A daily precipitation downscaling method and system based on a multi-task learning strategy improved residual network

Abstract

The application discloses a daily precipitation downscaling method and system based on a multi-task learning strategy improved residual network, which constructs input characteristics by using global climate model output and terrain information, inputs input data interpolated to a target spatial scale into a deep convolution downscaling model based on a multi-task learning strategy and a residual structure, jointly learns two types of tasks of precipitation occurrence discrimination and precipitation estimation, and outputs Gamma distribution parameters for representing precipitation occurrence probability and precipitation statistical characteristics. Based on the output, a high-resolution daily precipitation simulation result after downscaling is obtained. The application divides a precipitation process into two subtasks of precipitation occurrence discrimination and precipitation estimation by using a multi-task learning strategy for joint learning, effectively reduces the difficulty of directly simulating a daily precipitation process by a single model, and effectively improves the accuracy of daily precipitation downscaling.

Description

Technical Field

[0001] This invention relates to the field of daily precipitation downscaling, and in particular to a method and system for daily precipitation downscaling based on improving residual networks using a multi-task learning strategy. Background Technology

[0002] Global climate models (GCMs) are crucial techniques for simulating the evolution of the climate system and assessing the impacts of climate change. However, limited by computational capabilities and model structure, GCMs typically have coarse spatial resolutions, making them difficult to directly match with high spatial resolution hydrological models or regional-scale applications. To bridge the gap in spatial resolution between global climate models and regional hydrological simulations, downscaling the precipitation outputs of GCMs has become an effective and necessary approach.

[0003] In existing downscaling studies, daily precipitation processes present significant challenges for downscaling modeling due to their marked nonlinearity, discontinuity, and strong randomness. Influenced by complex topography, landforms, and underlying surface conditions, precipitation processes exhibit significant spatial heterogeneity and scale dependence, further exacerbating the uncertainty of daily precipitation downscaling. Existing downscaling methods suffer from insufficient accuracy in characterizing the spatial distribution of daily precipitation, making it difficult to meet the demands of high-precision hydrological simulations and related applications for precipitation data. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method and system for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy, thereby solving the problem of limited simulation accuracy in existing downscaling methods.

[0005] Technical solution: The present invention provides a method for downscaling daily precipitation based on a multi-task learning strategy to improve residual networks, comprising the following steps:

[0006] Acquire coarse-resolution daily meteorological data and DEM from GCMs and interpolate them to the target high-resolution grid. The interpolation results include first-resolution daily meteorological data and DEM data at the same resolution.

[0007] The interpolation results are input into a downscaling model based on a multi-task learning strategy for residual structures to obtain outputs of Gamma distribution parameters for determining whether precipitation has occurred and estimating precipitation amount. The outputs include precipitation probability, scale parameters of precipitation Gamma distribution, and shape parameters of precipitation Gamma distribution. The downscaling model performs joint learning for the two tasks of precipitation occurrence determination and precipitation estimation.

[0008] High-resolution downscaled daily precipitation results are obtained based on the parameters of the estimated Gamma distribution of precipitation and the precipitation probability.

[0009] Furthermore, the coarse-resolution daily meteorological data of the GCMs includes precipitation occurrence discrimination results, coarse-resolution precipitation, coarse-resolution temperature, coarse-resolution geopotential field, and coarse-resolution humidity data.

[0010] Furthermore, after interpolating the coarse-resolution daily precipitation data to the target high-resolution grid, it is standardized to eliminate dimensions.

[0011] Furthermore, the descaling model of the residual structure based on the multi-task learning strategy includes a residual module, a convolution module, and an output head;

[0012] The residual module includes several convolutional layers connected in sequence. The interpolation result is processed by the residual module to extract features to obtain a first feature vector. The interpolation result is fused with the first feature vector and then input into the convolutional module. The predicted precipitation probability, the scale parameter of the precipitation Gamma distribution, and the shape parameter of the precipitation Gamma distribution are output through three independent output heads.

[0013] Furthermore, the downscaling model performs joint learning on two tasks: precipitation occurrence discrimination and precipitation amount estimation, through a loss function. The loss function is a weighted sum of precipitation occurrence loss and precipitation amount loss. The precipitation occurrence loss is a binary cross-entropy loss function used for precipitation occurrence discrimination, and the precipitation amount loss is a negative log-likelihood loss function based on the Gamma distribution used for precipitation amount estimation.

[0014] Furthermore, the loss function is:

[0015] ;

[0016] ;

[0017] ;

[0018] in, Losses due to precipitation, For the loss of precipitation, This indicates whether precipitation has occurred; no rain indicates no precipitation. Rain is , To observe daily precipitation, It is the Gamma function; This represents the precipitation probability output by the downscaling model. The scaling parameters are the output of the downscaling model. These are the shape parameters output by the downscaling model.

[0019] Furthermore, the downscaled high-resolution daily precipitation results .

[0020] The present invention discloses a diurnal precipitation downscaling system based on an improved residual network, comprising:

[0021] Interpolation unit is used to acquire coarse-resolution daily meteorological data and DEM of GCMs and interpolate them to the target high-resolution grid. The interpolation results include first high-resolution daily meteorological data and DEM data of the same resolution.

[0022] The prediction unit is used to input the interpolation result into a downscaling model with a residual structure improved based on a multi-task learning strategy, and to obtain the output of the Gamma distribution parameters for determining whether precipitation has occurred and estimating the amount of precipitation; the output includes the precipitation probability, the scale parameter of the precipitation Gamma distribution, and the shape parameter of the precipitation Gamma distribution; the downscaling model performs joint learning on the two tasks of precipitation occurrence determination and precipitation estimation.

[0023] Downscaling output unit, used to obtain downscaled high-resolution daily precipitation results based on the parameters of the estimated precipitation Gamma distribution and the precipitation probability.

[0024] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the method for downscaling daily precipitation based on a multi-task learning strategy to improve residual networks.

[0025] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the method for downscaling daily precipitation based on the improved residual network according to the multi-task learning strategy.

[0026] Beneficial Effects: Compared with existing technologies, the advantages of this invention are as follows: Addressing the inherent characteristics of daily precipitation processes—strong nonlinearity, discontinuity, and significant probability distribution skewness—especially in complex terrain regions with large topographic relief and complex underlying surface conditions, traditional downscaling methods struggle to accurately characterize precipitation features. This invention introduces a multi-task learning mechanism to decouple precipitation occurrence discrimination from precipitation estimation, and integrates a loss function constrained by Gamma distribution, large-scale climate model output information, and topographic information to construct a high-precision daily precipitation downscaling method. This method not only reduces the overall learning difficulty of the model for complex precipitation processes and improves its ability to characterize the spatiotemporal variations of precipitation, but also enhances the model's applicability and stability in complex terrain regions, thereby effectively improving the accuracy and reliability of daily precipitation downscaling simulations. Specifically:

[0027] (1) This invention divides the precipitation process into two sub-tasks: precipitation occurrence discrimination and precipitation estimation, and uses a multi-task learning approach for joint modeling. This effectively reduces the difficulty of directly simulating daily precipitation processes with a single model, improves the model's ability to learn nonlinear and discontinuous precipitation features, and thus significantly enhances the overall accuracy and stability of daily precipitation downscaling simulation.

[0028] (2) The present invention introduces the Gamma distribution and improves the loss function, so that the model can more reasonably characterize the skewness and asymmetry of the daily precipitation distribution, avoid the problem that the traditional loss function is insufficient to characterize precipitation, and further improve the physical rationality and statistical consistency of the precipitation estimation results.

[0029] (3) This invention uses the precipitation occurrence discrimination results (rainy / no rain) of global climate models (GCMs), the coarse resolution precipitation, temperature, geopotential field and humidity output of GCMs, and digital elevation model (DEM) as network inputs. Among them, the GCMs output is used to characterize the large-scale atmospheric circulation background conditions, the DEM is used to introduce the influence of topography on precipitation formation and spatial distribution, and the precipitation occurrence discrimination results of GCMs are used to reduce the learning complexity of subsequent precipitation discrimination models. Thus, based on the comprehensive consideration of atmospheric circulation conditions and complex topographic factors, the accuracy of daily precipitation downscaling is effectively improved, especially the accuracy of daily precipitation downscaling in complex topographic areas. Attached Figure Description

[0030] Figure 1 This is a flowchart of the daily precipitation downscaling method according to an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the downscaling model network structure according to an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram of daily precipitation downscaled according to an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram illustrating the evaluation model effect of an embodiment of the present invention. Detailed Implementation

[0034] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, the method for downscaling daily precipitation includes the following steps.

[0036] Step 1: Obtain coarse-resolution daily meteorological data and DEM from GCMs and interpolate them to the target high-resolution grid. The interpolation results include the first high-resolution daily meteorological data and DEM data of the same resolution.

[0037] Specifically, in this embodiment, the coarse-resolution daily meteorological data of GCMs includes the precipitation occurrence discrimination results (rainy / no rain) of the global climate model, the coarse-resolution precipitation, temperature, geopotential field and humidity data output by GCMs, and the coarse-resolution daily meteorological data and digital elevation model (DEM) are used as input factors of the network for high spatial resolution daily precipitation simulation.

[0038] After determining the network input factors, the corresponding data is obtained and preprocessed according to the selected input factors, including: First, unifying the spatial resolution of each input data, and using spatial interpolation methods to interpolate the coarse resolution data of GCMs to the target high resolution spatial scale to obtain the interpolation result.

[0039] Furthermore, this embodiment also standardizes the interpolation results to eliminate dimensional differences and improve the stability of model training.

[0040] Step 2: Input the interpolation result into the downscaling model of the residual structure based on the multi-task learning strategy to obtain the output of the Gamma distribution parameters used to determine whether precipitation has occurred and to estimate the amount of precipitation; the output includes the precipitation probability, the scale parameter of the precipitation Gamma distribution, and the shape parameter of the precipitation Gamma distribution; the downscaling model performs joint learning on the two tasks of precipitation occurrence determination and precipitation estimation.

[0041] Specifically, in this embodiment, the downscaling model employs a multi-task learning strategy, decomposing the daily precipitation downscaling task into two sub-tasks: precipitation occurrence discrimination (rain / no rain) at a fine-resolution grid scale and precipitation estimation, which are then jointly learned. The downscaling model uses a convolutional neural network as its basic architecture, interpolation results as model input, and Gamma parameter estimation as model output, constructing a deep convolutional residual network downscaling model.

[0042] like Figure 2 As shown, the downscaling model includes a residual module with seven convolutional layers, whose kernel numbers are set to 16, 32, 32, 64, 64, 32, and 16 respectively. After feature extraction, deep features are fused with the initial input layer features through skip connections to enhance information transfer and improve the stability of the model training process. The fused features are further modeled and abstracted through two convolutional layers (16, 16). Three independent convolutional output heads are set at the end of the network to predict the probability of precipitation occurrence. The scale parameter of the Gamma distribution and shape parameters It is used to characterize the statistical distribution characteristics of precipitation.

[0043] The loss function of the downscaling model consists of two parts: first, a binary cross-entropy (BCE) loss function used to determine precipitation occurrence; and second, a negative log-likelihood loss function based on the Gamma distribution used for precipitation estimation, the expression of which is as follows:

[0044] ;

[0045] ;

[0046] ;

[0047] in This indicates whether actual precipitation occurred (0 for no rain, 1 for rain). Output the probability of precipitation for the model; and These represent the scale and shape parameters of the Gamma distribution output by the model, respectively. Represents the Gamma function; This indicates the observed daily precipitation.

[0048] When training the downscaling model, high-resolution daily precipitation data obtained from observations are used as observation samples for training the model's gamma loss function. Daily precipitation (occurrence discrimination and precipitation amount), temperature, humidity, geopotential field, and DEM output by GCMs are used as model input samples to construct a real sample set. The sample set is divided into a training set, a validation set, and a test set. The training set is used for model parameter fitting, the validation set is used to evaluate the model's simulation performance and select hyperparameters, and the test set is used to evaluate the final model's fitting effect. Evaluation metrics such as root mean square error and mean absolute error are used, and the model parameters and hyperparameters are optimized using the Adam optimization algorithm and an early stopping strategy.

[0049] Based on the trained downscaling model, inputting daily precipitation (occurrence discrimination and precipitation amount), temperature, humidity, geopotential field, and DEM data from GCMs, yields predicted values. , and .

[0050] Step 3: Obtain downscaled high-resolution daily precipitation results based on the estimated precipitation probability and the parameters of the Gamma distribution of precipitation amount.

[0051] Specifically, the downscaled high-resolution daily precipitation results .

[0052] The method described in this paper is verified through specific experiments using the daily precipitation in the upper reaches of the Yalong River and the Dadu River as examples.

[0053] (1) Measured data

[0054] Daily precipitation data are sourced from the daily surface precipitation values ​​for China published by the National Meteorological Information Center. Gridded dataset (V2.0). Factors used to establish relationships with precipitation simulations include humidity, precipitation (rain / no rain discrimination and precipitation amount), temperature, 500 hPa geopotential height, etc. Data are from the output product of the Canadian Earth System Model Version 5 (CanESM5) global climate model. All data spans from January 1961 to December 2016. The 1km digital elevation model data is sourced from the Resource and Environmental Science Data Center, Chinese Academy of Sciences.

[0055] (2) Input factor selection

[0056] The influencing factors selected for the model are elevation data, humidity output by GCMs, precipitation (rain / no rain discrimination and precipitation amount), temperature, and 500 hPa geopotential height data. Daily precipitation at high resolution is used as an observation sample to participate in the calculation of the gamma likelihood function.

[0057] (3) Data preprocessing:

[0058] The GCM and DEM data from step (1) above are scaled back using bilinear interpolation. Gridded datasets. For example... Figure 3 Showing from Grid interpolation A diagram illustrating the downscaling of gridded data.

[0059] (4) Model training

[0060] High-resolution daily precipitation data obtained from observations were used as the model observation samples. Daily precipitation (occurrence discrimination and precipitation amount), temperature, humidity, geopotential field and DEM output by GCMs were used as the model input samples to construct an actual sample set. The model parameters were optimized and the learning rate was set to 1e-3. The sample set was divided into training set, validation set and test set. The period from 1961 to 1992 was the model training period, the period from 1993 to 2000 was the validation period and the period from 2001 to 2014 was the model testing period.

[0061] (5) Model application and evaluation

[0062] The root mean square error and absolute error were used to evaluate the model performance. Specific results are shown in Table 1. Figure 4 As shown.

[0063] As can be seen from the above embodiments, the present invention utilizes a multi-task learning strategy and a convolutional residual network to perform high-precision prediction of interpolated high-resolution precipitation data, thereby achieving accurate simulation of daily precipitation.

[0064] Table 1. The effect of the constructed model on the downscaling of daily precipitation in the upper reaches of the Yalong River and Dadu River.

[0065]

[0066] The present invention discloses a diurnal precipitation downscaling system based on an improved residual network, comprising:

[0067] Interpolation unit is used to acquire coarse-resolution daily meteorological data and DEM of GCMs and interpolate them to the target high-resolution grid. The interpolation results include first high-resolution daily meteorological data and DEM data of the same resolution.

[0068] The prediction unit is used to input the interpolation result into a downscaling model based on a multi-task learning strategy for the residual structure, and to obtain the output of the Gamma distribution parameters for determining whether precipitation has occurred and estimating the amount of precipitation. The output includes the precipitation probability, the scale parameter of the precipitation Gamma distribution, and the shape parameter of the precipitation Gamma distribution. The downscaling model performs joint learning on the two tasks of precipitation occurrence determination and precipitation estimation.

[0069] Downscaling output unit, used to obtain downscaled high-resolution daily precipitation results based on the parameters of the estimated precipitation Gamma distribution and the precipitation probability.

[0070] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the method for downscaling daily precipitation based on a multi-task learning strategy to improve residual networks.

[0071] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the method for downscaling daily precipitation based on the improved residual network according to the multi-task learning strategy.

[0072] The computer-readable storage medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, flash memory, or any other media that can be used to store program code in the form of instructions or data structures and is accessible by a computer.

[0073] The processor is used to execute a computer program stored in memory to implement the various steps in the methods described in the above embodiments.

Claims

1. A method for downscaling daily precipitation based on a multi-task learning strategy to improve residual networks, characterized in that, Includes the following steps: Acquire coarse-resolution daily meteorological data and DEM from GCMs and interpolate them to the target high-resolution grid. The interpolation results include first high-resolution daily meteorological data and DEM data of the same resolution. The interpolation results are input into a downscaling model based on a multi-task learning strategy for residual structures to obtain outputs of Gamma distribution parameters for determining whether precipitation has occurred and estimating precipitation amount. The outputs include precipitation probability, scale parameters of precipitation Gamma distribution, and shape parameters of precipitation Gamma distribution. The downscaling model performs joint learning for the two tasks of precipitation occurrence determination and precipitation estimation. High-resolution downscaled daily precipitation results are obtained based on the parameters of the estimated Gamma distribution of precipitation and the precipitation probability.

2. The method for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy as described in claim 1, characterized in that, The coarse-resolution daily meteorological data of the GCMs includes precipitation occurrence discrimination results, coarse-resolution precipitation, coarse-resolution temperature, coarse-resolution geopotential field, and coarse-resolution humidity data.

3. The method for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy as described in claim 1, characterized in that, After interpolating coarse-resolution daily precipitation data to the target high-resolution grid, it is standardized to eliminate dimensions.

4. The method for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy according to claim 1, characterized in that, The descaling model of the residual structure based on the multi-task learning strategy includes a residual module, a convolution module, and an output head; The residual module includes several convolutional layers connected in sequence. The interpolation result is processed by the residual module to extract features to obtain a first feature vector. The interpolation result is fused with the first feature vector and then input into the convolutional module. The predicted precipitation probability, the scale parameter of the precipitation Gamma distribution, and the shape parameter of the precipitation Gamma distribution are output through three independent output heads.

5. The method for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy according to claim 1, characterized in that, The downscaling model performs joint learning on two tasks: precipitation occurrence discrimination and precipitation estimation, using a loss function. The loss function is a weighted sum of precipitation occurrence loss and precipitation loss. The precipitation occurrence loss is a binary cross-entropy loss function used for precipitation occurrence discrimination, and the precipitation loss is a negative log-likelihood loss function based on the Gamma distribution used for precipitation estimation.

6. The method for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy according to claim 5, characterized in that, The loss function is: ； ； ； in, Losses due to precipitation, For the loss of precipitation, This indicates whether precipitation has occurred; no rain indicates no precipitation. Rain is , To observe daily precipitation, It is the Gamma function; This represents the precipitation probability output by the downscaling model. The scaling parameters are the output of the downscaling model. These are the shape parameters output by the downscaling model.

7. The method for downscaling daily precipitation based on an improved residual network using a multi-task learning strategy according to claim 6, characterized in that, The downscaled high-resolution daily precipitation results .

8. A diurnal precipitation downscaling system based on a multi-task learning strategy to improve residual networks, characterized in that, include: Interpolation unit is used to acquire coarse-resolution daily meteorological data and DEM of GCMs and interpolate them to the target high-resolution grid. The interpolation results include first high-resolution daily meteorological data and DEM data of the same resolution. The prediction unit is used to input the interpolation result into a downscaling model with a residual structure improved based on a multi-task learning strategy, and to obtain the output of the Gamma distribution parameters for determining whether precipitation has occurred and estimating the amount of precipitation; the output includes the precipitation probability, the scale parameter of the precipitation Gamma distribution, and the shape parameter of the precipitation Gamma distribution; the downscaling model performs joint learning on the two tasks of precipitation occurrence determination and precipitation estimation. Downscaling output unit, used to obtain downscaled high-resolution daily precipitation results based on the parameters of the estimated precipitation Gamma distribution and the precipitation probability.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for downscaling daily precipitation based on a multi-task learning strategy to improve residual networks, as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for downscaling daily precipitation based on a multi-task learning strategy to improve residual networks, as described in any one of claims 1-7.

Images