A cloud classification quantitative precipitation estimation method based on a ResGAN network

Abstract

The application discloses a cloud classification quantitative precipitation estimation method based on a ResGAN network, relates to the technical field of quantitative precipitation estimation, and comprises the following steps: acquiring multi-source polarized radar parameters; pre-processing the multi-source polarized radar parameters to obtain pre-processed data; and inputting the pre-processed data into a cloud classification quantitative precipitation estimation model, wherein the cloud classification quantitative precipitation estimation model comprises a space-time feature extraction module, a cloud classification physical constraint module, a motion-intensity coupling module and an improved ResGAN generator. The cloud classification quantitative precipitation estimation method based on the ResGAN network considers the space-time heterogeneity of actual precipitation, can extract transient motion features of a precipitation system, takes cloud type probability as a physical constraint to prevent systematic deviation, and can effectively capture the coupling effect of motion-intensity.

Description

Technical Field

[0001] The present invention relates to the technical field of quantitative precipitation estimation, and in particular to a cloud classification quantitative precipitation estimation method based on a ResGAN network. Background Art

[0002] Weather radar is a crucial tool for monitoring and forecasting precipitation, providing high-temporal and spatial resolution precipitation information. It has widespread applications in meteorology, hydrology, and disaster warning. Quantitative precipitation estimation is a core task in meteorology and hydrology, and its accuracy directly impacts the reliability of flood warnings, water resources management, and disaster prevention and control. Traditional quantitative precipitation estimation methods are primarily based on empirical relationships between radar reflectivity (Z) and precipitation intensity (R), such as the ZR relationship. However, these methods suffer from three bottlenecks: 1. Physical simplification: the ZR relationship assumes a uniform distribution of precipitation particles, ignoring the spatiotemporal heterogeneity of actual precipitation (especially convective precipitation); 2. Cloud classification is inadequate, failing to distinguish between the physical processes of stratiform clouds (large-scale, stable precipitation) and convective clouds (localized, heavy precipitation), leading to systematic biases; 3. Dynamic feature decoupling: the radar reflectivity intensity (static feature) is separated from the motion of the precipitation system (such as speed and direction), failing to capture the coupled effects of motion and intensity (such as the short-duration, heavy precipitation of fast-moving convective clouds).

[0003] Therefore, a cloud classification and quantitative precipitation estimation method based on ResGAN network was developed to solve the above problems. Summary of the Invention

[0004] This paper proposes a cloud classification quantitative precipitation estimation method based on the ResGAN network to solve the problems that existing quantitative precipitation estimation methods ignore the temporal and spatial heterogeneity of actual precipitation, have systematic biases, and cannot capture the motion-intensity coupling effect.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] The present invention provides a cloud classification quantitative precipitation estimation method based on a ResGAN network, comprising:

[0007] Acquire multi-source polarimetric radar parameters, including the density and shape of precipitation particles in the precipitation system, path-accumulated precipitation, and radar beam observation height;

[0008] Preprocessing the multi-source polarization radar parameters to obtain preprocessed data;

[0009] The preprocessed data is input into a cloud classification quantitative precipitation estimation model, which includes a spatiotemporal feature extraction module, a cloud classification physical constraint module, a motion-intensity coupling module, and an improved ResGAN generator. The spatiotemporal feature extraction module is used to sequentially capture the dynamic evolution information of the precipitation system and compensate for long-distance signal attenuation based on the radar beam observation height on the preprocessed data to obtain a first output; the cloud classification physical constraint module is used to sequentially calculate the cloud type probability based on meteorological prior knowledge on the first output and generate a regional sensitive mask matching the cloud type to obtain a second output; the motion-intensity coupling module is used to fuse the second output with the spatiotemporal features and motion weights of the precipitation system to obtain a third output; and the improved ResGAN generator is used to process the third output and output the predicted precipitation intensity.

[0010] Furthermore, the multi-source polarization radar parameters are preprocessed, including:

[0011] performing data standardization and outlier trimming on the multi-source polarimetric radar parameters;

[0012] performing spatiotemporal alignment and sample matching on the multi-source polarimetric radar parameters;

[0013] Multi-elevation-angle fusion and radar beam observation height are introduced into the multi-source polarization radar parameters.

[0014] Furthermore, the spatiotemporal feature extraction module is used to process the preprocessed data in sequence through a 3D convolution layer, a cascaded star operation module, and a height attenuation compensation module.

[0015] Furthermore, the cascaded star operation module includes three Star Block modules connected in series.

[0016] Furthermore, the calculation formula of the height attenuation compensation module is:

[0017] ;

[0018] in, represents the first output, Represents the output of the cascaded star operation module, is the radar beam observation height, x, y represent two-dimensional data, x is the azimuth of the radar, and y is the radial distance library.

[0019] Furthermore, the cloud classification physical constraint module is used to sequentially perform global average pooling calculation, cloud type probability calculation, and spatially adaptive weighted mask calculation on the first output, wherein:

[0020] The calculation formula for average pooling is as follows:

[0021]

[0022] Among them, H×W represents the spatial dimension, Represents the value of the cth feature map at position (i, j), where i is the index from 1 to H along the height direction and j is the index from 1 to W along the width direction;

[0023] The cloud type probability calculation formula is as follows:

[0024] ;

[0025] in, represents the cloud type probability, represents the weight matrix of the fully connected layer of the cloud classifier, Represents the output obtained by inputting the global average pooling calculation result into the fully connected layer;

[0026] The spatially adaptive weighted mask calculation includes:

[0027] ;

[0028] in, For mask.

[0029] Furthermore, the third output is obtained by fusing the second output with the spatiotemporal characteristics and motion weights of the precipitation system, including:

[0030] Calculate the motion vector using the following formula:

[0031]

[0032] in, Represents the first output at the current moment, represents the first output at the next moment, represents the motion vector;

[0033] Calculate channel attention, the calculation formula is:

[0034] ;

[0035] ;

[0036] ;

[0037] in express, Represents the feature map output by the StarNet module, where c represents the cth channel, (i, j) represents the spatial position (row i, column j), and s represents the channel weight. represents the activation function, represents the weight matrix of the fully connected layer, a weight matrix representing a fully connected layer, a first output, spatio-temporal features;

[0038] a motion weight calculated according to the formula:

[0039]

[0040] wherein a motion weight, a convolutional layer;

[0041] feature enhancement:

[0042]

[0043] an enhancement coefficient, a feature enhancement result;

[0044] feature fusion:

[0045]

[0046] a feature fusion result.

[0047] Further, the improved ResGAN generator comprises three residual blocks, a multi-scale convolutional layer, a feature concatenation layer, a 1X1 convolutional layer, two layers of transpose convolutional layers, and an output layer arranged in sequence.

[0048] Further, the loss function of the cloud classification and quantitative precipitation estimation model is represented as:

[0049]

[0050] wherein a Huber loss, N represents the number of precipitation samples, i' represents the i'th precipitation sample, a real precipitation intensity, a predicted precipitation intensity output by the cloud classification and quantitative precipitation estimation model.

[0051] Further, the cloud classification and quantitative precipitation estimation model is trained based on a three-stage progressive training paradigm, comprising:

[0052] In the first stage, the spatio-temporal feature extraction module, the cloud classification physical constraint module, and the motion-intensity coupling module are trained;

[0053] In the second stage, the improved ResGAN generator is fine-tuned;

[0054] ​​​​In the third stage, the ResGAN generator is optimized through adversarial training.

[0055] The beneficial effects of the present invention are:

[0056] A cloud classification and quantitative precipitation estimation method based on the ResGAN network proposed in this paper takes into account the spatiotemporal heterogeneity of actual precipitation, can extract the transient motion characteristics of the precipitation system, uses cloud type probability as a physical constraint, prevents systematic bias, and can effectively capture the motion-intensity coupling effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a structural principle diagram of the spatiotemporal feature extraction module in an embodiment of the present application.

[0058] Figure 2 This is a structural principle diagram of the cloud classification physical constraint module in an embodiment of the present application.

[0059] Figure 3 This is a structural principle diagram of the motion-intensity coupling module in an embodiment of the present application.

[0060] Figure 4 This is a structural principle diagram of the improved ResGAN generator in the embodiment of this application.

[0061] Figure 5 This is a schematic diagram of the structural principle of the residual block in the improved ResGAN generator of the embodiment of the present application.

[0062] Figure 6 Flowchart of the three-stage progressive training paradigm in the embodiment of this application. DETAILED DESCRIPTION

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0064] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0065] It should be noted that like reference numerals and characters refer to like items throughout the drawings and that once an item is defined in one drawing, it is not necessary to further define and explain it in the subsequent drawings.

[0066] The "Star Block module" is a module in the StarNet neural network architecture for optimizing the computational efficiency and expressiveness of sequence modeling or graph modeling through a star structure. The "DROPS2.0" is an algorithm system specially used for weather radar data quality control and classification, and the full name is: Dual-polarization Radar Object-based Precipitation Segmentation (Version 2.0). The specific embodiments of the present application are described in detail below in conjunction with the drawings.

[0067] In order to reduce the influence of non-precipitation echoes in polarimetric radar data on precipitation estimation, the DROPS2.0 data quality control algorithm is used for filtering to remove non-precipitation echoes.

[0068] In order to ensure the accuracy, consistency and physical reasonableness of the input data of the cloud classification and quantitative precipitation estimation model, the present application carries out strict preprocessing and standardization operation on the weather radar observation data and ground rain gauge data before training the cloud classification and quantitative precipitation estimation model, especially the processing strategy of radar beam observation height and multi-channel radar parameters, which specifically includes the following preprocessing and standardization operation.

[0069] 1. Multi-source polarimetric radar parameter fusion input.

[0070] The input of the cloud classification and quantitative precipitation estimation model of the present application is composed of a plurality of polarimetric radar variables, including: Z: unit dBZ, reflecting the density of precipitation particles; Zdr: unit dB, reflecting the shape of precipitation particles; Kdp: unit ° / km, reflecting the path cumulative precipitation; radar beam observation height H: unit m, reflecting the vertical height from the ground.

[0071] The above four variables jointly constitute four-channel input data and are used in the subsequent three-dimensional residual coding and generation network, fully utilizing the vertical and horizontal distribution recognition ability of radar on precipitation system structure.

[0072] 2. Data standardization and outlier clipping.

[0073] Considering the large span of radar observation value range, linear normalization is used for uniform processing of variables to improve training stability.

[0074] The recommended ranges of the variables are as shown in Table 1:

[0075] Table 1

[0076]

[0077] Data beyond the interval will be clipped to avoid extreme outliers. For the non-echo region (no radar data coverage), a uniform assignment of -1 is made so that the cloud classification quantitative precipitation estimation model can distinguish between "no echo" and "very weak precipitation" regions, improving the boundary discrimination ability.

[0078] 3. Spatio-temporal alignment and sample matching.

[0079] Due to the inconsistency in time scale between radar data (5-7 min resolution) and rain gauge data (1 min resolution), the invention adopts the following methods for alignment: temporal alignment: for each frame of radar observation window, the average cumulative value of rain gauge data in the corresponding time range is taken as the "ground truth" label of the frame; spatial alignment: taking the geographic coordinates of the rain gauge as the center, a 9x9 pixel window is cut out on the radar image as the observation patch matched with the rain gauge. This "local radar window + center rain gauge live" pairing strategy makes the training data consider both spatial details and live accuracy, enhancing the model generalization.

[0080] 4. Multi-elevation angle fusion and height prior introduction

[0081] The invention introduces the echo data corresponding to multiple radar elevations (e.g. 0.5°, 0.9°, 1.3°, 1.8°) and their vertical height information, forming four layers of spatio-temporal grid feature maps:

[0082] ;

[0083] C: number of channels (Z, Zdr, Kdp, H);

[0084] T: continuous scanning time (representing different elevations);

[0085] HxW: spatial dimension (9x9 or larger range);

[0086] This design not only improves the model's ability to perceive vertical structure changes, but also provides physical constraints for the attenuation compensation of distant area precipitation.

[0087] The composition of the cloud classification quantitative precipitation estimation model is described in detail as follows:

[0088] The core goal of the spatiotemporal feature extraction module is to address shortcomings in spatiotemporal feature extraction: traditional 2D convolutions struggle to capture the dynamic evolution of precipitation systems (such as storm movement and intensity changes); and to mitigate radar echo attenuation: Atmospheric attenuation distorts the echo intensity of long-range radar signals, necessitating radar beam height compensation. Therefore, 3D convolutions are used to capture the dynamic evolution of precipitation systems, and radar beam height compensation is introduced to compensate for long-range signal attenuation.

[0089] The architecture of the spatiotemporal feature extraction module is as follows Figure 1 As shown,

[0090] Input data structure: 4D tensor (channel × time × distance × orientation)

[0091]

[0092] Z: reflectivity factor (dBZ);

[0093] Zdr: Differential reflectivity provides the radar beam observation height;

[0094] Kdp: differential phase shift rate;

[0095] H: radar beam observation height;

[0096] ;

[0097] Indicates the altitude of the radar detection unit, represents the altitude of the radar station itself, :max_elev is the maximum elevation angle of the radar (for example, 19.5°), and range is the slant range from the current detection unit to the radar.

[0098] Channel dimensions: reflectivity (Z), differential reflectivity (Zdr), relative differential phase (Kdp), height prior (H)

[0099] Time dimension: 4 consecutive scanning moments (about 20 minutes)

[0100] Spatial dimensions: 9×9 grid (approximately 20km×20km)

[0101] 3D convolutional layer:

[0102] ;

[0103] Where c represents the output channel index, c∈[0,63], dt represents the temporal convolution offset, dt∈[0,3], dx, dy represent the spatial convolution offset, (dx,dy)∈[0,2]. t represents time.

[0104] The Star Block calculation formula is as follows:

[0105] ;

[0106] .

[0107] Y represents the calculation result, and Output represents the output.

[0108] Height attenuation compensation:

[0109] ;

[0110] in, represents the first output, Represents the output of the cascaded star operation module, is the radar beam observation altitude.

[0111] in, (3D BatchNorm) standardizes data in the spatiotemporal dimension to accelerate training convergence and reduce internal covariate shift; introduces nonlinear ReLU activation function to enhance model expressiveness; uses 3×3×3 convolution kernel , layered extraction of local spatiotemporal features (such as storm edge movement and intensity changes); adding the input to the second convolution output to avoid gradient vanishing, support deep network training, and preserve original information to prevent feature degradation.

[0112] like Figure 1 As shown, the 4×4×9×9 input first passes through the convolution kernel Then, through a cascaded StarBlock structure, three StarBlocks are connected in series to gradually abstract high-level features: 64×4×7×7 input passes through StarBlock1 to represent basic motion patterns (such as echo movement direction), 64×4×7×7 input passes through StarBlock2 to represent medium-scale evolution (such as storm mergers and splitting), and 128×4×7×7 input passes through StarBlock3 to represent complex system dynamics (such as supercell rotational structure). The output is 256×4×7×7 (channel × time × height × width), highly compressing spatial dimensions and focusing on key spatiotemporal patterns. Finally, the height compensation module outputs the features.

[0113] like Figure 2As shown in the figure, the core goal of the cloud classification physical constraint module is to address the differences in precipitation mechanisms: the physical processes of stratus clouds (large-area stable precipitation) and convective clouds (local heavy precipitation) are different and require differentiated treatment; introduce physical constraints: integrate meteorological prior knowledge into the network structure to improve the physical rationality of precipitation intensity estimation; spatial adaptive weighting: generate regional sensitive masks that match the cloud type to guide precipitation intensity estimation.

[0114] Input: 256×7×7 high-level features output by StarNet ;

[0115] First, it passes through a 128-channel 3X3 convolution, a ReLU module (input size is 128×5×5), a 64-channel 3X3 convolution, and a ReLU module (input size is 64×3×3);

[0116] Then it goes through global average pooling and fully connected layers:

[0117] ;

[0118] Then after Softmax and cloud type probability output:

[0119] ;

[0120] Then the spatial weighted mask is obtained through mask generation:

[0121] ;

[0122] The input data comes from the 256×7×7 feature map of the spatiotemporal feature extraction module. This feature map already contains information about spatiotemporal evolution but does not distinguish between precipitation types and physical mechanisms. A two-branch convolutional processing is used: Branch 1 is a 128-channel convolutional layer that compresses the spatial dimension and maps high-dimensional features (256 channels) into a more physically meaningful 128-dimensional space. Feature abstraction uses 3×3 convolutions to capture local spatial patterns (such as precipitation core areas and edge gradients). Branch 2 is a 64-channel convolutional layer that further refines cloud physical features. The 64-dimensional features correspond to key precipitation process parameters. Channels 0-31 correspond to stratiform cloud characteristics (uniform water vapor distribution), and channels 32-63 correspond to coherent cloud characteristics (strong updrafts and ice phases). Global average pooling captures the overall cloud characteristics of the region, ignoring spatial details. The fully connected layer learns cloud type diagnosis rules (stratus weights: associated with stable stratification parameters, convective cloud weights: associated with unstable energy parameters). Softmax normalization outputs a probability vector [P(stratus), P(convective)]. Mask generation outputs a 7×7 spatially weighted mask (range [0,1]). High-value areas indicate precipitation core areas that need to be strengthened, while low-value areas indicate weak precipitation or noise areas.

[0123] The defects of traditional methods are: they can only treat all precipitation types uniformly, ignore spatial heterogeneity, are purely data-driven and lack physical constraints, and global averaging leads to loss of details.

[0124] The beneficial effects of the cloud classification physical constraint module in the present invention are: dual-path differentiation of stratiform / convective cloud physical processes, generation of a 7×7 fine-grained spatial mask, injection of meteorological prior knowledge into cloud type probabilities, and retention of the convective cloud core area structure in the mask.

[0125] Among them, altitude prior information: according to the radar wave propagation model, the signal attenuation coefficient (such as dB / km) of different distance circles (corresponding to different altitude layers) is calculated; compensation mechanism: weighted enhancement of long-distance (high elevation angle) feature map channels to offset the signal weakening caused by atmospheric attenuation; physical constraints: meteorological prior knowledge is embedded in the network to improve the prediction reliability of the model in long-distance areas.

[0126] like Figure 3 As shown in Figure 2, the core goal of the motion-intensity coupling module is to effectively fuse appearance information (Intensity, usually from RGB frames) and motion information (Motion, usually from optical flow or motion vectors) to generate a richer spatiotemporal feature representation.

[0127] Calculate the motion vector using the following formula:

[0128]

[0129] in, Represents the first output at the current moment, represents the first output at the next moment, represents the motion vector;

[0130] Calculate channel attention, the calculation formula is:

[0131] ;

[0132] ;

[0133] ;

[0134] in express, Represents the feature map output by the StarNet module, where c represents the cth channel, i.e., the output channel index, (i, j) represents the spatial position (row i, column j), and s represents the channel weight. represents the activation function, represents the weight matrix of the fully connected layer, represents the weight matrix of the fully connected layer, represents the first output, Represents spatiotemporal characteristics;

[0135] Calculate the motion weight using the following formula:

[0136] ;

[0137] in represents the motion weight, represents the convolutional layer;

[0138] Feature Enhancement:

[0139] ;

[0140] represents the enhancement coefficient;

[0141] Feature fusion:

[0142] .

[0143] The motion-intensity coupling module solves key problems in precipitation estimation by fusing multi-source features: motion-intensity decoupling, where traditional methods cannot correlate the movement of precipitation systems with intensity changes; lack of physical constraints, where the impact of cloud type on precipitation distribution is ignored; and feature alignment bias, where the scale of spatiotemporal features does not match that of physical features.

[0144] Channel attention mechanism focuses on key meteorological features and suppresses noise; global spatial compression compresses the 2D feature map of each channel into a scalar to capture global information; channel correlation learning, Dimensionality reduction fully connected layer (r=16 compression ratio), The fully connected layer is restored to its original dimension. Features are recalibrated, and the learned channel weights s∈[0,1] are used to enhance important features. Feature enhancement increases the feature values ​​of the moving severe convection zone by 1.8 times, while the feature values ​​of the stationary cloud zone remain unchanged.

[0145] like Figure 4 and Figure 5 As shown in the figure, the improved ResGAN generator includes three residual blocks, a multi-scale convolution layer, a feature splicing layer, a 1x1 convolution layer, two transposed convolution layers, and an output layer. The specific data processing mechanism includes:

[0146] The main path calculation inside the residual block is:

[0147] Indicates that the input feature map is subjected to 3D convolution operation, 3×3×3 convolution, and outputs 256 channels with unchanged size, that is, Figure 5 Conv3D in

[0148] represents batch normalization, i.e. Figure 5 BN3D in;

[0149] Represents the ReLU activation function, that is Figure 5 ReLU in

[0150] Indicates a 3D convolution operation, 3×3×3 convolution, outputting 256 channels with unchanged size;

[0151] Represents batch normalization.

[0152] Jump Path , here the input X feature map (256 channels) is converted to 128 channels through 1x1 convolution, that is, Figure 5 1X1Conv in , followed by batch normalization:

[0153] ;

[0154] Residual output Y':

[0155] ;

[0156] Atrous convolution, three branches, each branch uses atrous convolution, with dilation rates of 1, 2, and 4 respectively. The atrous convolution of each branch: convolution kernel size 3×3 (spatial size), dilation rate d, number of convolution kernel channels: input 256, output 128 (because it will be spliced ​​later, each branch outputs 128 channels, so after splicing it is 384 channels):

[0157] ;

[0158] Represents the result of the dilated convolution, d: dilation rate (1, 2, 4) weight Independent learning, c represents the channel index, k represents the input channel index, u, v represents the convolution kernel spatial position, Represents the value of the input feature map.

[0159] (d): Dilation rate identifier (different d use independent weights).

[0160] Multi-scale feature stitching:

[0161] ;

[0162] Represents the multi-scale feature splicing result, 、 、 They represent dilated convolutions with dilation rates of 1, 2, and 4, respectively.

[0163] Multi-scale perception is shown in Table 2:

[0164] Table 2

[0165]

[0166] Then, a 1x1 convolution is used to reduce the number of channels from 384 to 256, resulting in a 256x7x7 feature map.

[0167] Transposed convolution, first transposed convolution: input 256x7x7, output 128x14x14; second transposed convolution: input 128x14x14, output 64x28x28, same stride 2, kernel size 4x4:

[0168] ;

[0169] denotes the transposed convolution output, denotes the value in the weight tensor corresponding to the input channel c and the spatial offset (u, v), i, j denote the integer coordinates (row index and column index) of a certain spatial position on the output feature map, c: channel index of the input feature map, all input channels (0 to C_in-1) are traversed when summing, u: offset index of the convolution kernel in the height direction (0≤u≤3), v: offset index of the convolution kernel in the width direction (0≤v≤3), denotes the value of the input feature map.

[0170] Output layer, finally a 1x1 convolution is used to change 64 channels to 1 channel, then pass through the Sigmoid activation function, and then multiply by 150 (maximum precipitation intensity):

[0171] ;

[0172] denotes the 1x1 convolution, 150 mm / h: maximum measurable precipitation intensity, actual meaning is to map the network output to the real precipitation value.

[0173] ​The improved ResGAN generator passes through three residual blocks (the internal calculations in each block follow the modified formula above, keeping the size and number of channels unchanged). It then enters a multi-scale convolutional module (three dilated convolution branches with different dilation rates, each outputting 128 channels, concatenated to 384 channels, which are then reduced to 256 channels using a 1×1 convolution). This is followed by two upsampling operations using transposed convolutions (the first output is 128×14×14, the second output is 64×28×28). Finally, a 1×1 convolution is performed to obtain a 1×28×28 output. This is then passed through a sigmoid and multiplied by 150 to obtain the precipitation intensity estimate, converting the fused features into a high-resolution precipitation field. This innovative multi-scale residual structure achieves both spatiotemporal feature refinement and resolution improvement. A deep residual architecture is employed: three layers of residual blocks enhance feature propagation, while skip connections address the vanishing gradient problem. Transposed convolutions replace interpolation for efficient upsampling, and upsampling parameters are learned. A cascaded design balances computational effort and resolution.

[0174] The balanced Huber loss module rebalances the training weights of different types of precipitation samples through a dynamic weight mechanism. Because actual precipitation data exhibits a long-tail distribution, with a high proportion of weak precipitation and a low proportion of moderate and heavy precipitation, the use of traditional loss functions has the following drawbacks: 1. Treating all samples equally causes the model to overemphasize high-frequency weak precipitation; 2. The gradient of heavy precipitation samples is overwhelmed, leading to a systematic underestimation of extreme events. To this end, the present invention proposes the following loss function:

[0175] The final loss function calculation formula is:

[0176] ;

[0177] in, represents the Huber loss, N represents, i' represents the i'th precipitation sample, represents the true precipitation intensity, Represents the predicted precipitation intensity output by the cloud classification quantitative precipitation estimation model.

[0178] The training system architecture of the present invention is as follows:

[0179] Learning rate scheduling:

[0180]

[0181] The final learning rate, Minimum learning rate, Maximum learning rate, t current time step, T total time steps.

[0182] Adam optimizer:

[0183] ;

[0184] ;

[0185] ;

[0186] First-order moment decay rate, The second-order moment decay rate, Smoothness term, First-order moment estimate, Second-order moment estimation, learning rate, Parameter vector.

[0187] Gradient Clipping:

[0188]

[0189] Gradient vector.

[0190] like Figure 6 As shown, the present invention adopts a three-stage progressive training paradigm, combined with physical constraint guidance and adversarial learning optimization, to solve the problems of mode collapse and overfitting in traditional end-to-end training. The training process strictly follows the progressive principle of "basic features first, then generation capabilities, and finally discriminant optimization". After the raw data is pre-processed, the first stage uses StarNet and cloud classification physical constraint module and motion-coupling module for pre-training to establish the basic ability of physical feature extraction; the second stage is to train the ResGAN generator, i.e. Figure 6 Fine-tune the frozen generator in the CNN to optimize the quality of precipitation field generation; in stage 3, optimize the adversarial training to improve the physical rationality of the precipitation field.

[0191] Model Evaluation

[0192] This paper adopts five evaluation indicators to evaluate the model, namely, probability of detection (POD), false alarm rate (FAR), critical success index (CSI), Heidke skill score (HSS) and Gilbert skill score (GSS). The calculation method is as follows:

[0193]

[0194]

[0195]

[0196]

[0197]

[0198] If the radar precipitation estimate and the rain gauge measurement are greater than the threshold, they are considered valid; otherwise, they are considered invalid. S indicates a success (i.e., both the rain gauge measurement and the rainfall estimate are valid); F indicates a failure (i.e., the rain gauge measurement is valid, but the rainfall estimate is invalid); A indicates a false alarm (i.e., the rain gauge measurement is invalid, but the rainfall estimate is valid); and N indicates a true negative (i.e., both the rain gauge measurement and the rainfall estimate are invalid). Higher POD, CSI, HSS, and GSS values, or lower FAR values, indicate better estimation performance.

[0199] In Table 1, with the exception of the FAR parameter, Ours achieved the highest POD, CSI, HSS, and GSS values ​​at 2.5, 5.0, and 7.6 mm. For the 7.6 mm threshold, the Stratiform Z-R relationship had the lowest FAR value. However, it should be noted that CSI offers more comprehensive model evaluation capabilities than POD and FAR. Overall, Ours achieved the best overall performance.

[0200] The qualitative evaluation results of the five RQPE methods under different precipitation intensities are shown in Table 3 below:

[0201] Table 3

[0202]

[0203] Among them, Convective ZR represents the traditional convective cloud ZR relationship inversion, Stratiform ZR represents the traditional stratiform cloud ZR relationship inversion, ResNet101 represents the classic 101-layer residual network, and RQPENetD1 represents the RQPE network variant.

[0204] This paper proposes a quantitative precipitation estimation method for cloud classification based on the ResGAN network, which extracts the transient motion characteristics (speed and direction) of the precipitation system through a motion-intensity coupling module. The cloud classification physical constraint module uses a multi-task framework to connect cloud classification branches in parallel and output cloud type probabilities (stratus clouds / convective clouds) as physical constraints. The motion-intensity coupling module adjusts the spatial distribution of precipitation through adaptive convolution (such as enhancing the estimation of short-term heavy precipitation in fast-moving areas). The method also uses weighted optimization of type-related loss functions (such as assigning higher weights to convective cloud areas to alleviate the underestimation of heavy precipitation). The improved ResGAN generator introduces dilated convolution to capture multi-scale precipitation characteristics (particle scale → system scale). Physical prior enhancement improves the reliability of long-range estimation by compensating for radar wave attenuation through the height information channel in the spatiotemporal feature extraction module.

[0205] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A cloud classification quantitative precipitation estimation method based on ResGAN network, characterized by: include: Acquire multi-source polarimetric radar parameters, including the density and shape of precipitation particles in the precipitation system, path-accumulated precipitation, and radar beam observation height; Preprocessing the multi-source polarization radar parameters to obtain preprocessed data; The preprocessed data is input into the cloud classification quantitative precipitation estimation model, which includes a spatiotemporal feature extraction module, a cloud classification physical constraint module, a motion-intensity coupling module, and an improved ResGAN generator. The spatiotemporal feature extraction module is used to sequentially capture the dynamic evolution information of the precipitation system and compensate for long-distance signal attenuation based on the radar beam observation height on the preprocessed data to obtain the first output. The cloud classification physical constraint module is used to sequentially calculate the cloud type probability based on meteorological prior knowledge on the first output and generate a regional sensitivity mask that matches the cloud type to obtain the second output. The motion-intensity coupling module is used to fuse the second output with the spatiotemporal characteristics and motion weights of the precipitation system to obtain the third output; The improved ResGAN generator is used to process the third output, which outputs the predicted precipitation intensity.

2. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 1 is characterized in that: Preprocessing of multi-source polarimetric radar parameters, including: Perform data standardization and outlier trimming on multi-source polarimetric radar parameters; Perform spatiotemporal alignment and sample matching on multi-source polarimetric radar parameters; Multi-elevation-angle fusion of multi-source polarization radar parameters and introduction of radar beam observation height.

3. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 1 is characterized in that: The spatiotemporal feature extraction module is used to process the preprocessed data through the 3D convolution layer, the cascaded star operation module, and the height attenuation compensation module in sequence.

4. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 3 is characterized in that: The cascaded star operation module includes three Star Block modules connected in series.

5. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 3 or 4, characterized in that: The calculation formula of the height attenuation compensation module is: ； in, represents the first output, Represents the output of the cascaded star operation module, is the radar beam observation height, x, y represent two-dimensional data, x is the azimuth of the radar, and y is the radial distance library.

6. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 1, characterized in that: The cloud classification physical constraint module is used to perform global average pooling calculation, cloud type probability calculation, and spatial adaptive weighted mask calculation on the first output in sequence, where: The calculation formula for average pooling is as follows: ； Among them, H×W represents the spatial dimension, H represents the height, W represents the width, Represents the value of the c-th feature map at position (i, j), where i is the index from 1 to H along the height direction and j is the index from 1 to W along the width direction; The cloud type probability calculation formula is as follows: ； in, represents the cloud type probability, represents the weight matrix of the fully connected layer of the cloud classifier, Represents the output obtained by inputting the global average pooling calculation result into the fully connected layer; Spatially adaptive weighted mask calculation, including: ； in, For mask.

7. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 6, characterized in that: The third output is obtained by fusing the second output with the spatiotemporal characteristics and motion weights of the precipitation system, including: Calculate the motion vector using the following formula: ； in, Represents the first output at the current moment, represents the first output at the next moment, represents the motion vector; Calculate channel attention, the calculation formula is: ； ； ； in Represents the global average pooling result of the c-th channel, that is, the average of all spatial positions of the c-th channel of the feature map, H represents the height, W represents the width, Represents the feature map output by the StarNet module, where c represents the cth channel, (i, j) represents the spatial position row i and column j, and s represents the channel weight. represents the activation function, represents the weight matrix of the fully connected layer, represents the weight matrix of the fully connected layer, represents the first output, Represents spatiotemporal characteristics; Calculate the motion weight using the following formula: ； in represents the motion weight, represents the convolutional layer; Feature Enhancement: ； represents the enhancement coefficient, Represents the feature enhancement result; Feature fusion: ； Represents the feature fusion result.

8. The cloud classification and quantitative precipitation estimation method based on ResGAN network according to claim 1, characterized in that: The improved ResGAN generator consists of three residual blocks, a multi-scale convolutional layer, a feature splicing layer, a 1X1 convolutional layer, two transposed convolutional layers, and an output layer.

9. The cloud classification quantitative precipitation estimation method based on ResGAN network according to claim 1, characterized in that: The loss function of the cloud classification quantitative precipitation estimation model is expressed as: ； in, Represents Huber loss, N represents the number of precipitation samples, i' represents the i'th precipitation sample, represents the true precipitation intensity, Represents the predicted precipitation intensity output by the cloud classification quantitative precipitation estimation model.

10. A cloud classification and quantitative precipitation estimation method based on ResGAN network according to claim 1 or 9, characterized in that: A quantitative precipitation estimation model for cloud classification is trained based on a three-stage progressive training paradigm, including: The first stage trains the spatiotemporal feature extraction module, the cloud classification physical constraint module, and the motion-intensity coupling module; The second stage fine-tunes the improved ResGAN generator; In the third stage, the ResGAN generator is optimized through adversarial training.

Images