A system and method for fine-grained simulation of precipitation and runoff in mountainous watersheds

Abstract

This invention relates to the field of runoff simulation technology, specifically to a system and method for fine-grained simulation of precipitation and runoff in mountainous watersheds. The system comprises the following steps: Step 1: Dataset and data preprocessing; Step 2: Designing and training a U-Net super-resolution model (TA-UNet) based on a terrain attention mechanism as a precipitation downscaling model; Step 3: Accuracy verification; Step 4: Precipitation downscaling; and Step 5: Precipitation and runoff simulation. This system can better capture the spatial characteristics of precipitation in mountainous watersheds and simulate the propagation of precipitation and runoff in detail.

Description

Technical Field

[0001] The present invention relates to the technical field of runoff simulation, and in particular to a system and method for fine-grained simulation of precipitation and runoff in a mountainous watershed. Background Art

[0002] Heavy rainfall-induced runoff in small watersheds can lead to secondary hazards such as flash floods, landslides, and debris flows, which have caused significant casualties and economic losses worldwide. The triggering of these hazards depends on multiple factors, including hydrological conditions, geomorphological characteristics, slope characteristics, and sediment source volume, with precipitation being a key driver. Precipitation data observations directly impact the timeliness and accuracy of geological hazard forecasts. However, sparsely distributed surface rain gauges and coarse-resolution weather radars struggle to accurately reflect local precipitation patterns at the watershed scale. By analyzing field measurements and historical data, researchers have discovered that unique local precipitation patterns are shaped by microtopography. These characteristics suggest that precipitation intensities in debris flow source areas can be far higher than the hazard thresholds provided by local meteorological stations at lower elevations. Significant variability in local precipitation significantly increases the unpredictability of runoff-induced secondary hazards. Therefore, there is an urgent need to obtain high-resolution precipitation data that accurately reflects local precipitation patterns in small watersheds under complex terrain.

[0003] Recent advances in satellite remote sensing technology have made it an effective tool for obtaining spatially continuous and temporally comprehensive estimates of global precipitation. Numerous meteorological reanalysis products derived from satellite remote sensing, such as TRMM, IMERG GPM, ERA5-Land, and CLDAS, effectively perform statistical analysis of large-scale precipitation distribution and trends, serving as key resources for regional precipitation analysis. Researchers have subsequently utilized dynamical and statistical downscaling techniques to improve the spatial resolution of meteorological satellite observations. Dynamical downscaling techniques, known as regional climate models, integrate topography, meteorological factors, and other mathematical and physical factors to construct and simulate atmospheric thermodynamic equations to calculate local climate fields. While capable of accurately simulating climate models with well-defined physical mechanisms, they require significant computational resources and have limited spatial generalization capabilities. In contrast, statistical downscaling methods utilize machine learning methods, such as random forests, support vector machines, and neural networks, based on the assumption of spatial correlation and scale invariance between precipitation and various meteorological elements. Numerous statistical models have been developed to describe nonlinear relationships between coarse and fine resolutions. These models employ machine learning methods to extract characteristic relationships between coarse-resolution meteorological underlying surface factors and precipitation, thereby predicting the spatial distribution of precipitation at fine-resolution meteorological underlying surface factors. Statistical models based on point-to-point precipitation prediction often ignore the spatial information between environmental factors and precipitation. Deep learning methods have proven to be more capable of capturing high-dimensional and complex spatial features. Single-image super-resolution models, most notably SRCNN, generate super-resolution images by learning the intrinsic image features and the mapping relationship between coarse and fine resolution. Considering meteorological downscaling as an image super-resolution task, researchers proposed DeepSD, which uses stacked SRCNNs to demonstrate superior downscaling performance on Earth system models (ESMs), outperforming other statistical downscaling models. Subsequent research has made significant progress in extreme precipitation and precipitation forecasting by integrating additional meteorological auxiliary information into neural network models. Furthermore, researchers have scientifically explained how CNN-based downscaling models learn the physical relationship between the large-scale atmospheric environment and extreme local precipitation events. More improved downscaling models based on super-resolution algorithms, such as FSRCNN, EPSCN, and ResLap, have outperformed traditional statistical models in computational efficiency, prediction accuracy, and feature extraction. Related research has primarily focused on downscaling precipitation to annual and monthly timescales, where precipitation interacts significantly with meteorological factors. However, as precipitation resolution increases, its spatial discontinuity increases, making downscaling to daily or hourly precipitation more challenging. Meanwhile, runoff simulations currently rely on coarse-resolution rain gauge observations, meteorological bureau forecasts, and simple interpolation methods (such as bilinear interpolation and Kriging interpolation). Research has shown that such precipitation data significantly limits computational accuracy.Due to the different focus scales of meteorological and hydrological research, few studies have considered satellite precipitation downscaling based on basin climate characteristics and applied it to runoff simulation. Summary of the Invention

[0004] The present invention provides a system and method for fine-grained simulation of precipitation and runoff in mountainous watersheds, which can better capture the spatial precipitation characteristics of mountainous watersheds and simulate the precipitation and runoff propagation process in detail.

[0005] A method for fine-tuning precipitation-runoff simulation in a mountainous watershed according to the present invention comprises the following steps:

[0006] Step 1: Dataset and data preprocessing;

[0007] Step 2: Design a U-Net super-resolution model TA-UNet based on terrain attention mechanism as a precipitation downscaling model and train it;

[0008] Step 3: Accuracy verification;

[0009] Step 4: downscaling precipitation;

[0010] Step 5: Precipitation-runoff simulation.

[0011] Preferably, in step 1, the half-hourly IMERG 0.1° precipitation dataset of recent years, as well as relevant topographic meteorological elements and ground rain gauge observation data, are collected and collated; the satellite precipitation data and topography are resampled to 0.5° to establish a training dataset.

[0012] As a preference, in step 2, the TA-UNet model is improved as follows:

[0013] 1) Add an additional upsampling module to the Attention U-Net model to upsample the coarse-resolution data to match the fine-resolution data grid;

[0014] 2) Incorporating multi-scale terrain data through attention gates; in the TA-UNet model, the attention gate acts as a self-attention feature layer between the encoder and decoder feature layers, as well as an attention feature layer between terrain and precipitation information. The representation of the attention gate is as follows:

[0015] q att =ψ(σ1((w x x input +b x )+(w g g+b g )))+b ψ

[0016] α=σ2(q att (x input ,g;Θatt ))

[0017] σ1(x)=max(0,x)

[0018]

[0019] where q att represents the attention feature layer; σ1 and σ2 are ReLU and Sigmoid activation functions respectively, x var is the independent variable, c is a constant; Θ att Represents the trainable parameters of the neural network, i.e. ψ,w x ,b x ,w g and b g ;q att The attention coefficient α is obtained by normalizing the sigmoid function. The attention mechanism is aimed at the feature layer input x input The attention coefficient is g, further weighting the feature layer x input ;

[0020] 3) Replace the deconvolution layer used for upsampling with pixel reorganization PixelShuffle.

[0021] Preferably, in step 2, the training process is as follows:

[0022] The dataset was used to train a 5x precipitation downscaling model until the error converged. The model was saved and a 0.02° downscaling prediction was performed. The PyTorch deep learning framework was used with a training batch size of 64 and epochs of 50. The ADAM optimizer was used with a learning rate of 1e-4. The Exponential LR learning rate adjustment strategy was set, and the gamma value was set to 0.95.

[0023] Preferably, in step 4, specifically: using satellite precipitation data at 0.1° as model input, predicting 0.02° precipitation downscaled data.

[0024] Preferably, in step 5, specifically: coupling the precipitation downscaling model and the hydrodynamic model to perform runoff forecast simulation on a 10m Cartesian coordinate grid.

[0025] As a preferred method, the hydrodynamic model is based on the two-dimensional depth-averaged shallow water equations, assuming that the pressure distribution of the water body is hydrostatic pressure, the channel slope is minimal, the fluid density is uniform, and there is a free surface and a fixed bottom boundary. In matrix form, the shallow water equations SWEs are expressed as:

[0026]

[0027] Where t represents time; x and y are Cartesian coordinates; U is the vector of conservative physical variables; F and G are the flux vectors along the x and y directions respectively; the source vectors include precipitation R, riverbed slope S b Friction slope S with riverbed f ; These vectors are represented as follows:

[0028]

[0029] Where h represents the water depth; u and v are the depth-averaged velocity components in the x and y directions, respectively; I represents the precipitation intensity; g is the acceleration of gravity; Z b is the slope height; S fx and S fy is the friction stress estimated by Manning's formula, which can be expressed as n b is the Manning friction coefficient, set to 0.01.

[0030] As a preferred method, the hydrodynamic numerical simulation method uses the Godunov finite volume method to solve the shallow water equation. The Godunov finite volume method can handle the exact or approximate Riemann problem at the boundary of the calculation unit; the HLLC approximate method is selected to handle the Riemann problem at the unit interface; the finite volume format derived from the shallow water equation is as follows:

[0031]

[0032] That is, the differential form of the time step after numerical discretization is:

[0033]

[0034] Where, the superscript n represents the time step; the subscript i represents the grid unit number, and f E and f W corresponds to the flux to the east and west respectively, and g N and g S represent the flux to the north and south respectively, Δt is the time step selected by the CFL condition, and its expression is Δt=cflmin(Δt x ,Δt y ),in cfl∈[0,1]; the iterative process is performed uniformly over the overall time T, and the precipitation intensity R(I,t) is synchronized on the time scale based on the downscaled precipitation of each consecutive half-hour.

[0035] The present invention provides a system for fine-grained simulation of precipitation and runoff in a mountainous area, which adopts the above-mentioned method for fine-grained simulation of precipitation and runoff in a mountainous area.

[0036] This paper proposes a refined precipitation-runoff simulation system and method for mountainous watersheds, suitable for small watersheds in mountainous areas. By coupling an improved downscaling model with a hydrodynamic model, this method can better capture the spatial characteristics of precipitation in mountainous watersheds and simulate the rainfall-runoff propagation process in detail. By introducing an attention mechanism that considers topographic features, a TA-UNet downscaling model based on the terrain attention mechanism is proposed, which efficiently and conveniently downscaling IMERG satellite precipitation data to 0.02°. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of a method for fine-tuning precipitation-runoff simulation in a mountainous watershed according to an embodiment;

[0038] Figure 2 This is a structural diagram of the TA-UNet model in the embodiment;

[0039] Figure 3 This is a comparison chart of the TA-UNet downscaling results and the Sichuan Province ground station verification in the embodiment;

[0040] Figure 4 This is a comparison chart of the TA-UNet downscaling results in the embodiment and the Jiangjiagou rain gauge verification. DETAILED DESCRIPTION

[0041] In order to further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and embodiments. It should be understood that the embodiments are merely for explaining the present invention and are not intended to limit the present invention.

[0042] Example

[0043] like Figure 1 As shown, this embodiment provides a method for fine-grained simulation of precipitation and runoff in a mountainous watershed. Specifically, a method for fine-grained simulation of precipitation and runoff in a mountainous watershed based on satellite precipitation downscaling and a hydrodynamic model is provided, which includes the following steps:

[0044] Step 1: Dataset and data preprocessing;

[0045] In step 1, the half-hourly IMERG 0.1° precipitation dataset from recent years, along with relevant topographic meteorological elements and ground-based rain gauge observation data, was collected and collated. The satellite precipitation data and topography were resampled to 0.5° to establish a training dataset.

[0046] Step 2: Design a U-Net super-resolution model TA-UNet based on terrain attention mechanism as a precipitation downscaling model and train it;

[0047] In step 2, the improvements to the TA-UNet model are as follows:

[0048] 1) The Attention U-Net model was originally designed for classification tasks to identify spatially distributed features in images, but it lacks super-resolution capabilities. To adapt to downscaling tasks, an additional upsampling module (nearest neighbor interpolation) is added to the original Attention U-Net model to upsample the coarse-resolution data to match the fine-resolution data grid.

[0049] 2) Incorporating multi-scale terrain data through an attention gate; by applying the attention mechanism, the network's ability to focus on spatial features in terrain and precipitation data is enhanced to improve the efficiency and accuracy of feature extraction; the modified attention gate can specifically extract features from terrain elements based on the attention mechanism.

[0050] The attention gate acts as a self-attention feature layer between the encoder and decoder feature layers in the TA-UNet model, as well as an attention feature layer between terrain and precipitation information. The representation of the attention gate is as follows:

[0051] q att =ψ(σ1((w x x input +b x )+(w g g+b g )))+b ψ

[0052] α=σ2(q att (x input ,g;Θ att ))

[0053] σ1(x)=max(0,x)

[0054]

[0055] where q att represents the attention feature layer, σ1 and σ2 are ReLU and Sigmoid activation functions respectively, Θ att Represents the trainable parameters of the neural network, i.e. ψ,w x ,b x ,w g and b g ;q att The attention coefficient α is obtained by normalizing the sigmoid function. The attention mechanism is aimed at the feature layer input x input The attention coefficient is g, further weighting the feature layer x input ;

[0056] Compared with the original Attention U-Net model structure, the attention gate, as a self-attention mechanism, has excellent performance in identifying the latent inherent features between feature layers;

[0057] 3) The "deconvolution layer" used for upsampling usually produces discontinuous data and leads to a series of strange checkerboard artifacts. This embodiment replaces the deconvolution layer used for upsampling with pixel reorganization PixelShuffle, enabling the model to generate smoother and higher-quality downscaled prediction data.

[0058] like Figure 2 Figure 1 shows a schematic diagram of the TA-UNet architecture. "Addition" represents element-wise addition of tensors, "multiplication" represents the Hadamard product of tensors, and "concatenation" represents the concatenation of multiple tensors along a given dimension. "Downsampling" reduces the spatial resolution of the input tensor, using a max pooling algorithm for 2x downsampling. Similarly, "upsampling" uses a 5x nearest neighbor interpolation algorithm and a 2x pixel shuffle algorithm.

[0059] The training process is as follows:

[0060] The dataset was used to train a 5x precipitation downscaling model until the error converged. The model was saved and a 0.02° downscaling prediction was performed. The PyTorch deep learning framework was used with a training batch size of 64 and epochs of 50. The ADAM optimizer was used with a learning rate of 1e-4. The Exponential LR learning rate adjustment strategy was set, and the gamma value was set to 0.95.

[0061] Step 3: Accuracy verification;

[0062] The TA-UNet model outperforms bilinear interpolation methods and benchmark models such as SRCNN, FSRCNN, and EPSCN. Furthermore, validation using observational data from meteorological stations in Sichuan Province and a ground-based rain gauge in Jiangjiagou demonstrates that the TA-UNet downscaling results are more accurate than raw satellite precipitation data.

[0063] Step 4: downscaling precipitation;

[0064] In step 4, specifically: use the satellite precipitation data at 0.1° as the model input and predict the precipitation downscaled data at 0.02°.

[0065] Step 5: precipitation-runoff simulation;

[0066] In step 5, specifically: couple the precipitation downscaling model with the hydrodynamic model and perform runoff forecast simulation on a 10m Cartesian coordinate grid.

[0067] The hydrodynamic model is based on the two-dimensional depth-averaged shallow water equations, assuming that the pressure distribution of the water body is hydrostatic pressure, the channel slope is minimal, the fluid density is uniform, and there is a free surface and a fixed bottom boundary. In matrix form, the shallow water equations SWEs are expressed as:

[0068]

[0069] Where t represents time; x and y are Cartesian coordinates; U is the vector of conservative physical variables; F and G are the flux vectors along the x and y directions respectively; the source vectors include precipitation R, riverbed slope S b Friction slope S with riverbed f ; These vectors are represented as follows:

[0070]

[0071] Where h represents the water depth; u and v are the depth-averaged velocity components in the x and y directions, respectively; I represents the precipitation intensity; g is the acceleration of gravity; Z b is the slope height; S fx and S fy is the friction stress estimated by Manning's formula, which can be expressed as n b is the Manning friction coefficient, set to 0.01.

[0072] The numerical simulation method of hydrodynamics uses the Godunov finite volume method to solve the shallow water equation. The Godunov finite volume method can handle the exact or approximate Riemann problem at the boundary of the computational unit. In order to improve the efficiency and robustness of solving the Riemann problem, the HLLC approximation method is selected to handle the Riemann problem at the unit interface. The finite volume format derived from the shallow water equation is as follows:

[0073]

[0074] That is, the differential form of the time step after numerical discretization is:

[0075]

[0076] Where, the superscript n represents the time step; the subscript i represents the grid unit number, and f E and f W corresponds to the flux to the east and west respectively, and g N and g S represent the flux to the north and south respectively, Δt is the time step selected by the CFL condition, and its expression is Δt=cflmin(Δt x ,Δt y ),in cfl∈[0,1]; the iterative process is performed uniformly over the overall time T, and the precipitation intensity R(I,t) is synchronized on the time scale based on the downscaled precipitation of each consecutive half-hour.

[0077] This embodiment provides a system for fine-grained simulation of precipitation and runoff in a mountainous watershed, which adopts the above-mentioned method for fine-grained simulation of precipitation and runoff in a mountainous watershed.

[0078] This embodiment proposes a refined precipitation-runoff simulation system and method that combines a precipitation downscaling model and a hydrodynamic model. By introducing an attention mechanism that focuses on terrain features, the U-Net model (TA-UNet) based on the terrain attention mechanism realizes an efficient and convenient data preprocessing process for satellite precipitation downscaling from 0.1° to 0.02°. Compared with bilinear interpolation, TA-UNet improves the mean square error (MSE), mean absolute error (MAE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) by 34.0%, 29.9%, 4.4% and 0.6%, respectively. In addition, it performs well in meteorological forecast indicators such as VHI, VMI, VFAR and VCSI, as shown in Table 1.

[0079] Table 1 Performance evaluation index of downscaling methods

[0080]

[0081]

[0082] By verifying the ground observation data of 156 national meteorological stations in Sichuan Province and 8 rain gauges in Jiangjiagou, TA-UNet can effectively downscale the IMERG GPM products to provide higher-precision local precipitation. Figure 3 As shown in the figure, the TA-UNet downscaling results are compared with the Sichuan Province ground station verification. Figure 4 The figure shows the comparison between the TA-UNet downscaling results and the Jiangjiagou rain gauge verification.

[0083] Furthermore, the high-precision downscaled precipitation data demonstrated a high sensitivity to runoff responses in the Jiangjiagou rainfall-runoff simulation. This refined approach better captures the spatial characteristics of precipitation in mountainous watersheds and more accurately simulates the rainfall-runoff propagation process, providing an innovative and efficient forecasting strategy for hydrological management.

[0084] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A method for fine-tuning precipitation-runoff simulation in a mountainous watershed, characterized by: The following steps are involved: Step 1: Dataset and data preprocessing; Step 2: Design a U-Net super-resolution model TA-UNet based on terrain attention mechanism as a precipitation downscaling model and train it; In step 2, the improvements to the TA-UNet model are as follows: 1) Add an additional upsampling module to the AttentionU-Net model to upsample the coarse-resolution data to match the fine-resolution data grid; 2) Incorporating multi-scale terrain data through attention gates; The attention gate acts as a self-attention feature layer between the encoder and decoder feature layers in the TA-UNet model, as well as an attention feature layer between terrain and precipitation information. The representation of the attention gate is as follows: q att =ψ(σ1((w x x input +b x )+(w g g+b g )))+b ψ α=σ2(q att (x input ,g;Θ att )) σ1(x var )=max(0,x var ) where q att Represents the attention feature layer; σ1 and σ2 are ReLU and Sigmoid activation functions respectively, x var is the independent variable, c is a constant; Θ att Represents the trainable parameters of the neural network, i.e. ψ,w x ,b x ,w g and b g ;q att The attention coefficient α is obtained by normalizing the sigmoid function. The attention mechanism is aimed at the feature layer input x input The attention coefficient is g, further weighting the feature layer x input ; 3) Replace the deconvolution layer used for upsampling with pixel reorganization PixelShuffle; Step 3: Accuracy verification; Step 4: downscaling precipitation; Step 5: Precipitation-runoff simulation.

2. The method for fine-tuning precipitation and runoff simulation in a mountainous watershed according to claim 1, characterized in that: In step 1, the half-hourly IMERG 0.1° precipitation dataset from recent years, along with relevant topographic meteorological elements and ground-based rain gauge observation data, was collected and collated. The satellite precipitation data and topography were resampled to 0.5° to establish a training dataset.

3. The method for fine-tuning precipitation and runoff simulation in a mountainous watershed according to claim 2, characterized in that: In step 2, the training process is as follows: The dataset was used to train a 5x precipitation downscaling model until the error converged. The model was saved and a 0.02° downscaling prediction was performed. The PyTorch deep learning framework was used with a training batch size of 64 and epochs of 50. The ADAM optimizer was used with a learning rate of 1e-4. The Exponential LR learning rate adjustment strategy was set, and the gamma value was set to 0.

95.

4. The method for fine-tuning precipitation and runoff simulation in a mountainous watershed according to claim 3, characterized in that: In step 4, specifically: use the satellite precipitation data at 0.1° as the model input and predict the precipitation downscaled data at 0.02°.

5. The method for fine-tuning precipitation and runoff simulation in a mountainous watershed according to claim 4, characterized in that: In step 5, specifically: couple the precipitation downscaling model with the hydrodynamic model and perform runoff forecast simulation on a 10m Cartesian coordinate grid.

6. The method for fine-tuning precipitation and runoff simulation in a mountainous watershed according to claim 5, characterized in that: The hydrodynamic model is based on the two-dimensional depth-averaged shallow water equations, assuming that the pressure distribution of the water body is hydrostatic pressure, the channel slope is minimal, the fluid density is uniform, and there is a free surface and a fixed bottom boundary. In matrix form, the shallow water equations SWEs are expressed as: Where t represents time; x and y are Cartesian coordinates; U is the vector of conservative physical variables; F and G are the flux vectors along the x and y directions respectively; the source vectors include precipitation R, riverbed slope S b Friction slope S with riverbed f ; These vectors are represented as follows: Where h represents the water depth; u and v are the depth-averaged velocity components in the x and y directions, respectively; I represents the precipitation intensity; g is the acceleration of gravity; z b is the slope height; S fx and S fy is the friction stress estimated by Manning's formula, which can be expressed as n b is the Manning friction coefficient, set to 0.

01.

7. The method for fine-tuning precipitation and runoff simulation in a mountainous watershed according to claim 6, characterized in that: The numerical simulation method of hydrodynamics uses the Godunov finite volume method to solve the shallow water equations. The Godunov finite volume method can handle the exact or approximate Riemann problem at the boundary of the computational unit. The HLLC approximation method is selected to handle the Riemann problem at the unit interface. The finite volume format derived from the shallow water equations is as follows: That is, the differential form of the time step after numerical discretization is: Where, the superscript n represents the time step; the subscript i represents the grid unit number, and f E and f W corresponds to the flux to the east and west respectively, and g N and g S represent the flux to the north and south respectively, Δt is the time step selected by the CFL condition, and its expression is Δt=cflmin(Δt x ,Δt y ),in The iterative process is performed uniformly over the overall time T, and the precipitation intensity R(I,t) is synchronized over the time scale based on the downscaled precipitation of each consecutive half-hour.

8. A refined simulation system for precipitation and runoff in mountainous watersheds, characterized by: It adopts a mountain basin precipitation-runoff refinement simulation method as described in any one of claims 1-7.

Images