A short-term precipitation prediction method based on frequency domain representation and space-time decoupling

Abstract

The application discloses a short-term precipitation prediction method based on frequency domain representation and space-time decoupling, and belongs to the technical field of image data processing, and comprises the following steps: acquiring a data set D for short-term precipitation prediction, wherein samples are sequences composed of continuous N frames of radar echo maps; constructing a short-term precipitation prediction network, including a frequency domain coding module, a conditional diffusion training module and a frequency domain decoding module; training the short-term precipitation prediction network with the data set D until convergence, so as to obtain a short-term precipitation prediction model for short-term precipitation prediction. The application proposes a method for converting an image to a frequency domain space, using a conditional diffusion training module and a space-time decoupling state space unit to learn and predict a generation distribution of a future sequence in the frequency domain space, and then inversely transforming and restoring to a pixel domain to perform short-term precipitation prediction, and the method has the advantages of clear prediction result details, less noise and high prediction precision.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling. Background Technology

[0002] The core task of radar echo nowcasting is to predict the echo distribution for the next few frames based on past radar echo or precipitation intensity images. This task requires the model to determine the overall movement and intensity changes of the rain area, while also preserving strong echo boundaries, local textures, and small-scale precipitation cores as much as possible. Currently, common radar echo prediction methods can be broadly classified into three categories: extrapolation methods, numerical weather prediction methods, and deep learning methods.

[0003] Extrapolation methods typically estimate motion vectors from continuous echo images and then translate or deform the current echo field into the future. Optical flow, correlation tracking, and Lagrange extrapolation all belong to this category. These methods are computationally fast and have stable short-term performance, but they assume that the echo structure does not change much within the prediction window. When strong convective systems undergo new formation, merging, splitting, or rapid decay, extrapolation alone is insufficient to describe the true evolution.

[0004] Numerical weather prediction methods, based on atmospheric dynamic equations and physical parameterization schemes, can explain precipitation formation from physical processes. However, in nowcast scenarios, they suffer from slow start-up, high computational load, and sensitivity to the initial field. For localized severe convection, small-to-medium-scale rain clusters, and short-duration sudden precipitation, numerical models often struggle to meet the requirements for high-time-precipitation forecasts.

[0005] Deep learning methods learn future patterns of change by analyzing historical echo sequences, and can be further subdivided into the following categories: (1) Recurrent structures such as ConvLSTM (Convolutional Long Short-Term Memory), PredRNN (Predictive Recurrent Neural Network), and TrajGRU (Trajectory Gated Recurrent Unit) can handle spatial features and temporal dependencies at the same time, but error accumulation is likely to occur when predicting long sequences, and the predicted image is also likely to become smooth.

[0006] (2) The Transformer-based method has strong global modeling capabilities, but the attention computation cost is high and the preservation of local boundaries and small-scale strong echo details is not stable.

[0007] (3) The diffusion model generates prediction results by gradually denoising, and the generation quality and stability are good. However, if the model is directly modeled in the pixel domain, the noise prediction network needs to learn large-scale rainband structure and high-frequency details at the same time, which is more difficult to train.

[0008] Radar echo images inherently possess a distinct frequency domain structure. Low-frequency components typically correspond to large-scale precipitation patterns and rainband morphology, while high-frequency components contain edges, abrupt changes, textures, and small-scale strong echo information. Existing methods mostly model directly in the pixel domain or ordinary feature space, failing to fully utilize this frequency domain separability. Furthermore, echo sequence prediction involves two distinct problems: spatial structure modeling within a single frame and temporal evolution modeling across frames. Directly mixing spatial and temporal aspects, while enabling end-to-end model training, results in unclear internal representations and increased computational and optimization complexity.

[0009] Definitions: Radar echo map: The pixel value essentially corresponds to the reflectivity factor (Z), which is measured in dBZ (decibels). The larger the value, the stronger the reflected electromagnetic wave energy, meaning there are more and larger precipitation particles (raindrops, snowflakes, hail, etc.) in the atmosphere. Therefore, the pixel value in the radar echo map can express the precipitation at the corresponding location.

[0010] Precipitation nowcasting: Based on current and past short-term observational data, the task of predicting changes in the location, extent, and intensity of precipitation in the near future.

[0011] Frequency domain normalization: The frequency domain coefficients are standardized based on the mean and standard deviation of the frequency domain channels obtained from the training set statistics. This reduces the amplitude differences between low-frequency and high-frequency coefficients, preventing model training from being dominated by certain frequency bands.

[0012] Diffusion Models (DM) are generative models that progressively add and remove noise. During training, the model learns how to predict noise from noisy samples; during inference, it starts from random noise and progressively removes noise to generate the target sample.

[0013] Conditional Diffusion Model (CDM): This model incorporates additional conditional information into the diffusion model. In this invention, historical frequency domain echo sequences are used as conditions, and future frequency domain echo sequences are used as the generation target.

[0014] Haar wavelet transform: Performing Haar wavelet transform on the original image yields four sub-band images, namely low-frequency approximation (LL), horizontal high frequency (LH), vertical high frequency (HL), and diagonal high frequency (HH). Summary of the Invention

[0015] The purpose of this invention is to provide a short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling, which solves the problems in the prior art, such as error accumulation causing smoothing of predicted images during long-sequence prediction, high attention cost of Transformer method, instability in maintaining local boundaries and small-scale strong echo details, complex spatiotemporal hybrid modeling, and unclear internal scaling.

[0016] To achieve the above objectives, the technical solution adopted by this invention is as follows: a short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling, comprising the following steps: S1. Obtain the dataset D for short-term precipitation prediction, where the sample is a sequence of N consecutive radar echo images. The first N1 frames of the sample are labeled as historical sequence Y, and the last N2 frames are labeled as future sequence X, N = N1 + N2. S2, construct a short-term precipitation prediction network, including a frequency domain coding module, a conditional diffusion training module, and a frequency domain decoding module; The frequency domain coding module is used to perform Haar wavelet transform on Y and X frame by frame, decomposing each frame of single-channel radar echo image into four sub-bands in the frequency domain, to obtain the historical frequency domain sequence Y corresponding to Y and X. frep and future frequency domain sequence X frep Then Y frep X frep Normalize them separately to obtain the normalized historical frequency domain sequence Y. norm Normalized future frequency domain sequence X norm ; The conditional diffusion training module includes a noise-adding unit and a conditional denoising unit. The noise-adding unit is used to direct X... norm Add noise step by step to generate a noisy sequence X at time step t. t The conditional denoising unit is used to input X t Y norm and t, and with Y norm As a condition, from X t By predicting and removing noise step by step, a normalized future frequency domain sequence is obtained. ; The conditional denoising unit includes a splicing layer, an embedding layer, a spatiotemporal decoupling state space unit, a feature rearrangement layer, and a denoising unit. The splicing layer is used to integrate Y... norm and X t spliced ​​into an image sequence of length N , For L t The i-th frame image; The embedding layer is used to... The image is divided into J image blocks. Each image block is encoded and then encoded with spatial location encoding and temporal step t encoding to obtain the corresponding embedding vector, where the image block s at position j in the i-th frame is an integer. ij The embedding vector is e ij e ij The dimension is d; The spatiotemporal decoupling state-space unit is used for e ij Spatial state modeling is performed to obtain spatial augmentation features. and all The sequences are spliced ​​into an N×J×d spatial enhancement sequence FS1, and then N frames at the same position in FS1 are processed. Perform time-state modeling to generate J×N×d spatiotemporal enhanced features FS2; The feature rearrangement layer is used to first rearrange FS2 to FS1 scale, and then reverse the operation back to L according to the embedding layer. t The scale is used to construct the prediction noise for time step t, taking the last N2 frames. The noise reduction unit is used to remove X t By subtracting the prediction noise step by step, we obtain the predicted normalized future frequency domain sequence. ; The frequency domain decoding module is used for... By performing inverse normalization and Haar wavelet transform operations, the predicted future sequence is obtained. ; S3, train the imminent precipitation prediction network with dataset D until convergence to obtain the imminent precipitation prediction model; S4 generates corresponding predicted future sequences based on historical sequences using the imminent precipitation prediction model.

[0017] Preferably, in S1, the method for constructing the dataset D includes S11 to S13; S11, Load the radar echo sequence, which is composed of radar echo maps arranged in time sequence; S12, randomly select N consecutive radar echo images to form a sample, and label the first N1 frames of the sample as the historical sequence Y, and the last N2 frames as the target sequence X, N=N1+N2; S13, repeat S12 to generate multiple samples, all of which constitute dataset D.

[0018] Preferably, the frequency domain coding module includes a frequency domain transformation unit and a frequency domain normalization unit; The frequency domain transformation unit is used to perform Haar wavelet transform on Y and X frame by frame, transforming the 1×H×W radar echo map of each frame into... Wavelet feature maps are used to obtain the historical frequency domain sequence Y of Y. frep and the future frequency domain sequence X frepWhere H and W are the height and width of the radar echo map, respectively, and 4 is the number of feature channels in the wavelet feature map, corresponding to the four sub-band features LL, LH, HL, and HH after frequency domain decomposition. , ; The frequency domain normalization unit is used to statistically analyze the mean and standard deviation of each channel of all wavelet feature maps for Y. frep X frep Channel-by-channel normalization yields the corresponding normalized historical frequency domain sequence Y. norm and normalized future frequency domain sequence X norm .

[0019] Preferably, the spatiotemporal decoupled state space unit includes a spatial modeling unit and a temporal modeling unit; The spatial modeling unit is used for... J e ij Perform forward and reverse scans respectively to obtain forward and reverse sequences. Concatenate the j-th element of the two sequences and map it to dimension d to obtain s. ij Spatial enhancement features Then J Rearranged by position into J×d spatial enhancement features; finally, N frames... The spatial enhancement features are spliced ​​together to form an N×J×d spatial enhancement sequence FS1; The time modeling unit is used to input FS1 and obtain the time series at position j. Time j ∈R N×d The feature is reconstructed into J×N×d based on its position, and then processed by the time-dimensional selective state-space modeling module to become the spatiotemporal enhanced feature FS2.

[0020] As a preferred option, when training the precipitation prediction network in S3, the loss function is the MSE error between the actual noise and the predicted noise.

[0021] Preferably, S4 specifically involves: during prediction, obtaining the historical sequence Y' of the future sequence to be predicted, and generating a normalized historical frequency domain sequence of Y' through the frequency domain coding module. ,by As a condition, a conditional denoising unit is used to denoise Gaussian noise and generate a normalized future frequency domain sequence. Then, the corresponding predicted future sequence is obtained through the frequency domain decoding module.

[0022] Compared with the prior art, the advantages of the present invention are as follows: (1) This invention does not directly predict future images in the pixel domain, but first decomposes the echo image into the frequency domain space for modeling. Since the low-frequency part usually corresponds to the main body of large-scale precipitation and the shape of rainbands, while the high-frequency part contains edge, abrupt change, texture and small-scale strong echo information, this invention expresses the main structure of the rain area through the low-frequency sub-band and the boundary and local details through the high-frequency sub-band, making the model learning target clearer.

[0023] (2) This invention introduces frequency domain statistical normalization of the dataset / training set. It iterates through the samples in the dataset / training set, calculates the mean and standard deviation of each frequency domain channel, and saves them as a statistical file. During training and inference, the frequency domain sequence is normalized according to this statistical file; after the model outputs, inverse normalization is performed according to the same statistical file. This method avoids training bias caused by excessively large amplitudes of low-frequency coefficients and small amplitudes of high-frequency coefficients, making frequency domain diffusion training more stable.

[0024] (3) The present invention uses conditional diffusion to generate predicted normalized future frequency domain sequences. Normalized historical frequency domain sequence Y norm It is not simply used as an additional input, but as a conditional constraint for the entire denoising process, so that the generated results maintain a continuous relationship with historical observations.

[0025] (4) Spatiotemporal decoupling state space modeling is introduced in the denoising process. Spatial modeling is responsible for the structure and boundary relationships of rain areas in a single frame, while temporal modeling is responsible for cross-frame movement and intensity evolution. The two have a clear division of labor, avoiding the expression confusion caused by direct spatiotemporal mixing. Moreover, the use of forward and reverse sequence modeling in the spatial modeling stage can obtain bidirectional spatial context at the same time, which is more friendly to irregular rain areas, strip-shaped echoes and the edges of local strong echoes.

[0026] In summary, this invention converts images to the frequency domain and uses a conditional diffusion training module and a spatiotemporally decoupled state space unit to predict near-term precipitation, which has the advantages of clear prediction results, low noise, and high prediction accuracy. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram of the short-term precipitation forecasting network structure. Figure 3 This is a flowchart of the data processing for the short-term precipitation forecasting network. Figure 4 Data processing flowchart for spatiotemporally decoupled state space units; Figure 5 This is a comparison chart of the predicted results. Detailed Implementation

[0028] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0029] Example 1: See Figures 1 to 4 A short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling includes the following steps: S1. Obtain the dataset D for short-term precipitation prediction, where the sample is a sequence of N consecutive radar echo images. The first N1 frames of the sample are labeled as historical sequence Y, and the last N2 frames are labeled as future sequence X, N = N1 + N2. S2, construct a short-term precipitation prediction network, including a frequency domain coding module, a conditional diffusion training module, and a frequency domain decoding module; The frequency domain coding module is used to perform Haar wavelet transform on Y and X frame by frame, decomposing each frame of single-channel radar echo image into four sub-bands in the frequency domain, to obtain the historical frequency domain sequence Y corresponding to Y and X. frep and future frequency domain sequence X frep Then Y frep X frep Normalize them separately to obtain the normalized historical frequency domain sequence Y. norm Normalized future frequency domain sequence X norm ; The conditional diffusion training module includes a noise-adding unit and a conditional denoising unit. The noise-adding unit is used to direct X... norm Add noise step by step to generate a noisy sequence X at time step t. t The conditional denoising unit is used to input X t Y norm and t, and with Y norm As a condition, from X t By predicting and removing noise step by step, a normalized future frequency domain sequence is obtained. ; The conditional denoising unit includes a splicing layer, an embedding layer, a spatiotemporal decoupling state space unit, a feature rearrangement layer, and a denoising unit. The splicing layer is used to integrate Y... norm and X t spliced ​​into an image sequence of length N , For L t The i-th frame image; The embedding layer is used to... The image is divided into J image blocks. Each image block is encoded and then encoded with spatial location encoding and temporal step t encoding to obtain the corresponding embedding vector, where the image block s at position j in the i-th frame is an integer. ij The embedding vector is e ij e ij The dimension is d; The spatiotemporal decoupling state-space unit is used for e ij Spatial state modeling is performed to obtain spatial augmentation features. and all The sequences are spliced ​​into an N×J×d spatial enhancement sequence FS1, and then N frames at the same position in FS1 are processed. Perform time-state modeling to generate J×N×d spatiotemporal enhanced features FS2; The feature rearrangement layer is used to first rearrange FS2 to FS1 scale, and then reverse the operation back to L according to the embedding layer. t The scale is used to construct the prediction noise for time step t, taking the last N2 frames. The noise reduction unit is used to remove X t By subtracting the prediction noise step by step, we obtain the predicted normalized future frequency domain sequence. ; The frequency domain decoding module is used for... By performing inverse normalization and Haar wavelet transform operations, the predicted future sequence is obtained. ; S3, train the imminent precipitation prediction network with dataset D until convergence to obtain the imminent precipitation prediction model; S4 generates corresponding predicted future sequences based on historical sequences using the imminent precipitation prediction model.

[0030] In this embodiment S1, the method for constructing dataset D includes S11~S13; S11, Load the radar echo sequence, which is composed of radar echo maps arranged in time sequence; S12, randomly select N consecutive radar echo images to form a sample, and label the first N1 frames of the sample as the historical sequence Y, and the last N2 frames as the target sequence X, N=N1+N2; S13, repeat S12 to generate multiple samples, all of which constitute dataset D.

[0031] The frequency domain coding module includes a frequency domain transformation unit and a frequency domain normalization unit; The frequency domain transformation unit is used to perform Haar wavelet transform on Y and X frame by frame, transforming the 1×H×W radar echo map of each frame into... Wavelet feature maps are used to obtain the historical frequency domain sequence Y of Y. frep and the future frequency domain sequence X frep Where H and W are the height and width of the radar echo map, respectively, and 4 is the number of feature channels in the wavelet feature map, corresponding to the four sub-band features LL, LH, HL, and HH after frequency domain decomposition. , The frequency domain normalization unit is used to statistically analyze the mean and standard deviation of each channel of all wavelet feature maps, for Y. frep X frep Channel-by-channel normalization yields the corresponding normalized historical frequency domain sequence Y. normand normalized future frequency domain sequence X norm .

[0032] The spatiotemporal decoupling state space unit includes a spatial modeling unit and a temporal modeling unit; the spatial modeling unit is used for... J e ij Perform forward and reverse scans respectively to obtain forward and reverse sequences. Concatenate the j-th element of the two sequences and map it to dimension d to obtain s. ij Spatial enhancement features Then J Rearranged by position into J×d spatial enhancement features; finally, N frames... The spatial augmentation features are concatenated into an N×J×d spatial augmentation sequence FS1; the temporal modeling unit is used to input FS1 to obtain the time sequence at position j. Time j ∈R N×d The feature is reconstructed into J×N×d based on its position, and then processed by the time-dimensional selective state-space modeling module to become the spatiotemporal enhanced feature FS2.

[0033] When training the precipitation prediction network in S3, the loss function is the MSE error between the actual noise and the predicted noise.

[0034] S4 specifically involves: during prediction, obtaining the historical sequence Y' of the future sequence to be predicted, and generating a normalized historical frequency domain sequence of Y' through the frequency domain coding module. ,by As a condition, a conditional denoising unit is used to denoise Gaussian noise and generate a normalized future frequency domain sequence. Then, the corresponding predicted future sequence is obtained through the frequency domain decoding module.

[0035] Example 2: See Figures 1-4 A short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling includes the following steps: S1, same as step S1 in embodiment 1, where N1=N2=6, N=12, then a sample consists of 12 consecutive radar echo images. Assume the scale of each radar echo image is C×H×W=1×128×128, then the shape of the historical sequence Y formed by the first 6 frames is 6×1×128×128, and the shape of the future sequence X formed by the last 6 frames is 6×1×128×128. S2. Construct a short-term precipitation prediction network, including a frequency domain coding module, a conditional diffusion training module, and a frequency domain decoding module. This embodiment takes processing a batch of B samples as an example to describe the processing methods of each module in detail: Obtain a batch of samples, in which the shape of the historical sequence and the future sequence is B×6×1×128×128.

[0036] The frequency domain coding module performs Haar wavelet transform on each radar echo image frame, changing each frame from 1×128×128 to 4×64×64. Then Y... frep X frep The shape of each element is B×6×4×64×64, and the Y element is obtained by normalizing each channel. norm and X norm The shape is also B×6×4×64×64, and normalization can be achieved using the following formula: , Where Std and mean represent the standard deviation and mean of the wavelet feature maps corresponding to all samples in the dataset in each channel, respectively; Std' represents the standard deviation after lower limit clipping; and max(∙) is the max function. The conditional diffusion training module includes a noise-adding unit and a conditional denoising unit. The noise-adding unit is used to input X norm Add noise step by step to generate a noisy sequence X at time step t. t ,according to Add noise, where t is the time step in the diffusion process. This represents the cumulative signal preservation coefficient determined by the diffuse noise scheduling function; Corresponding cumulative noise variance; Epsilon represents the variance relative to X. norm Standard Gaussian noise with the same shape; The conditional denoising unit is used to input X. t Y norm and t, output Its structure includes a splicing layer, an embedding layer, a spatiotemporally decoupled state space unit, a feature rearrangement layer, and a denoising unit; The splicing layer is used to integrate Y... norm and X t spliced ​​into an image sequence of length N The shape becomes B×12×4×64×64; For each 4×64×64 image The embedding layer divides the image into blocks. If a block embedding of size 2×2 with a stride of 2 is used, 32×32=1024 image blocks are obtained. After encoding, there are a total of J=1024 embedding vectors. Each embedding vector has a dimension d=512, so the shape of each frame image becomes 1024×512. norm The corresponding embedding vector has an overall shape of B×6×1024×512, X t The corresponding embedding vector has an overall shape of B×6×1024×512, and the image patch s at position j in the i-th frame... ij The embedding vector is e ij ; The spatiotemporal decoupling state-space unit is used for L t e of each image block ij To perform spatial state modeling, the specific implementation method is as follows: (1) Obtain For its J e ij A forward scan yields the forward sequence e. i1 ~e i1024 Reverse scanning yields the reverse sequence e i1024 ~e i1 Then, for one position, first put e i1 e i1024 After concatenation and mapping to dimension d=512, image patch s at position 1 is obtained. i1 Spatial enhancement features Process each image block sequentially to obtain ~ The overall shape of the spatial enhancement sequence FS1 obtained by splicing the spatial enhancement features of 12 frames is N×J×d=12×1024×512; (2) For N frames at the same position in FS1 Perform time-state modeling, time series of position j Each Time j The shape is 12×512, which is reconstructed into J×N×d=1024×12×512 according to j=1~1024, and then processed by the time dimension selective state space modeling module into spatiotemporal enhanced feature FS2, with a shape of 1024×12×512. The feature rearrangement layer is used to first rearrange FS2 from 1024×12×512 to 12×1024×512, and then reverse the embedding layer operation back to L. t The shape is B×12×4×64×64, and the last 6 frames are used to form the prediction noise at time step t. Its shape is B×6×4×64×64; The noise reduction unit is used to remove X t By subtracting the prediction noise step by step, we obtain the predicted normalized future frequency domain sequence. Its shape is B×6×4×64×64; The frequency domain decoding module is used for... By performing inverse normalization and Haar wavelet transform operations, the predicted future sequence is obtained. The shape is B×6×1×128×128; S3, train the imminent precipitation prediction network with dataset D until convergence to obtain the imminent precipitation prediction model; S4, During prediction, the historical sequence Y' of the future sequence to be predicted is obtained, and the normalized historical frequency domain sequence of Y' is generated by the frequency domain coding module. ,by As a condition, a conditional denoising unit is used to denoise Gaussian noise and generate a normalized future frequency domain sequence. Then, the corresponding predicted future sequence is obtained through the frequency domain decoding module.

[0037] This invention is not limited to 6-frame input and 6-frame output, nor is it limited to 128×128 resolution. The number of historical frames, the number of predicted frames, the image size, the image patch size, the number of wavelet decomposition layers, and the network depth of the spatiotemporally decoupled state space unit can all be adjusted according to actual needs.

[0038] Example 3: See Figure 5 To evaluate the performance of this invention in radar echo prediction tasks, comparative experiments were conducted to verify its effectiveness. Currently recognized classic models in the field were selected as benchmarks: Smart-Unet, PhyDNet, ConvlSTM, PredRNN, and Rainformer. The experiments evaluated prediction performance under multiple metrics, with two key metrics: The first category is used to evaluate the quality of image reconstruction, including the structural similarity index SSIM, peak signal-to-noise ratio PSNR, and perceptual similarity index LPIPS; The second category is used to evaluate the accuracy of precipitation forecasts, including the Critical Success Index (CSI), the Hyderick Skill Score (HSS), and the Fair Threat Score (ETS), with thresholds of 10 dBz, 20 dBz, 30 dBz, and 40 dBz, respectively. The results are shown in Tables 1 to 4 below: Table 1. Image Reconstruction Quality Comparison Table Smart-Unet 0.5941 22.1460 0.4375 PhyDNet 0.6301 23.1269 0.4221 Rainformer 0.6043 22.7086 0.4321 ConvlSTM 0.6311 22.9411 0.4163 PredRNN 0.6325 23.0226 0.3994 This invention 0.6452 22.8713 0.3255 Table 2. Critical Success Index (CSI) Comparison Table Smart-Unet 0.4629 0.1584 0.0178 0.0004 Rainformer 0.4750 0.1634 0.0185 0.0008 PhyDNet 0.4868 0.1822 0.0232 0.0013 ConvlSTM 0.6793 0.3722 0.1307 0.0202 PredRNN 0.6900 0.3971 0.1434 0.0265 Our Model 0.6961 0.5213 0.1761 0.0790 Table 3. Comparison of Heidrick Skills Scores (HSS) Smart-Unet 0.2613 0.1380 0.0189 0.0005 Rainformer 0.2812 0.1466 0.0200 0.0009 PhyDNet 0，3096 0.1695 0.0244 0.0016 ConvlSTM 0.5954 0.4248 0.1795 0.0301 PredRNN 0.6069 0.4519 0.1955 0.0390 This invention 0.7103 0.6092 0.2560 0.1205 Table 4. Fairness Threat Score ETS Comparison Table Smart-Unet 0.2574 0.1054 0.0144 0.0003 Rainformer 0.2764 0.1113 0.0153 0.0007 PhyDNet 0.2995 0.1308 0.0191 0.0012 ConvlSTM 0.4590 0.2979 0.1224 0.0196 PredRNN 0.4703 0.3199 0.1351 0.0258 This invention 0.5726 0.4649 0.2732 0.0753 The results show that this invention achieves the best performance across all key indicators. Regarding image reconstruction quality, compared to other models, this invention significantly improves the structural and perceptual similarity of the generated images, indicating that its predictions are closer to the actual echoes, with clearer details and less noise. In terms of precipitation prediction accuracy, which has greater practical application value, this invention also demonstrates a significant advantage, with its CSI value being significantly higher than all comparative models. This indicates that it has higher accuracy in identifying and locating meteorological targets of varying intensities, such as areas of heavy rainfall. Overall, the experimental data fully validates the superior performance of this invention in radar echo prediction tasks, demonstrating not only comprehensive leadership in image reconstruction quality but also outstanding performance in the crucial ability to detect meteorological targets.

[0039] See Figure 5 , Figure 5 This is a set of prediction performance comparison images obtained during actual training, containing 3 rows and 6 columns of images. The first row is the normalized historical frequency domain sequence Y. norm The second line is the normalized future frequency domain sequence X. norm The third line predicts future sequences. It can be seen that the model's predicted results are basically consistent with the actual future results in terms of the main location, spatial distribution range, and changing trend of strong echo areas of precipitation echoes. As the prediction time progresses, the prediction results still retain the main outline of the rainband and the local strong echo structure well, without significant overall positional shifts or excessive smoothing. This result demonstrates that the present invention, by jointly modeling the low-frequency rainband structure and high-frequency edge details through frequency domain representation, and by combining spatiotemporally decoupled state spatial units to capture intra-frame spatial relationships and cross-frame temporal evolution relationships, can improve the accuracy of echo morphology preservation and strong precipitation area localization in short-term precipitation prediction.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling, characterized in that, Includes the following steps: S1. Obtain the dataset D for short-term precipitation prediction, where the sample is a sequence of N consecutive radar echo images. The first N1 frames of the sample are labeled as historical sequence Y, and the last N2 frames are labeled as future sequence X, N = N1 + N2. S2, construct a short-term precipitation prediction network, including a frequency domain coding module, a conditional diffusion training module, and a frequency domain decoding module; The frequency domain coding module is used to perform Haar wavelet transform on Y and X frame by frame, decomposing each frame of single-channel radar echo image into four sub-bands in the frequency domain, to obtain the historical frequency domain sequence Y corresponding to Y and X. frep and future frequency domain sequence X frep Then Y frep X frep Normalize them separately to obtain the normalized historical frequency domain sequence Y. norm Normalized future frequency domain sequence X norm ; The conditional diffusion training module includes a noise-adding unit and a conditional denoising unit. The noise-adding unit is used to direct X... norm Add noise step by step to generate a noisy sequence X at time step t. t The conditional denoising unit is used to input X t Y norm and t, and with Y norm As a condition, from X t By predicting and removing noise step by step, a normalized future frequency domain sequence is obtained. ; The conditional denoising unit includes a splicing layer, an embedding layer, a spatiotemporal decoupling state space unit, a feature rearrangement layer, and a denoising unit. The splicing layer is used to integrate Y... norm and X t spliced ​​into an image sequence of length N , For L t The i-th frame image; The embedding layer is used to... The image is divided into J image blocks. Each image block is encoded and then encoded with spatial location encoding and temporal step t encoding to obtain the corresponding embedding vector, where the image block s at position j in the i-th frame is an integer. ij The embedding vector is e ij e ij The dimension is d; The spatiotemporal decoupling state-space unit is used for e ij Spatial state modeling is performed to obtain spatial augmentation features. and all The sequences are spliced ​​into an N×J×d spatial enhancement sequence FS1, and then N frames at the same position in FS1 are processed. Perform time-state modeling to generate J×N×d spatiotemporal enhanced features FS2; The feature rearrangement layer is used to first rearrange FS2 to FS1 scale, and then reverse the operation back to L according to the embedding layer. t The scale is used to construct the prediction noise for time step t, taking the last N2 frames. The noise reduction unit is used to remove X t By subtracting the prediction noise step by step, we obtain the predicted normalized future frequency domain sequence. ; The frequency domain decoding module is used for... By performing inverse normalization and Haar wavelet transform operations, the predicted future sequence is obtained. ; S3, train the short-term precipitation prediction network with dataset D until convergence to obtain the short-term precipitation prediction model; S4 generates corresponding predicted future sequences based on historical sequences using a short-term precipitation prediction model.

2. The short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling according to claim 1, characterized in that, In S1, the methods for constructing dataset D include S11~S13; S11, Load the radar echo sequence, which is composed of radar echo maps arranged in time sequence; S12, randomly select N consecutive radar echo images to form a sample, and label the first N1 frames of the sample as the historical sequence Y, and the last N2 frames as the target sequence X, N=N1+N2; S13, repeat S12 to generate multiple samples, all of which constitute dataset D.

3. The short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling according to claim 1, characterized in that, The frequency domain coding module includes a frequency domain transformation unit and a frequency domain normalization unit; The frequency domain transformation unit is used to perform Haar wavelet transform on Y and X frame by frame, transforming the 1×H×W radar echo map of each frame into... Wavelet feature maps are used to obtain the historical frequency domain sequence Y of Y. frep and the future frequency domain sequence X frep Where H and W are the height and width of the radar echo map, respectively, and 4 is the number of feature channels in the wavelet feature map, corresponding to the four sub-band features LL, LH, HL, and HH after frequency domain decomposition. , ; The frequency domain normalization unit is used to statistically analyze the mean and standard deviation of each channel of all wavelet feature maps for Y. frep X frep Channel-by-channel normalization yields the corresponding normalized historical frequency domain sequence Y. norm and normalized future frequency domain sequence X norm .

4. The short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling according to claim 1, characterized in that, The spatiotemporal decoupled state space unit includes a spatial modeling unit and a temporal modeling unit; The spatial modeling unit is used for... J e ij Perform forward and reverse scans respectively to obtain forward and reverse sequences. Concatenate the j-th element of the two sequences and map it to dimension d to obtain s. ij Spatial enhancement features ; Then J Spatial enhancement features rearranged by position into J×d; Finally, N frames The spatial enhancement features are spliced ​​together to form an N×J×d spatial enhancement sequence FS1; The time modeling unit is used to input FS1 and obtain the time series at position j. Time j ∈R N ×d The feature is reconstructed into J×N×d based on its position, and then processed by the time-dimensional selective state-space modeling module to become the spatiotemporal enhanced feature FS2.

5. The short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling according to claim 1, characterized in that, When training the precipitation prediction network in S3, the loss function is the MSE error between the actual noise and the predicted noise.

6. The short-term precipitation prediction method based on frequency domain representation and spatiotemporal decoupling according to claim 1, characterized in that, S4 specifically involves: during prediction, acquiring the historical sequence Y', and generating a normalized historical frequency domain sequence of Y' through the frequency domain coding module. ,by As a condition, a conditional denoising unit is used to denoise Gaussian noise and generate a normalized future frequency domain sequence. Then, the corresponding predicted future sequence is obtained through the frequency domain decoding module.

Images