis ( Figure 5 ): Figure 5 The predictive performance of the model of this invention was compared with that of the traditional model. Under the same forecast lead time, the image generated by the traditional model was relatively blurry and lost the extreme value information of the center of heavy precipitation (i.e., the "smoothing effect"); while the present invention (Bi-Mamba-DA) benefits from the multi-scale CNN decoder and the strong modeling capability of Mamba, generating radar echo textures with clear and sharp edges, which can better reflect the true structure of extreme weather.

[0041] S5-3: Quantitative Analysis ( Figure 6 ,7): Figure 6 The model's classification forecast indices are shown under different precipitation thresholds.

[0042] POD (Pick-up Rate): The bar chart shows that the POD of this model is higher than that of the comparison model at all levels, especially in the extreme precipitation range of >30mm / h, where the advantage is more obvious.

[0043] FAR (False Alarm Rate): The curve shows that the FAR of this model remains at a low level.

[0044] This invention effectively solves the problems of spatiotemporal alignment of multi-source meteorological data and low efficiency of long sequence modeling by introducing a dual-stream Mamba architecture and a deformable cross-attention mechanism, and significantly improves the accuracy of short-term forecasts of extreme precipitation.

[0045] Example 2 The present invention also provides a device for forecasting small- and medium-scale extreme precipitation, comprising: The first processing module is used to obtain the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence based on the multimodal meteorological observation data of the area to be measured. The second processing module is used to obtain water vapor feature maps and radar feature maps containing long-range dependence based on the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence. The third processing module is used to perform deformable cross-attention fusion on water vapor feature map and radar feature map to generate spatiotemporally aligned fused features. The fourth processing module is used to decode and enhance the details of the fused features to obtain multi-scale features; The fifth processing module is used to map multi-scale features into precipitation intensity distribution maps for future continuous time periods.

[0046] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for forecasting small- and medium-scale extreme precipitation, characterized in that, include Step S1: Based on the multimodal meteorological observation data of the area to be measured, obtain the GNSS PWV image sequence of atmospheric precipitable water and the meteorological radar reflectivity image sequence; Step S2: Based on the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence, obtain water vapor feature map and radar feature map containing long-range dependence. Step S3: Perform deformable cross-attention fusion on the water vapor feature map and the radar feature map to generate spatiotemporally aligned fused features; Step S4: Decode and enhance the details of the fused features to obtain multi-scale features; Step S5: Map the multi-scale features into a precipitation intensity distribution map for future continuous time periods.

2. The method for forecasting small- and medium-scale extreme precipitation as described in claim 1, characterized in that, In step S2, the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence are input into two independent Bi-Mamba branches. The spatiotemporal feature encoding with linear complexity is performed using the bidirectional state model SSM to obtain water vapor feature maps and radar feature maps containing long-range dependencies.

3. The method for forecasting small- and medium-scale extreme precipitation as described in claim 2, characterized in that, Bi The Mamba branch contains stacked patch embedding layers and several Bi-Mamba modules. These Bi-Mamba modules capture global spatiotemporal dependencies in long-sequence meteorological data through forward and backward scanning mechanisms.

4. The method for forecasting small- and medium-scale extreme precipitation as described in claim 3, characterized in that, In step S3, PWV features are used as the query, radar features are used as the key and value, and a sampling offset is generated by an offset prediction network. Based on the sampling offset, the radar features are dynamically sampled, and the interaction weights of multimodal features are calculated by a deformable attention mechanism to generate spatiotemporally aligned fused features.

5. The method for forecasting small- and medium-scale extreme precipitation as described in claim 4, characterized in that, In step S3, the offset prediction network simultaneously receives PWV features and radar features, learns the spatial misalignment information between features through convolutional layers, and outputs a two-dimensional spatial offset for each sampling point.

6. The method for forecasting small- and medium-scale extreme precipitation as described in claim 5, characterized in that, In step S3, the generated spatial offset is used to perform non-grid aligned dynamic interpolation sampling on the radar feature map, so that the sparsely distributed GNSSPWV features can adaptively aggregate effective contextual information in the high-resolution radar echo.

7. The method for forecasting small- and medium-scale extreme precipitation as described in claim 6, characterized in that, In step S4, the fused features are reshaped and upsampled, while local detail features of different receptive fields are captured by a multi-scale convolutional neural network (CNN).

8. A device for forecasting small- and medium-scale extreme precipitation, characterized in that, include: The first processing module is used to obtain atmospheric precipitable water GNSSPWV image sequences and meteorological radar reflectivity image sequences based on multimodal meteorological observation data of the area to be measured. The second processing module is used to obtain water vapor feature maps and radar feature maps containing long-range dependence based on the atmospheric precipitable water GNSS PWV image sequence and the meteorological radar reflectivity image sequence. The third processing module is used to perform deformable cross-attention fusion on water vapor feature map and radar feature map to generate spatiotemporally aligned fused features. The fourth processing module is used to decode and enhance the details of the fused features to obtain multi-scale features; The fifth processing module is used to map multi-scale features into precipitation intensity distribution maps for future continuous time periods.

Images