A spatial downscaling method for precipitation fields

Abstract

The present invention discloses a method for spatial downscaling of precipitation fields, which relates to the technical field of meteorological data processing and includes: obtaining high-resolution terrain information data and precipitation paired data composed of high-resolution and low-resolution precipitation data, and performing normalization processing on both the high-resolution terrain information data and the precipitation paired data; inputting the normalized precipitation paired data and high-resolution terrain information data into a pre-trained precipitation field spatial downscaling model to output a high-resolution precipitation field; wherein, the precipitation field spatial downscaling model includes a convolutional neural network and a denoising diffusion probabilistic model, and the denoising diffusion probabilistic model is constructed based on residual calculation, forward noise addition, a multi-scale feature extraction and fusion network, and backward denoising. The present invention can reconstruct a high-resolution precipitation field from a low-resolution precipitation field.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological data processing, and more specifically, to a method for spatially downscaling precipitation fields, which is applicable to scenarios such as high-resolution precipitation forecasting and disaster prevention. Background Art

[0002] Precipitation forecasting is crucial for multiple fields such as disaster prevention and mitigation, urban management, and agricultural production. With the development of remote sensing technology and ground observation technology, the resolution and accuracy of meteorological data have been gradually improved. However, due to the strong nonlinearity, local suddenness, and complex spatial distribution of precipitation data itself, traditional downscaling methods have significant limitations in capturing the fine structure and local changes of heavy precipitation, often lacking accurate restoration of details and local features, and it is difficult to accurately reflect the microscopic distribution characteristics and short-term change trends of heavy precipitation. In addition, although models such as generative adversarial networks (GANs) have achieved certain results in image super-resolution, there are problems such as unstable optimization and noise amplification in the application of precipitation spatial downscaling. Summary of the Invention

[0003] The present invention aims to provide a method for spatially downscaling precipitation fields to reconstruct high-resolution precipitation fields.

[0004] The technical solution adopted by the present invention is as follows:

[0005] A method for spatially downscaling precipitation fields includes: obtaining high-resolution terrain information data and precipitation paired data composed of high-resolution and low-resolution precipitation data, and performing normalization processing on both the high-resolution terrain information data and the precipitation paired data; inputting the normalized precipitation paired data and high-resolution terrain information data into a pre-trained precipitation field spatial downscaling model to output a high-resolution precipitation field; wherein, the precipitation field spatial downscaling model includes a convolutional neural network and a denoising diffusion probabilistic model, and the denoising diffusion probabilistic model is constructed based on residual calculation, forward noise addition, backward denoising, and a multi-scale feature extraction and fusion network based on the UNET architecture.

[0006] Preferably, the low resolution refers to 3000-meter and 1000-meter resolutions, and the high resolution refers to 750-meter resolution.

[0007] Specifically, the convolutional neural network is used to process the low-resolution precipitation data in the precipitation paired data to obtain initial high-resolution precipitation data x cnn ; during the model training process, the frequency domain hybrid loss function loss frequency is used to calculate the training loss of the convolutional neural network.

[0008] loss frequency = αloss FFT + βloss DWT

[0009] Among them, M and are the frequency domain values obtained by applying the fast Fourier transform to the true value and predicted value of the high-resolution precipitation field, respectively; H i , V i , D i are the high-frequency components after the i-th HAAR wavelet transform of the true value; is the high-frequency component after the i-th HAAR wavelet transform of the predicted value.

[0010] Preferably, the optimization objective of the denoising diffusion probability model is defined as follows:

[0011]

[0012] Among them, cond terrain is the high-resolution terrain information data.

[0013] Specifically, the residual calculation is based on two Markov chains to generate the residual x0 = x hr of the true value x of the high-resolution precipitation data and the initial high-resolution precipitation data x cnn - x hr - x cnn .

[0014] Specifically, the forward noise addition formula is:

[0015]

[0016] Among them, β t is a hyperparameter representing the level of noise added to the data, related to the time step t, β t ∈(0,1), t = 1,…,T; α t = 1-β,

[0017] Specifically, the multi-scale feature extraction and fusion network is used to fuse the data x t with added noise and the high-resolution terrain information data cond terrain at the time step t.

[0018] Specifically, the reverse denoising formula is:

[0019]

[0020] Among them, ∈ t_pred is the predicted value of the noise added at the forward time step t corresponding to the UNET output, and ∈ is the randomly sampled Gaussian white noise;

[0021] By randomly sampling a Gaussian white noise and performing denoising through T-step iteration, using the initial high-resolution precipitation data x cnn and the high-resolution terrain information cond terrain as a guide to gradually restore the high-frequency details of the accurate high-resolution precipitation data After that, use the generated high-frequency details of the high-resolution precipitation and the initial high-resolution precipitation data x cnn to synthesize an accurate high-resolution precipitation field with high-frequency details.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] Using a pre-trained convolutional neural network (CNN) to estimate the initial high-resolution precipitation field, making full use of its advantages in spatial feature extraction to support multi-resolution downscaling, including non-integer multiple downscaling; then, using the residual learning ability of the residual network (ResNet), with the encoder-decoder structure of UNET, capturing context information at different scales, guiding with frequency domain information and terrain information, generating the residual between the true value of the high-resolution precipitation field and the preliminary result, and supplementing details to the preliminary result to improve the accuracy of the high-resolution precipitation field data distribution; during this process, through the denoising process, effectively removing noise to ensure that the finally generated high-resolution precipitation field has higher accuracy and reliability.

[0024] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following specifically gives embodiments of the present invention and, in conjunction with the accompanying drawings, makes a detailed description as follows. Brief Description of the Drawings

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0026] Figure 1 is the overall framework diagram of the present invention;

[0027] Figure 2 is the schematic diagram of the architecture of the precipitation field spatial downscaling model of the present invention;

[0028] Figure 3 is the schematic diagram of precipitation field data at different resolutions;

[0029] Figure 4 is the comparison diagram of super-resolution effects. Detailed Embodiments

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention.

[0031] Please refer to Figure 1 and Figure 2 , the present invention discloses a method for spatial downscaling of precipitation fields, which is specifically as follows:

[0032] 1. Obtain high-resolution terrain information data and precipitation pairing data composed of high-resolution and low-resolution precipitation data.

[0033] Among them, 750-meter high-resolution precipitation field data is obtained from radar echo quantitative precipitation estimation products, and precipitation pairing data is constructed by interpolation downsampling method, including the pairing of 3000-meter low-resolution precipitation data and 750-meter high-resolution precipitation data, and the pairing of 1000-meter low-resolution precipitation data and 750-meter high-resolution precipitation data, as shown in Figure 3 shown.

[0034] The high-resolution terrain information data is extracted from 90-meter resolution elevation data through data processing such as projection transformation and data resampling to extract terrain feature information such as slope, aspect, and terrain relief, and generate 750-meter resolution altitude, slope, slope change rate, aspect, aspect change rate, terrain relief, and surface roughness data.

[0035] 2. Perform min-max normalization processing on both the high-resolution terrain information data and the precipitation pairing data to improve the stability and generalization ability of the model.

[0036] 3. Input the normalized precipitation pairing data and high-resolution terrain information data into a pre-trained precipitation field spatial downscaling model, and output a high-resolution precipitation field. Among them, the precipitation field spatial downscaling model includes a convolutional neural network and a denoising diffusion probabilistic model, as shown in Figure 1 , which is divided into two stages. The first stage is to generate an initial high-resolution precipitation field by a convolutional neural network based on a frequency domain loss function, and the second stage is to generate an accurate high-resolution precipitation field that conforms to the true data distribution by a denoising diffusion probabilistic model guided by high-resolution terrain information and high-frequency information.

[0037] To capture the global information and high-frequency information of the precipitation field, in the model training process of the convolutional neural network, the prior art usually uses the difference between the true value and the model output value at the pixel level as the main loss function, while the present invention abandons the loss term of pixel difference and directly uses the frequency domain hybrid loss function loss frequency (including the loss function loss based on the fast Fourier transformFFT and the loss function loss based on the discrete wavelet transform DWT ) Calculate the training loss of the convolutional neural network, enhance the ability of the convolutional neural network to capture global information and initial high-frequency information, and improve the accuracy of subsequent generation of high-resolution precipitation data;

[0038] loss frequency = αloss FFT + βloss DWT

[0039] where

[0040]

[0041] α and β are adjustable hyperparameters, L is the number of discrete wavelet transforms; M and are the frequency domain values obtained by applying the fast Fourier transform to the true value and predicted value of the high-resolution precipitation field, including the real part and the imaginary part; H i , V i , D i are the high-frequency components after the i-th HAAR wavelet transform of the true value; are the high-frequency components after the i-th HAAR wavelet transform of the predicted value.

[0042] The precipitation pairing dataset after normalization processing (such as: 3000m - 750m dataset) includes the low-resolution precipitation field x lr , the high-resolution precipitation field x hr , x lr as the input of the one-stage convolutional neural network, and x hr is used as the target.

[0043] The output of the convolutional neural network is x cnn , construct the dataset corresponding to the spatial scale (x cnn , x hr ), and avoid the occupation of time and resources by the convolutional neural network.

[0044] The denoising diffusion probability model is constructed based on residual calculation, forward denoising, reverse denoising and a multi-scale feature extraction fusion network based on the UNET architecture. The second-stage denoising diffusion probability model takes the initial high-resolution precipitation field output by the first-stage convolutional neural network as input, introduces a multi-scale feature extraction mechanism, and combines the convolutional neural network (CNN) structure with the channel attention mechanism to extract high-resolution terrain information and spatial features of the initial high-resolution precipitation field. The residual convolution structure is used to extract different scale features with the help of the jump connection of the corresponding layer of UNET. Accurate prediction of the noise at the corresponding time step t is achieved. Gaussian white noise is randomly sampled from the standard normal distribution. The resolution of the Gaussian white noise is consistent with the resolution of the high-resolution precipitation field. Through the iterative denoising process, the noise is gradually reduced in each iteration, and the residual between the high-resolution precipitation field and the initial convolutional neural network output result is gradually restored. The residual is the high-frequency details of the high-resolution precipitation field, which is a detailed supplement to the initial prediction result.

[0045] The present invention uses high-resolution terrain information to guide the reverse denoising process to generate a more accurate high-resolution precipitation field. The optimization objective of the denoising diffusion probability model is defined as follows:

[0046]

[0047] Among them, cond terrain It is high-resolution terrain information data.

[0048] The residual calculation is based on two Markov chains to generate the true value x of high-resolution precipitation data. hr With the initial high-resolution precipitation data x cnn The residual x0 = x hr -x cnn .

[0049] The forward noise addition process gradually adds noise to the data:

[0050]

[0051] Among them, β t is a hyperparameter, indicating the level of noise added to the data, which is related to the time step t, β t ∈(0,1), t=1,…,T; α t =1-β t , The forward noise addition process is a parameter-free process. When T is large enough, x0 will be approximately Gaussian noise sampled from a standard normal distribution.

[0052] The multi-scale feature extraction fusion network is used to add noise to the data x t and high-resolution terrain information datacond terrain Fusion is performed at time step t.

[0053] Inverse denoising is used to process the initial high-resolution precipitation data x cnn and the high-resolution terrain information cond terrain to gradually restore accurate high-resolution precipitation data After that, the accurate high-resolution precipitation data and the initial high-resolution precipitation data x cnn are used to generate an accurate high-resolution precipitation field with high-frequency details.

[0054] The inverse denoising process is a process of reversely restoring the data to the true data distribution:

[0055]

[0056] where θ represents the model parameters, p θ is the conditional probability transition density function, represents the normal distribution, μ θ , Σ θ are the mean and variance respectively, and I is the identity matrix.

[0057] The multi-scale feature extraction and fusion UNET designed in the present invention uses multi-scale convolutional kernels and channel prior convolutional attention mechanisms to extract and fuse the features of high-resolution terrain information and high-resolution precipitation field information, so as to accurately predict the corresponding time step t timestep The noise ∈ to be removed t_pred , and this UNET uses the noise ∈ added at the time step t timestep added in the forward noise addition process as the true value, and takes the absolute error between ∈ t and the model output ∈ t as the loss function: t_pred

[0058]

[0059] where t is equivalent to t timestep , and ∈ is randomly sampled Gaussian white noise.

[0060] The inverse denoising process is executed as follows:

[0061]

[0062] where ∈ is randomly sampled Gaussian white noise. Through the above process, after iterating T steps, starting from a randomly sampled Gaussian noise, the high-frequency details of the high-resolution precipitation field are gradually restored under the guidance of the initial high-resolution precipitation field and high-resolution terrain information to generate an accurate high-resolution precipitation field with high-frequency details. The results of the high-resolution precipitation field generated in the embodiment can be seen Figure 4 . ​

[0063] The method of the present invention has been verified in actual cases. The results show that the precipitation forecast after spatial downscaling is significantly superior to the traditional method in terms of resolution and accuracy, and can effectively meet the needs of the spatial distribution of high-resolution precipitation. Through the present invention, more refined support for the spatial distribution of precipitation can be provided for fields such as large-scale events, urban waterlogging, and geological disaster prevention. The present invention has good application prospects and market value, can provide important technical guarantees for the intelligent processing and efficient utilization of meteorological data, and helps to improve the urban management and disaster response capabilities.

[0064] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A spatial downscaling method for precipitation fields, characterized in that Including: Obtain high-resolution terrain information data and precipitation paired data composed of high-resolution and low-resolution precipitation data, and perform normalization processing on both the high-resolution terrain information data and the precipitation paired data; input the normalized precipitation paired data and high-resolution terrain information data into a pre-trained precipitation field spatial downscaling model to output a high-resolution precipitation field; wherein, the precipitation field spatial downscaling model includes a convolutional neural network and a denoising diffusion probability model, and the denoising diffusion probability model is constructed based on residual calculation, forward noise addition, backward denoising, and a multi-scale feature extraction and fusion network based on the UNET architecture; The convolutional neural network is used to process the low-resolution precipitation data in the precipitation paired data to obtain initial high-resolution precipitation data ; during the model training process, a frequency-domain hybrid loss function is used to calculate the training loss of the convolutional neural network; ； Among them, , ; and are the frequency domain values obtained by applying the fast Fourier transform to the true value and the predicted value of the high-resolution precipitation field, respectively; , , are the high-frequency components after times wavelet transform of the true value; , , are the high-frequency components after times wavelet transform of the predicted value.

2. The precipitation field spatial downscaling method according to claim 1, characterized in that Low resolution refers to 3000-meter and 1000-meter resolutions, and high resolution refers to 750-meter resolution.

3. The precipitation field spatial downscaling method according to claim 1, wherein The optimization objective of the denoising diffusion probability model is defined as follows: ； Among them, is high-resolution terrain information data.

4. The precipitation field spatial downscaling method according to claim 3, characterized in that The residual calculation is based on two Markov chains to generate the true value of high-resolution precipitation data and the initial high-resolution precipitation data residuals .

5. The precipitation field spatial downscaling method according to claim 4, characterized in that The forward noise addition formula is: ； ； Among them, is a hyperparameter representing the level of noise added to the data, related to the time step and , , , is an adjustable hyperparameter; , , when is large enough, after the -step forward noise addition process, it will approach a Gaussian white noise sampled from a standard normal distribution.

6. The precipitation field spatial downscaling method according to claim 5, characterized in that The multi-scale feature extraction and fusion network is used to process the data with added noise , the initial high-resolution precipitation data and the high-resolution terrain information data are fused at each time step .

7. The precipitation field spatial downscaling method according to claim 6, characterized in that The backward denoising formula is: ； Among them, is the predicted value of the noise added to the corresponding forward time step of the UNET output ; is randomly sampled Gaussian white noise; By randomly sampling a Gaussian white noise and going through steps of iterative denoising, using the initial high-resolution precipitation data and the high-resolution terrain information as a guide to gradually restore the high-frequency details of the accurate high-resolution precipitation data , and then using the generated high-frequency details of the high-resolution precipitation and the initial high-resolution precipitation data to synthesize an accurate high-resolution precipitation field with high-frequency details.

Images