A method, equipment and medium for downscaling precipitation

Abstract

This invention belongs to the field of meteorological forecasting technology and discloses a precipitation downscaling method, equipment, and medium. The method includes: acquiring a low-resolution precipitation forecast field of a target area and upsampling it to obtain a basic high-resolution precipitation forecast field; acquiring the topographic features of the target area, classifying the target area into topographic types based on the topographic features, and generating topographic classification masks corresponding to each topographic type; fusing the topographic features and the topographic classification masks to obtain a topographic feature map; inputting the fused topographic feature map and the basic high-resolution precipitation forecast field into a topographic-aware variational autoencoder to obtain a latent space vector; using the topographic feature map as a condition, denoising the latent space vector using a diffusion model; and reconstructing the high-resolution precipitation forecast field from the denoised latent space vector using a variational autoencoder. By integrating topographic features into the downscaling process, the deep utilization of topographic information is achieved during the downscaling process, overcoming the problem that traditional methods are insufficient in depicting the details of precipitation in complex topographic areas.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to a precipitation downscaling method, equipment, and medium. Background Technology

[0002] Downscaling techniques aim to generate high-resolution information from low-resolution meteorological or climate data, and are an indispensable part of fields such as energy forecasting, hydrological simulation, and disaster management. Traditional downscaling approaches mainly fall into two categories: one is dynamic downscaling, which relies on high-resolution regional models for physical simulations. While it can ensure physical consistency, it consumes enormous computational resources and is sensitive to initial and boundary conditions. The other is statistical downscaling, which constructs a mapping by mining statistical correlations between low-resolution and high-resolution observations in historical data. This method is computationally efficient, but may perform poorly when dealing with complex terrain influences or non-stationary climate states, and it easily smooths out local details and extreme values.

[0003] Topography has a significant impact on precipitation: windward slopes generate orographic lifting and precipitation, while leeward slopes create a rain shadow effect; valley channels create a funneling effect that enhances local precipitation; and thermal and dynamic processes in plateau regions alter precipitation distribution patterns. Traditional downscaling methods often ignore or simplify the effects of topography, leading to systematic biases in precipitation forecasts for complex terrain regions and making it difficult to accurately characterize topographically induced enhancements or reductions in local precipitation. Furthermore, while existing deep learning downscaling methods can learn statistical patterns in data, most use uniform model parameters and generation strategies to process all regions, failing to differentiate precipitation characteristics for different terrain types. This results in models performing reasonably well in plains areas, but lacking the ability to depict detailed precipitation patterns in mountainous and plateau regions, and may even generate unreasonable results that contradict the physical relationship between topography and precipitation.

[0004] Currently, precipitation forecasts provided by meteorological numerical weather prediction models at global and regional scales still face the challenge of insufficient spatial resolution, which is particularly prominent in regions with diverse topography and extremely uneven precipitation distribution. Traditional downscaling methods have inherent limitations in capturing small-scale precipitation characteristics, local heavy precipitation centers, and extreme events. While deep learning techniques, especially convolutional neural networks (CNNs), have shown potential, there is still room for improvement in generating precipitation fields with high realism and fine texture; the output results may be blurry or overly smooth. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a precipitation downscaling method, device, and medium that integrates topographic features into the downscaling process. This enables in-depth utilization of topographic information during the generation of high-resolution precipitation forecast fields, allowing the final output high-resolution precipitation forecast field to not only match the macroscopic trends of low-resolution precipitation forecast fields but also depict the details of precipitation under complex terrain. This overcomes the problem of traditional methods' insufficient ability to depict the details of precipitation in complex terrain areas.

[0006] This invention provides the following technical solution:

[0007] The first objective of this invention is to provide a precipitation downscaling method, comprising:

[0008] Obtain the low-resolution precipitation forecast field of the target area, and upsample the low-resolution precipitation forecast field to the target resolution to obtain the basic high-resolution precipitation forecast field.

[0009] The terrain features of the target area are acquired, and the target area is divided into different terrain types according to the terrain features. Terrain classification masks corresponding to each terrain type are generated.

[0010] The terrain features and terrain classification mask are fused to obtain a terrain feature map, which includes multiple feature channels;

[0011] The terrain feature map is fused with the basic high-resolution precipitation forecast field and then input into the terrain-aware variational autoencoder to obtain the latent space vector.

[0012] Using the terrain feature map as a condition, the latent space vector is denoised using a diffusion model to obtain the denoised latent space vector.

[0013] The denoised latent space vectors are reconstructed into a high-resolution precipitation forecast field using a variational autodecoder.

[0014] As a further improvement of the present invention, the terrain features include altitude, slope, aspect, terrain relief, terrain location index and terrain humidity index;

[0015] The target area is divided into different terrain types based on altitude and slope, including plains, hills, mountains and plateaus.

[0016] As a further improvement of the present invention, the step of fusing terrain features and terrain classification masks to obtain a terrain feature map includes:

[0017] The terrain features are standardized, and the terrain classification mask is 01 encoded.

[0018] The standardized terrain features and the encoded terrain classification mask are resampled to the target resolution and stacked and fused in the channel dimension to obtain a multi-channel terrain feature tensor, which is the terrain feature map.

[0019] As a further improvement of the present invention, the basic high-resolution precipitation forecast field and the topographic feature map are stitched together in the channel dimension to form a joint input as input data;

[0020] The terrain-aware variational autoencoder is a convolutional neural network with an embedded attention mechanism. It extracts the spatial features of the input data layer by layer through the convolutional neural network, and uses the attention mechanism to dynamically calculate the correlation weights between the precipitation features of the precipitation forecast field and each feature channel of the terrain feature map, so as to compress the input data to a low-dimensional latent space to obtain the latent space vector.

[0021] As a further improvement of the present invention, the diffusion model is a conditional denoising network of the U-Net architecture, and the denoising condition is a multi-scale terrain feature pyramid obtained from the terrain feature map; wherein, the multi-scale terrain feature pyramid includes multi-scale terrain feature maps, and each scale corresponds to the spatial resolution of each decoding layer of the U-Net architecture.

[0022] As a further improvement of the present invention, the multi-scale terrain feature pyramid is obtained from the terrain feature map by a terrain encoder;

[0023] The terrain encoder is a lightweight multi-branch convolutional downsampling network, wherein the downsampling ratio of each convolutional branch corresponds to the spatial resolution of each decoding layer of the U-Net architecture, and each convolutional branch processes the terrain feature map in parallel to obtain a multi-scale terrain feature pyramid.

[0024] As a further improvement of the present invention, in each decoding layer of the U-Net architecture, a terrain feature map of the corresponding scale is introduced as a denoising condition through a cross-attention mechanism.

[0025] As a further improvement of the present invention, the cross-attention mechanism includes:

[0026] The feature maps obtained from each decoding layer of the U-Net architecture are used as query vectors, and the corresponding scale terrain feature maps are used as key vectors and value vectors. Attention is calculated, and the result of the attention calculation is used as an embedding vector to be introduced into the denoising process.

[0027] A second objective of this invention is to provide a computer device comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the aforementioned method.

[0028] A third objective of this invention is to provide a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] By explicitly incorporating topographic features into the downscaling process, the topographic feature map and precipitation forecast field are jointly encoded in the encoding stage. In the diffusion model denoising stage, the topographic feature map is introduced as a denoising condition. This enables in-depth utilization of topographic information in the generation of high-resolution precipitation forecast fields. The final output high-resolution precipitation forecast field can not only match the macroscopic trend of low-resolution precipitation forecast fields, but also depict the precipitation details under complex terrain. This overcomes the problem that traditional methods are not good at depicting the precipitation details in complex terrain areas.

[0031] By simultaneously characterizing precipitation and topographic features in a low-dimensional latent space and performing complex generation calculations in the latent space, the computational cost is significantly reduced while ensuring the quality of high-resolution precipitation forecast field generation, making high-resolution downscaling with topographic information feasible in practical operations.

[0032] During the encoding stage, an attention mechanism is used to dynamically calculate the correlation weights between precipitation and various topographic features, enabling adaptive fusion of precipitation forecast fields and topographic feature maps at the feature level. In the latent space vector denoising stage, a cross-attention mechanism is used to introduce topographic feature maps of corresponding scales into each decoding layer of the U-Net architecture, guiding the diffusion model to generate precipitation details based on topographic context information at different resolutions. This phased, multi-scale fusion mechanism ensures that the generation process of high-resolution precipitation forecast fields is subject to precise topographic constraints, enabling more accurate depiction of precipitation details in complex topographic regions. Attached Figure Description

[0033] Figure 1 A flowchart of the precipitation downscaling method provided by the present invention;

[0034] Figure 2 A model structure diagram of the precipitation downscaling method provided by this invention;

[0035] Figure 3 A flowchart for obtaining multi-scale terrain feature towers from terrain feature maps. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0037] The present invention will now be described in further detail with reference to the accompanying drawings:

[0038] like Figure 1 As shown, this embodiment provides a precipitation downscaling method, including:

[0039] The process involves acquiring a low-resolution precipitation forecast field for the target area, upsampling the low-resolution precipitation forecast field to the target resolution, and obtaining a base high-resolution precipitation forecast field. Specifically, acquiring the low-resolution precipitation forecast field for the target area includes: acquiring meteorological data, inputting the meteorological data into the Fuxi meteorological model, generating an initial forecast field based on the meteorological data, separating the low-resolution precipitation forecast field from the initial forecast field, and cropping the target area from the low-resolution precipitation forecast field to obtain a 0.25° low-resolution precipitation forecast field for the target area.

[0040] The Fuxi meteorological model is a deep learning-based meteorological forecasting system that employs a cascaded machine learning architecture to efficiently forecast global meteorological fields. Its main features include:

[0041] Multi-scale feature extraction: Atmospheric dynamic features at different scales are captured through multi-layer convolutional neural networks;

[0042] Physical constraint fusion: Incorporating physical constraints during model training to ensure that forecast results conform to atmospheric dynamics.

[0043] Provides global forecasts: Supports global grid forecasts of 0.25°×0.25° with time resolution down to the hourly level.

[0044] In this embodiment, the Fuxi meteorological model is used to generate an initial forecast field. Then, a 0.25° low-resolution precipitation forecast field for the target region is separated and cropped from the initial forecast field. The 0.25° low-resolution precipitation forecast field for the target region is then upsampled to the target resolution to obtain a basic high-resolution precipitation forecast field, which provides input for subsequent downscaling. In this embodiment, the target resolution is 0.01°.

[0045] The terrain features of the target area are acquired, and the target area is divided into different terrain types according to the terrain features. Terrain classification masks corresponding to each terrain type are generated.

[0046] The terrain features are extracted from the high-precision digital elevation model and include:

[0047] Elevation; slope, indicating the degree of inclination of the land surface, which affects precipitation runoff and local circulation; aspect, i.e. the orientation of the slope, which determines the windward / leeward slope effect; topographic relief, the degree of drastic change in local surface elevation; topographic position index, indicating the current location relative to the surrounding terrain; topographic humidity index, reflecting the potential for water accumulation.

[0048] After extracting terrain features, the target area is divided into different terrain types based on elevation and slope. The criteria for terrain type division are as follows:

[0049] Plains: Elevation <200m, slope <2°; characterized by flat terrain, with precipitation mainly controlled by large-scale weather systems;

[0050] Hilly area: The elevation ranges from 200 to 500 meters, and the slope ranges from 2 to 8 degrees; the terrain is characterized by relatively small undulations, which has a certain moderating effect on precipitation.

[0051] Mountainous areas: elevation ranges from 500 to 3000 meters, slope > 8°; characterized by large topographic relief and significant orographic lifting effect.

[0052] Plateau region: Altitude > 3000m; characterized by a cold climate, large diurnal temperature range, and high precipitation variability, mostly in the form of intermittent precipitation.

[0053] The terrain features and terrain classification masks are fused to obtain a terrain feature map, specifically including:

[0054] Terrain features are processed: All terrain features are standardized, and the standardized terrain features are uniformly resampled to 0.01° resolution using bilinear interpolation.

[0055] The terrain classification mask is processed by 0-1 encoding to convert the terrain classification information into a numerical terrain classification feature map, and the terrain classification feature map is resampled to 0.01° resolution.

[0056] The processed terrain features are stacked and fused with the terrain classification mask in the channel dimension to obtain a multi-channel terrain feature tensor with a total number of channels C_t. The multi-channel terrain feature tensor is the terrain feature map.

[0057] The terrain feature map is fused with the basic high-resolution precipitation forecast field and then input into the terrain-aware variational autoencoder to obtain the latent space vector. Specifically, the basic high-resolution precipitation forecast field and the terrain feature map are concatenated in the channel dimension to form a joint input as input data. The total number of channels of the input data is 1+C_t.

[0058] The terrain-aware variational autoencoder is a convolutional neural network with an embedded attention mechanism. It extracts the spatial features of the input data layer by layer through the convolutional neural network, and uses the attention mechanism to dynamically calculate the correlation weights between the precipitation features of the precipitation forecast field and each feature channel of the terrain feature map, so as to compress the input data into a low-dimensional latent space and thus obtain the latent space vector.

[0059] The terrain-aware variational autoencoder can not only extract precipitation and terrain features from the input data, but also actively and selectively fuse precipitation features with terrain features according to the terrain context. Finally, it maps the high-dimensional input data into a low-dimensional, highly condensed latent space vector, which simultaneously encodes the initial precipitation distribution pattern and deep terrain constraint information.

[0060] like Figure 2 As shown, using the terrain feature map as a condition, a diffusion model is used to denoise the latent space vector to obtain the denoised latent space vector. Specifically, the diffusion model is a conditional denoising network of the U-Net architecture, and the multi-scale terrain feature pyramid obtained from the terrain feature map serves as the sole external condition to guide the denoising process, providing static spatial physical constraints for the denoising process; wherein, the multi-scale terrain feature pyramid includes multi-scale terrain feature maps, and each scale corresponds to the spatial resolution of each decoding layer of the U-Net architecture.

[0061] Furthermore, the multi-scale terrain feature pyramid is obtained from the terrain feature map by the terrain encoder; such as Figure 3 As shown, the terrain encoder is a lightweight multi-branch convolutional downsampling network, wherein the downsampling ratio of each convolutional branch corresponds to the spatial resolution of each decoding layer of the U-Net architecture, and each convolutional branch processes the terrain feature map in parallel to obtain a multi-scale terrain feature pyramid.

[0062] During denoising, in each decoding layer of the U-Net architecture, a corresponding scale terrain feature map is introduced as a denoising condition through a cross-attention mechanism; wherein, the cross-attention mechanism includes:

[0063] The feature maps obtained from each decoding layer of the U-Net architecture are used as query vectors, and the corresponding scale terrain feature maps are used as key vectors and value vectors. Attention is calculated, and the attention calculation results are used as embedding vectors to be introduced into the denoising process.

[0064] After a complete denoising iteration, the denoised latent space vector is obtained. The denoised latent space vector is rich in detail and highly coordinated with the micro-topographic features.

[0065] Finally, the variational autodecoder is used to reconstruct the high-resolution precipitation forecast field from the denoised latent space vector. The variational autodecoder is a standard convolutional neural network consisting of a series of upsampling layers and convolutional layers, which is responsible for mapping the denoised low-dimensional latent space vector back to the high-dimensional pixel space.

[0066] In this process, the variational autodecoder operates independently without introducing external terrain information. Its goal is to reconstruct the precipitation pattern, which has been optimized through a diffusion denoising process that incorporates terrain features, as accurately as possible from the denoised latent space vector. By performing a complex generation process in the latent space rather than the original pixel space, computational efficiency can be significantly improved while ensuring generation quality, making it suitable for high-resolution downscaling tasks.

[0067] This embodiment provides a specific implementation process for the above-mentioned precipitation downscaling method.

[0068] Obtain global meteorological data X at two consecutive time steps (time t-1 and time t). t-1 and X t The data contains dynamic and thermodynamic information driving atmospheric evolution, with dimensions of 70×721×1440; where 70 represents the number of meteorological variable layers, 721 represents the number of latitude grid points, and 1440 represents the number of longitude grid points, and the resolution of the data is 0.25°.

[0069] The Fuxi meteorological model is based on X t-1 and X t It uses its deep learning network to predict the global weather conditions at the next time step (t+1), thus obtaining deterministic forecast results. Deterministic forecast results The dimensions are 70×721×1440, and the resolution is 0.25°.

[0070] From deterministic forecast results The low-resolution (LR) precipitation forecast field is extracted and the target region is cropped to obtain the target region's low-resolution precipitation forecast field, which is a tensor with a resolution of 0.25° and dimensions of 1×180×280. Bilinear interpolation is then performed on the target region's low-resolution precipitation forecast field to upscale it to the target high-resolution (HR) field, resulting in the base high-resolution precipitation forecast field HR_interp, which is a tensor with a resolution of 0.01° and dimensions of 1×4500×7000.

[0071] The terrain features of the target area are extracted from the high-precision digital elevation model. The terrain features include altitude, slope, aspect, terrain relief, terrain location index and terrain humidity index. Based on the altitude and slope, the target area is divided into four terrain types: plain, hill, mountain and plateau. A terrain classification mask corresponding to each terrain type is generated.

[0072] All terrain features are standardized, and bilinear interpolation is used to uniformly resample the standardized terrain features to a resolution of 0.01°. The terrain classification mask is 0-1 encoded to convert the terrain classification information into a numerical terrain classification feature map, and the terrain classification feature map is resampled to a resolution of 0.01°. The processed terrain features and the terrain classification mask are stacked and fused in the channel dimension to obtain a multi-channel terrain feature tensor, i.e., the terrain feature map Terrain_features, with a resolution of 0.01° and a dimension of 7×4500×7000.

[0073] HR_interp (1×4500×7000) and Terrain_features (7×4500×7000) are concatenated along the channel dimension to form the joint input Input_combined, which has a dimension of 8×4500×7000.

[0074] Input_combined is fed into the terrain-aware variational autoencoder, which uses an attention mechanism to learn the interaction between precipitation and terrain, mapping high-dimensional data into a low-dimensional latent space vector Z_interp, which captures both precipitation patterns and terrain constraint information.

[0075] The latent space vector Z_interp is denoised using a diffusion model. The diffusion model is a conditional denoising network based on the U-Net architecture. Through iterative denoising, it generates a more detailed and structurally sound refined latent space representation Z_refined with dimensions of 4×563×875.

[0076] During the denoising process, the terrain feature map is downsampled to different scales, and the terrain features at different scales are injected into the corresponding decoding layers of the U-Net architecture through a cross-attention mechanism to guide the diffusion model to generate precipitation details that conform to the terrain constraints.

[0077] The refined latent space representation Z_refined is input into the variational autodecoder, which is a standard convolutional network, to reconstruct the final high-resolution precipitation forecast field HR_final from Z_refined. The dimension of HR_final is 1×4500×7000.

[0078] The final output high-resolution precipitation forecast field HR_final for the target area not only improves spatial resolution, but more importantly, enhances the detail realism, spatial texture representation, and local feature characterization of the high-resolution precipitation forecast field through the generative capabilities of the latent space diffusion model. Simultaneously, it ensures that the output high-resolution precipitation forecast field maintains consistency with the input low-resolution precipitation forecast field in overall trend. The final generated high-resolution precipitation forecast data is stored in a standard format (such as NetCDF) for distribution to downstream applications such as flood warning, water resource management, and agricultural meteorology through a dedicated data distribution system.

[0079] This embodiment provides a computer device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described precipitation downscaling method.

[0080] This embodiment provides a computer-readable storage medium storing a computer program executable by an electronic device. When the program is run on the electronic device, it causes the electronic device to perform the above-described precipitation downscaling method.

[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A precipitation downscaling method, characterized in that, The method comprises the following steps: obtaining a low-resolution precipitation forecast field of a target area, upsampling the low-resolution precipitation forecast field to a target resolution to obtain a basic high-resolution precipitation forecast field; obtaining terrain features of the target area, dividing the target area into different terrain types according to the terrain features, and generating a terrain classification mask corresponding to each terrain type; fuse the terrain features and the terrain classification mask to obtain a terrain feature map, the terrain feature map comprising a plurality of feature channels; fuse the terrain feature map and the basic high-resolution precipitation forecast field to input a terrain-aware variational autoencoder to obtain a latent space vector, comprising: concatenating the basic high-resolution precipitation forecast field and the terrain feature map in the channel dimension to form a joint input as input data; the terrain-aware variational autoencoder is a convolutional neural network embedded with attention mechanism, which extracts spatial features of the input data layer by layer through the convolutional neural network, and dynamically calculates the correlation weight between the precipitation features of the precipitation forecast field and each feature channel of the terrain feature map through the attention mechanism, so as to compress the input data to a low-dimensional latent space to obtain a latent space vector; use the diffusion model to denoise the latent space vector under the condition of the terrain feature map to obtain a denoised latent space vector; use the variational auto-decoder to reconstruct the denoised latent space vector into a high-resolution precipitation forecast field.

2. The method of claim 1, wherein, The terrain features include elevation, slope, aspect, terrain relief, terrain position index and terrain humidity index. According to the elevation and slope, the target area is divided into different terrain types, including plain area, hilly area, mountainous area and plateau area.

3. The method of claim 1, wherein, fuse the terrain features and the terrain classification mask to obtain a terrain feature map, comprising: standardize the terrain features, and 01 encode the terrain classification mask; resample the standardized terrain features and the encoded terrain classification mask to the target resolution, and stack and fuse in the channel dimension to obtain a multi-channel terrain feature tensor, which is the terrain feature map.

4. The method of claim 1, wherein, The diffusion model is a conditional denoising network with U-Net architecture, and the denoising condition is a multi-scale terrain feature pyramid obtained from the terrain feature map; wherein the multi-scale terrain feature pyramid comprises multi-scale terrain feature maps, and each scale corresponds to the spatial resolution of each decoding layer of the U-Net architecture.

5. The method of claim 4, wherein, The multi-scale terrain feature pyramid is obtained from the terrain feature map by a terrain encoder; The terrain encoder is a lightweight multi-branch convolution downsampling network, wherein the downsampling rate of each convolution branch corresponds to the spatial resolution of each decoding layer of the U-Net architecture, and each convolution branch processes the terrain feature map in parallel to obtain the multi-scale terrain feature pyramid.

6. The method of claim 4, wherein, In each decoding layer of the U-Net architecture, the corresponding scale terrain feature map is introduced as the denoising condition through the cross-attention mechanism.

7. The method of claim 6, wherein, The cross-attention mechanism comprises: using the feature map obtained from each decoding layer of the U-Net architecture as the query vector, using the corresponding scale terrain feature map as the key vector and the value vector, performing attention calculation, and introducing the attention calculation result as the embedding vector into the denoising process.

8. A computer device, comprising: A computer program product comprising at least one storage medium to store the computer program according to claim 8, wherein said at least one storage medium is a non-transitory storage medium.

9. A computer-readable storage medium, characterized in that, A computer program product comprising at least one storage medium to store the computer program according to claim 8, wherein said at least one storage medium is a non-transitory storage medium.

Images