Quantitative precipitation inversion method and system based on brown bridge condition diffusion and multi-modal fusion

Abstract

The present application provides a quantitative precipitation inversion method and system based on Brown bridge condition diffusion and multi-modal fusion, belonging to the technical field of radar image sequence and quantitative precipitation estimation, a multi-modal paired data set containing weather radar volume scan data, ground precipitation records and ERA5 reanalysis data is constructed, and spatio-temporal alignment and preprocessing are completed; three types of heterogeneous data are mapped to a unified latent space to provide standardized input for the diffusion model; a Brown bridge diffusion model is trained with radar features as the starting point, precipitation latent representation as the endpoint, and ERA5 features as the condition; a U-Net noise prediction network with a multi-scale condition injection mechanism is designed, and end-to-end training is performed through a joint loss function; finally, based on the trained model, a precipitation estimation map is generated for the target time, and precipitation inversion is realized. The present application solves the technical problems of insufficient multi-source heterogeneous data fusion, high cross-modal mapping uncertainty and single condition information injection mechanism, and realizes intelligent inversion of quantitative precipitation with high precision and high physical consistency.

Description

Technical Field

[0001] This invention belongs to the field of radar image sequence and quantitative precipitation estimation technology, specifically, it relates to a quantitative precipitation inversion method and system based on Brownian bridge conditional diffusion and multimodal fusion. Background Technology

[0002] Accurate quantitative precipitation estimation (QPE) is a core task in meteorological disaster prevention and mitigation. Currently, precipitation retrieval based on Doppler weather radar is the main technical means. However, from a physical mechanism perspective, the occurrence and evolution of precipitation not only depend on the microphysical structure within clouds observed by radar, but are also significantly modulated by the atmospheric environment field, such as water vapor transport, thermal instability, and vertical motion. Single-mode radar observations are insufficient to fully characterize the coupling mechanisms of these multi-scale, multi-physical processes, especially in scenarios of severe convective precipitation, where models often deviate from the extreme precipitation intensity and location due to the lack of atmospheric background constraints.

[0003] While existing deep learning methods have made some progress in radar precipitation retrieval, the following technical bottlenecks still exist: Insufficient fusion of multi-source information: Existing models mostly use radar data as the sole or dominant input, failing to fully utilize the physical information such as temperature, humidity, and wind field contained in atmospheric reanalysis data (such as ERA5). How to effectively fuse radar structural information with atmospheric thermodynamic background is a key challenge to improve inversion accuracy.

[0004] High uncertainty in cross-modal mapping: The mapping from radar echoes to precipitation fields exhibits high nonlinearity and uncertainty. Traditional "one-step" regression models struggle to characterize the distributional properties of this complex mapping. While standard diffusion models can generate data from noise, their starting point is unstructured random noise, making them difficult to learn and prone to producing results that do not conform to physical laws.

[0005] The conditional information injection mechanism is too simple: In the conditional generation model, if multiple source conditions are simply spliced ​​together at the input layer, the conditional information in the deep network is easily diluted and it is difficult to play a guiding role throughout the generation process, resulting in a lack of physical consistency in the generation results.

[0006] Therefore, how to construct a model that can efficiently integrate radar observations and atmospheric environmental fields and transform precipitation inversion into a deterministic conditional mapping rather than an unconstrained generation model is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a quantitative precipitation inversion method and system based on Brownian bridge conditional diffusion and multimodal fusion. This invention transforms the precipitation inversion task into a Brownian bridge diffusion process that starts with radar features, ends with the potential representation of precipitation, and is conditional on the atmospheric environment field, thereby achieving the fusion of multi-source information and the generation of a physically consistent precipitation field in the potential space.

[0008] This invention is achieved through the following technical solution: A quantitative precipitation inversion method based on Brownian bridge conditional diffusion and multimodal fusion: The method specifically includes the following steps: S1, Construction and Spatiotemporal Alignment of Multimodal Conditional Dataset: Acquire 3D weather radar volume scan data, ground automatic weather station precipitation records and ERA5 atmospheric reanalysis data within the target area and time period, perform spatiotemporal alignment and preprocessing, and construct radar-precipitation-ERA5 multimodal paired dataset. S2, Unified Latent Space Coding for Multi-Source Heterogeneous Data: Maps three types of heterogeneous data—radar, ERA5, and precipitation—to a dimension-aligned unified latent space, providing a standardized input format for the Brownian Bridge diffusion model; S3. Construct and train a Brownian bridge conditional diffusion generation model: Train a Brownian bridge diffusion model in the potential space, starting with radar features, ending with the potential representation of precipitation, and conditional with ERA5 features; S4. Design of Multi-Scale Conditional Injection and Noise Prediction Network: Design a U-Net noise prediction network that includes a multi-scale conditional injection mechanism, injecting spatial splicing mechanism, cross-attention mechanism and time-conditional joint modulation; S5. Design a joint loss function to perform end-to-end training of the model: S6. Model Inference and Application: After training is completed, the model is applied to actual precipitation inversion. For any given target time, the precipitation estimation map is output as the final inversion result.

[0009] Furthermore, in S1, Radar data processing: Using the multi-frame sequence input method, a continuous 6-frame three-dimensional radar data consisting of the target time and its 5 preceding times is constructed; ERA5 data processing: Six variables closely related to precipitation physical processes were selected, including relative humidity, specific cloud liquid water content, specific humidity, temperature, zonal wind, and meridional wind; each variable was taken at 5 isobaric levels, for a total of 30 channels; all variables were resampled to the same spatial grid as the radar data through bilinear interpolation and aligned with the radar samples through a nearest neighbor time matching strategy. Precipitation labeling: Hourly precipitation records from automatic weather stations are collected. After screening for climate extremes, phase consistency tests, and spatial consistency checks, two types of labels are generated: a two-dimensional precipitation grid map obtained through Kriging interpolation, and the original station observation dictionary set.

[0010] Furthermore, S2 includes, S2.1, Radar Feature Extraction and Temporal Fusion: A pre-trained 3D mask autoencoder is used as the radar encoder to extract the spatial structure features of a single frame of radar data; for a continuous 6-frame radar sequence, the temporal dependency model of each spatial location is independently performed through the point-by-point time Transformer module, and the radar fusion features that integrate dynamic evolution information are output. S2.2, ERA5 Conditional Encoding: Design a lightweight convolutional encoder to map 30-channel ERA5 data to the same latent spatial dimension as radar features to obtain conditional features; S2.3, Precipitation label encoding: A variational autoencoder (VAE) is used to compress high-resolution precipitation images into continuous and regularized latent representations. The parameters of the VAE are frozen after separate pre-training.

[0011] Furthermore, S3 includes, S3.1, Forward Diffusion Construction: Define a random bridging path starting from radar fusion features and ending with VAE-encoded precipitation potential representation, set the total number of diffusion steps and adopt a cosine noise scheduling strategy; during the training phase, define interpolation coefficients and noise intensity for random sampling time steps, and obtain intermediate states based on the Brownian bridge forward process. S3.2, Inverse Denoising Process: Construct a denoising network with ERA5 conditions as input, learn the inverse process of recovering from the noisy intermediate state to the target state; by predicting and removing the noise added during the forward diffusion process, gradually generate the potential representation of precipitation.

[0012] Furthermore, in S4, The noise prediction network adopts a U-Net architecture with symmetric encoding and decoding. It extracts high-dimensional semantic features step by step through the downsampling path, and then restores spatial details through the upsampling path. Skip connections are used to compensate for the loss of details. Injection of spatial stitching mechanism: In each downsampling and upsampling stage of the network, the ERA5 conditional features are adjusted to the current feature map size by bilinear interpolation through spatial stitching, and then convolved and fused with the main features in the channel dimension to guide the generation of local precipitation structures; A cross-attention mechanism is introduced at the bottleneck layer, using the main feature as the query and the condition feature as the key to achieve adaptive aggregation of global physical information; Time-conditional joint modulation: The diffusion time step is combined with the global statistics of the conditional features. The feature linear modulation FiLM parameters are generated by the multilayer perceptron. The main features are scaled and shifted channel by channel, so that the network can dynamically adjust its dependence on ERA5 information according to the denoising stage.

[0013] Furthermore, in S5, The total loss function is defined as a weighted sum of the diffusion loss and the image reconstruction loss:

[0014] Among them, diffusion loss To minimize the mean square error between the noise predicted by the noise prediction network and the actual forward noise; Image reconstruction loss The method involves back-calculating the potential precipitation estimate based on the predicted noise, reconstructing the precipitation image using a frozen VAE decoder, and calculating the mean square error between the reconstructed image and the actual precipitation image.

[0015] A quantitative precipitation inversion system based on Brownian bridge conditional diffusion and multimodal fusion; The system includes a module for constructing and spatiotemporally aligning a multimodal conditional dataset, a module for unified latent spatial coding of multi-source heterogeneous data, a Brownian bridge conditional diffusion generation model, a module for designing a multi-scale conditional injection and noise prediction network, a joint loss module, and an application module. The module for constructing and aligning the multimodal conditional dataset is used to acquire three-dimensional weather radar volume scan data, precipitation records from ground automatic weather stations, and ERA5 atmospheric reanalysis data within the target area and time period, perform spatiotemporal alignment and preprocessing, and construct a radar-precipitation-ERA5 multimodal paired dataset. The unified latent space coding module for multi-source heterogeneous data is used to map three types of heterogeneous data—radar, ERA5, and precipitation—to a dimension-aligned unified latent space, providing a standardized input format for the Brownian bridge diffusion model. The Brownian bridge conditional diffusion generation model is used to train a Brownian bridge diffusion model in the potential space, starting with radar features, ending with the potential representation of precipitation, and conditional with ERA5 features. The multi-scale conditional injection and noise prediction network design module is used to design a U-Net noise prediction network that includes a multi-scale conditional injection mechanism, injecting spatial splicing mechanism, cross-attention mechanism and time-conditional joint modulation. The joint loss module performs end-to-end training of the model: The application module is used to perform actual precipitation inversion after training is completed; for any given target time, it outputs a precipitation estimation map as the final inversion result.

[0016] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0017] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: The inversion accuracy is greatly improved: By constraining the generation process to a deterministic mapping path from radar to precipitation through the Brownian bridge mechanism, the uncertainty of the generation space is significantly reduced. It is significantly better than existing models in terms of root mean square error, correlation coefficient and critical success index for heavy precipitation.

[0019] Efficient fusion of multi-source physical information: For the first time, the ERA5 multivariable and multi-isobaric atmospheric background field is introduced as a condition into the diffusion model, and the physical information plays a continuous role in the entire generation process through a multi-scale injection mechanism, which enhances the physical consistency of the inversion results.

[0020] Enhanced generation quality and robustness: The joint optimization strategy balances the distribution consistency of the latent space with the pixel accuracy of the image space, enabling the model to accurately reproduce the peak structure of the heavy precipitation center while maintaining the overall shape of the precipitation system, effectively avoiding the problems of over-smoothing or positional shift in traditional methods.

[0021] The framework is highly versatile: the proposed Brownian bridge conditional diffusion framework can be extended to other cross-modal meteorological variable inversion tasks (such as inverting surface temperature from satellite observations and reconstructing high-resolution wind fields from reanalysis data), and has good scalability.

[0022] The two-stage training framework and Brownian bridge conditional diffusion model proposed in this invention effectively solve technical problems such as insufficient fusion of multi-source heterogeneous data, high uncertainty in cross-modal mapping, and a single conditional information injection mechanism, and realizes high-precision and high-physical-consistency quantitative precipitation intelligent inversion. Attached Figure Description

[0023] Figure 1 This is a diagram illustrating the overall architecture of the method of the present invention; Figure 2 This is a structural diagram of the precipitation feature encoder model of the present invention; Figure 3 This is a structural diagram of the precipitation feature decoder model of the present invention; Figure 4This is a flowchart of the atmospheric reanalysis data feature encoder-decoder structure of the present invention; Figure 5 This is a schematic diagram of the Brownian bridge diffusion process of the present invention. Detailed Implementation

[0024] 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.

[0025] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0026] Example 1, combined with Figures 1 to 5 This embodiment provides a quantitative precipitation inversion method based on Brownian bridge conditional diffusion and multimodal fusion, using a precipitation inversion task in Guangdong Province, China as an example for detailed explanation. The method specifically includes the following steps: S1, Construction and Spatiotemporal Alignment of Multimodal Conditional Datasets: We acquire 3D weather radar volume scan data, precipitation records from ground automatic weather stations, and ERA5 atmospheric reanalysis data for the target area and time period. We then perform rigorous spatiotemporal alignment and preprocessing to construct a spatiotemporally aligned radar-precipitation-ERA5 multimodal paired dataset.

[0027] (1) Radar data processing: Using the multi-frame sequence input method, a continuous 6-frame three-dimensional radar data (each frame contains 6 height layers) consisting of the target time and its 5 preceding times (12 minutes apart) is constructed as the main feature source of the model.

[0028] Obtain Doppler weather radar contour mosaic products covering Guangdong Province (longitude 108.50°E–117.50°E, latitude 19.04°N–26.04°N). Select height layers 1–6 closely related to precipitation processes. For each target time (00:00 and 30:00 per hour), construct a contour mosaic based on the target time. and its 5 previous moments ( , , , , A continuous sequence of 6 radar frames (minutes) is formed. Each frame contains reflectivity values ​​for 6 altitude levels, with a spatial size of 700×900 pixels.

[0029] (2) ERA5 data processing: Six variables closely related to precipitation physical processes were selected (relative humidity, specific cloud liquid water content, specific humidity, temperature, zonal wind, and meridional wind), and five isobaric surface levels were taken for each variable (30 channels in total). All variables were resampled to the same spatial grid (700×900) as the radar data through bilinear interpolation, and aligned with the radar samples through the nearest neighbor time matching strategy as atmospheric background conditions.

[0030] ERA5 reanalysis data for the corresponding time period and region were obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF). The following six variables closely related to precipitation physical processes were selected: relative humidity, specific cloud liquid water content, specific humidity, temperature, zonal wind, and meridional wind. Each variable was analyzed using five standard isobaric levels (300 hPa, 500 hPa, 700 hPa, 850 hPa, and 925 hPa), for a total of 30 channels. All variables were resampled to the same 700×900 grid as the radar data using bilinear interpolation. Time alignment employed a nearest neighbor matching strategy: for precipitation labels at 00 minutes, the ERA5 analysis field at the corresponding hour was used; for precipitation labels at 30 minutes, the ERA5 data from the hour preceding that hour was used.

[0031] (3) Precipitation label processing: Two types of labels are retained: a two-dimensional precipitation grid map obtained by Kriging interpolation (for global monitoring) and the original station observation dictionary set (for precise station monitoring).

[0032] Hourly precipitation records from 2,856 automatic weather stations in Guangdong Province were collected. After screening for climate extremes, phase consistency tests, and spatial consistency checks, two types of labels were generated: one is a two-dimensional precipitation grid map (700×900) generated by Kriging interpolation, used for global monitoring; the other is a dictionary set of original station observations, used for precise monitoring of the station's neighborhood.

[0033] S2, Unified Latent Spatial Coding of Multi-Source Heterogeneous Data: To facilitate the conditional generation of the Brownian bridge diffusion model, this step aims to map the three heterogeneous data types—radar, ERA5, and precipitation—to a unified latent space with strictly aligned dimensions, thus providing a standardized input format for the Brownian bridge diffusion model.

[0034] (1) Radar feature extraction and temporal fusion: A pre-trained 3D mask autoencoder (3DMAE) is used as the radar encoder to extract the spatial structure features of a single-frame radar data. For a continuous 6-frame radar sequence, the temporal dependency model of each spatial location is independently performed by the point-by-point time Transformer module, and the radar fusion feature that integrates dynamic evolution information is output as the starting point of the Brownian bridge diffusion process.

[0035] The parameters of the pre-trained 3D mask autoencoder (3DMAE) are frozen and used as a fixed radar feature extractor. The six radar sequences constructed in S1 are input into this encoder, resulting in six feature sequences, each containing 2408 image patch tokens, each with a dimension of 768. To preserve temporal information, learnable temporal location encodings are added to each time step. Subsequently, the spatial and batch dimensions are merged, and a point-to-point temporal Transformer module containing two layers of Transformer blocks (each with eight self-attention heads) is independently applied to each spatial location to learn the feature evolution of that location over six historical time points. The output of the last frame is taken as the radar fusion feature after incorporating temporal information. F The fused characteristic will serve as the starting point for the Brownian bridge diffusion process. x 0.

[0036] (2) ERA5 conditional coding: A lightweight convolutional encoder is designed to map 30-channel ERA5 data to the same latent spatial dimension as radar features to obtain conditional features, which are used to provide continuous physical guidance during the diffusion process; Design and implement a lightweight convolutional encoder, with the following structure: Figure 2 As shown. The encoder input is 30 channels of ERA5 data (shape: B The feature size (30×700×900) comprises four downsampling stages. Each stage consists of two 3×3 convolutional layers with a stride of 2, with the number of channels increasing sequentially to 128, 256, 512, and 768. Each convolutional layer is followed by a GELU activation function and residual connections, and then batch normalization is performed. After four downsampling passes, the spatial size is reduced from 700×900 to 44×57. After center cropping to 43×56 and flattening, a conditional feature with dimensions completely consistent with the radar feature size is obtained. c (shape is) B ×2408×768).

[0037] (3) Precipitation label encoding: A variational autoencoder (VAE) is used to compress high-resolution precipitation images into a continuous and regularized latent representation, which serves as the endpoint of the Brownian bridge diffusion process. The parameters of the VAE are frozen after separate pre-training to ensure the stability of the latent space.

[0038] The precipitation field is compressed and encoded using a variational autoencoder (VAE), with the structure as follows: Figure 5 As shown. The VAE encoder consists of four 3×3 convolutional layers with a stride of 2, which process the input precipitation image ( B ×1×700×900) progressively downsampled to the potential space ( B ×768×43×56). The latent mean is predicted using two independent convolutional layers. μp Sum of logarithms and variances σp 2. Precipitation potential representation obtained through reparameterized sampling zT This representation will serve as the endpoint of the Brownian bridge diffusion process. The VAE decoder employs a symmetrical upsampling structure, progressively restoring the original resolution through bilinear interpolation and 3×3 convolution. The VAE module is pre-trained separately for 200 epochs before the diffusion model training, with the optimization objective being a weighted sum of the reconstruction loss (mean squared error) and the KL divergence loss (weights of 1.0 and 0.01, respectively). All parameters are frozen after pre-training.

[0039] S3. Construct and train the Brownian bridge conditional diffusion generative model: The purpose of this step is to train a Brownian bridge diffusion model in the potential space, starting with radar features, ending with the potential representation of precipitation, and conditioned on ERA5 features.

[0040] (1) Forward diffusion construction: Define a random bridging path that starts with radar fusion features and ends with VAE-encoded precipitation potential representation, and set the total number of diffusion steps. T =1000, using a cosine noise scheduling strategy. During the training phase, for the random sampling time step... t ∈[0, T Define interpolation coefficients. and noise intensity (according to calculate, (is a constant).

[0041] The intermediate states are directly calculated using the formula for the forward process of the Brownian bridge. ;

[0042] in For radar fusion features, A potential representation of precipitation encoded by VAE; It follows a Gaussian distribution with mean value of linear interpolation at both endpoints that varies with time.

[0043] (2) Reverse denoising process: The goal of the model is to learn from noisy intermediate states. xt The reverse process to restore the target state. This process is modeled as a denoising network with ERA5 conditions as input, which gradually generates a potential representation of precipitation by predicting and removing noise added in the forward process.

[0044] S4. Multi-scale conditional injection and noise prediction network design: In order to enable ERA5 atmospheric background information to continuously guide the generation process at different scales and different denoising stages, a U-Net noise prediction network with a multi-scale conditional injection mechanism is designed.

[0045] (1) Noise prediction network design (network main structure): The U-Net architecture with encoding-decoding symmetry is adopted. High-dimensional semantic features are extracted step by step through the downsampling path, and spatial details are restored through the upsampling path. Skip connections are used to make up for the loss of details.

[0046] use Figure 4 The U-Net architecture shown serves as a noise prediction network. The network consists of a downsampling path, a bottleneck layer, and an upsampling path. The downsampling path comprises four stages, each consisting of two residual convolutional blocks (3×3 convolution + group normalization + SiLU activation + residual connection) and a downsampling convolution with a stride of 2, with channel numbers of 256, 512, 1024, and 1024 respectively. The bottleneck layer introduces two Transformer modules, enhancing global modeling capabilities through self-attention and achieving global semantic alignment with ERA5 conditional features through cross-attention. The upsampling path is symmetrical to the encoder, progressively restoring spatial resolution through bilinear interpolation and fusing features from the corresponding encoding stages through skip connections.

[0047] (2) Multi-scale conditional injection: including spatial splicing injection, cross-attention injection and time-conditional joint modulation; Spatial stitching injection: In each downsampling and upsampling stage of the network, ERA5 conditional features are spatially stitched together. c After being adjusted to the current feature map size by bilinear interpolation, it is convolved and fused with the main feature in the channel dimension to guide the generation of local precipitation structures.

[0048] Cross-attention injection: A cross-attention mechanism is introduced at the bottleneck layer, using the main feature as the query and the condition feature as the key / value pair to achieve adaptive aggregation of global physical information.

[0049] Time-conditional joint modulation: simultaneously controlling the diffusion time step t The sinusoidal positional encoding is concatenated with the global average pooling result of the conditional features, and then the scale parameters of Feature Linear Modulation (FiLM) are generated by a multilayer perceptron. γ and offset parameters β The output of each residual block is modulated channel by channel (the main features are scaled and offset channel by channel), as shown in the formula. F ′= γ ⊙ F + β This enables the network to dynamically adjust its dependence on ERA5 information based on the denoising stage.

[0050] S5. Joint Loss Function Optimization: Design a joint loss function for end-to-end training of the model. The total loss function is defined as a weighted sum of the diffusion loss and the image reconstruction loss:

[0051] Among them, diffusion loss Noise predicted by the noise prediction network ^ and real noise The mean square error between them ensures the consistency of the distribution of the generated results.

[0052] Image reconstruction loss The calculation method is based on prediction noise. ^Estimated value of potential precipitation representation x ^ T Inputting it into the frozen VAE decoder yields a reconstructed precipitation map. y ^, calculate its relationship with the actual precipitation grid map y The mean square error within the effective precipitation area is used to directly constrain pixel-level numerical accuracy.

[0053] Weighting coefficient λ The value was determined to be 0.1 through grid search. The optimizer used was AdamW, with an initial learning rate set to 1×10⁻⁶. -4 The cosine annealing process reduces the temperature to 1×10⁻¹⁰ within 300 rounds. -6 The weight decay coefficient is 1×10 -4 The batch size is set to 8, and gradient accumulation (accumulation steps of 2) is used to effectively increase the batch size.

[0054] Through the above steps, this invention transforms the complex cross-modal precipitation inversion problem into a structured conditional evolution process constrained by both radar observation and atmospheric background, thus achieving intelligent generation of high-precision, physically consistent precipitation fields.

[0055] S6. Model Reasoning and Application After completing the above training, the model can be used for actual precipitation inversion applications. For any given target time, the following inference process is executed: (1) Press S1 to obtain the 6-frame radar sequence and corresponding ERA5 data at that moment.

[0056] (2) Extract radar fusion features according to S2 and ERA5 conditional features c .

[0057] (3) From the standard normal distribution N(0, I A vector is randomly sampled from the sample and used as an approximate endpoint. (The actual precipitation during the inference phase is unknown, so random noise is used as a substitute).

[0058] (4) From t = T Begin by iteratively executing the inverse denoising process. For each time step... t , the current state xt Time step t and conditional features c Input the trained noise prediction network , to obtain predicted noise ^, and calculate using the Brownian bridge inverse update formula. Repeat this process until... t =0, thus obtaining the final potential precipitation representation. x ^0.

[0059] (5) x The input is a frozen VAE decoder, which is then used to recover a high-resolution precipitation estimation map with a spatial size of 700×900. The output is the final inversion result.

[0060] In summary, the two-stage training framework and Brownian bridge conditional diffusion model proposed in this invention effectively solve the technical problems of insufficient fusion of multi-source heterogeneous data, high uncertainty in cross-modal mapping, and single conditional information injection mechanism, and realize high-precision and high-physical-consistency quantitative precipitation intelligent inversion.

[0061] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0062] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0063] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0064] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium such as a floppy disk, hard disk, magnetic tape; an optical medium such as a high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.

[0065] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0066] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0067] The quantitative precipitation inversion method and system based on Brownian bridge conditional diffusion and multimodal fusion proposed in this invention have been described in detail above. The principles and implementation methods of this invention have been explained. The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A quantitative precipitation inversion method based on Brownian bridge conditional diffusion and multimodal fusion, characterized in that: The method specifically includes the following steps: S1, Construction and Spatiotemporal Alignment of Multimodal Conditional Dataset: Acquire 3D weather radar volume scan data, ground automatic weather station precipitation records and ERA5 atmospheric reanalysis data within the target area and time period, perform spatiotemporal alignment and preprocessing, and construct radar-precipitation-ERA5 multimodal paired dataset. S2, Unified Latent Space Coding for Multi-Source Heterogeneous Data: Maps three types of heterogeneous data—radar, ERA5, and precipitation—to a dimension-aligned unified latent space, providing a standardized input format for the Brownian Bridge diffusion model; S3. Construct and train a Brownian bridge conditional diffusion generation model: Train a Brownian bridge diffusion model in the potential space, starting with radar features, ending with the potential representation of precipitation, and conditional with ERA5 features; S4. Design of Multi-Scale Conditional Injection and Noise Prediction Network: Design a U-Net noise prediction network that includes a multi-scale conditional injection mechanism, injecting spatial splicing mechanism, cross-attention mechanism and time-conditional joint modulation; S5. Design a joint loss function to perform end-to-end training of the model: S6. Model Inference and Application: After training is completed, the model is applied to actual precipitation inversion. For any given target time, the precipitation estimation map is output as the final inversion result.

2. The method according to claim 1, characterized in that: In S1, Radar data processing: Using the multi-frame sequence input method, a continuous 6-frame three-dimensional radar data consisting of the target time and its 5 preceding times is constructed; ERA5 data processing: Six variables closely related to precipitation physical processes were selected, including relative humidity, specific cloud liquid water content, specific humidity, temperature, zonal wind, and meridional wind; each variable was taken at 5 isobaric levels, for a total of 30 channels; all variables were resampled to the same spatial grid as the radar data through bilinear interpolation and aligned with the radar samples through a nearest neighbor time matching strategy. Precipitation labeling: Hourly precipitation records from automatic weather stations are collected. After screening for climate extremes, phase consistency tests, and spatial consistency checks, two types of labels are generated: a two-dimensional precipitation grid map obtained through Kriging interpolation, and the original station observation dictionary set.

3. The method according to claim 2, characterized in that: S2 includes, S2.1, Radar Feature Extraction and Temporal Fusion: A pre-trained 3D mask autoencoder is used as the radar encoder to extract the spatial structural features of a single frame of radar data. For a continuous 6-frame radar sequence, the temporal dependency model of each spatial location is independently performed by the point-by-point time Transformer module, and the radar fusion feature that incorporates dynamic evolution information is output. S2.2, ERA5 Conditional Encoding: Design a lightweight convolutional encoder to map 30-channel ERA5 data to the same latent spatial dimension as radar features to obtain conditional features; S2.3, Precipitation label encoding: A variational autoencoder (VAE) is used to compress high-resolution precipitation images into continuous and regularized latent representations. The parameters of the VAE are frozen after separate pre-training.

4. The method according to claim 3, characterized in that: S3 includes, S3.1, Forward Diffusion Construction: Define a random bridging path starting from radar fusion features and ending with VAE-encoded precipitation potential representation, set the total number of diffusion steps and adopt a cosine noise scheduling strategy; during the training phase, define interpolation coefficients and noise intensity for random sampling time steps, and obtain intermediate states based on the Brownian bridge forward process. S3.2, Inverse Denoising Process: Construct a denoising network with ERA5 conditions as input, learn the inverse process of recovering from the noisy intermediate state to the target state; by predicting and removing the noise added during the forward diffusion process, gradually generate the potential representation of precipitation.

5. The method according to claim 4, characterized in that: In S4, The noise prediction network adopts a U-Net architecture with symmetric encoding and decoding. It extracts high-dimensional semantic features step by step through the downsampling path, and then restores spatial details through the upsampling path. Skip connections are used to compensate for the loss of details. Injection of spatial stitching mechanism: In each downsampling and upsampling stage of the network, the ERA5 conditional features are adjusted to the current feature map size by bilinear interpolation through spatial stitching, and then convolved and fused with the main features in the channel dimension to guide the generation of local precipitation structures; A cross-attention mechanism is introduced at the bottleneck layer, using the main feature as the query and the condition feature as the key to achieve adaptive aggregation of global physical information; Time-conditional joint modulation: The diffusion time step is combined with the global statistics of the conditional features. The feature linear modulation FiLM parameters are generated by the multilayer perceptron. The main features are scaled and shifted channel by channel, so that the network can dynamically adjust its dependence on ERA5 information according to the denoising stage.

6. The method according to claim 5, characterized in that: In S5, The total loss function is defined as a weighted sum of the diffusion loss and the image reconstruction loss: Among them, diffusion loss To minimize the mean square error between the noise predicted by the noise prediction network and the actual forward noise; Image reconstruction loss The method involves back-calculating the potential precipitation estimate based on the predicted noise, reconstructing the precipitation image using a frozen VAE decoder, and calculating the mean square error between the reconstructed image and the actual precipitation image.

7. A quantitative precipitation inversion system based on Brownian bridge conditional diffusion and multimodal fusion, characterized in that: The system is used to execute the quantitative precipitation inversion method based on Brownian bridge conditional diffusion and multimodal fusion as described in any one of claims 1 to 6; The system includes a module for constructing and spatiotemporally aligning a multimodal conditional dataset, a module for unified latent spatial coding of multi-source heterogeneous data, a Brownian bridge conditional diffusion generation model, a module for designing a multi-scale conditional injection and noise prediction network, a joint loss module, and an application module. The module for constructing and aligning the multimodal conditional dataset is used to acquire three-dimensional weather radar volume scan data, precipitation records from ground automatic weather stations, and ERA5 atmospheric reanalysis data within the target area and time period, perform spatiotemporal alignment and preprocessing, and construct a radar-precipitation-ERA5 multimodal paired dataset. The unified latent space coding module for multi-source heterogeneous data is used to map three types of heterogeneous data—radar, ERA5, and precipitation—to a dimension-aligned unified latent space, providing a standardized input format for the Brownian bridge diffusion model. The Brownian bridge conditional diffusion generation model is used to train a Brownian bridge diffusion model in the potential space, starting with radar features, ending with the potential representation of precipitation, and conditional with ERA5 features. The multi-scale conditional injection and noise prediction network design module is used to design a U-Net noise prediction network that includes a multi-scale conditional injection mechanism, injecting spatial splicing mechanism, cross-attention mechanism and time-conditional joint modulation. The joint loss module performs end-to-end training of the model: The application module is used to perform actual precipitation inversion after training is completed; for any given target time, it outputs a precipitation estimation map as the final inversion result.

8. A computer device system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program instructions implement the steps of the method according to any one of claims 1 to 6.

Images