Radar short-time heavy rainfall estimation method based on classification echo and environmental physical constraint

Abstract

The invention discloses a radar short-time heavy rainfall estimation method based on classification echoes and environmental physical constraints, and belongs to the technical field of meteorological detection, and the method comprises the following steps: S1, multi-source data integrated fusion and cooperative gridding preprocessing; s2, rainfall type dynamic identification and Z-R relation self-adaptive primary selection based on multi-feature fusion; s3, adaptive correction of the estimation result driven by the environmental physical process is carried out; s4, estimating sequence optimization and systematic deviation correction based on a sliding time window and live feedback; and S5, multi-source information optimal fusion and refined heavy rainfall product generation. According to the method, the problems that a traditional fixed Z-R relation and single data source estimation method is insufficient in precision and insufficient in physical mechanism consideration are effectively solved, the accuracy of short-time heavy rainfall estimation is remarkably improved, and the method has obvious service application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological detection, and particularly relates to a radar short-time heavy rainfall estimation method based on classified echoes and environmental physical constraints. BACKGROUND

[0002] Short-time heavy rainfall is an important weather phenomenon with suddenness, locality and disaster-causing property. Its accurate monitoring and prediction is one of the core challenges in the meteorological field. As a key tool for detecting the three-dimensional structure of rainfall, the precision of radar quantitative rainfall estimation directly affects the input quality of hydro-meteorological models and the timeliness of disaster warning.

[0003] Traditional radar quantitative rainfall estimation generally uses a fixed Z-R relationship, i.e., an empirical relationship between reflectivity factor Z and rainfall rate R. However, the particle size distribution of actual rainfall is significantly affected by rainfall type, thermodynamic environment, etc. For example, the raindrop spectrum characteristics of stratiform cloud precipitation and severe convective precipitation are quite different, and using the same Z-R relationship will introduce systematic bias. In addition, microphysical processes such as evaporation and collision will greatly change the rainfall rate under certain environmental conditions such as low-level dry air intrusion, and the fixed Z-R relationship cannot respond to changes in such physical processes.

[0004] In recent years, radar echo extrapolation models based on deep learning (such as ConvLSTM, U-Net) have shown advantages in short-term prediction, but their essence is data-driven pattern recognition, lacking explicit expression of rainfall microphysical mechanisms, resulting in insufficient generalization ability in complex weather systems or rare individual cases, and poor physical interpretability of estimation results. Existing technologies attempt to introduce environmental parameters for correction, but are mostly limited to single factors or static lookup tables, failing to build a closed-loop optimization system from "echo recognition-physical correction-bias correction-fusion output", making it difficult to achieve the robustness and precision required for business.

[0005] Therefore, there is an urgent need for a new method of short-time heavy rainfall estimation that can deeply couple radar observations, environmental parameters and physical mechanisms, achieve dynamic, adaptive and high-precision short-time heavy rainfall estimation. SUMMARY

[0006] The present application aims to overcome the defects of the prior art and provide a radar short-time heavy rainfall estimation method based on classified echoes and environmental physical constraints, improving the accuracy, stability and physical consistency of rainfall estimation under complex weather conditions.

[0007] A radar short-time heavy rainfall estimation method based on classified echoes and environmental physical constraints, comprising the following steps: Step S1, multi-source data integration and collaborative gridding preprocessing: A multi-source data input layer is constructed to synchronously access radar echo extrapolation prediction data, multi-dimensional environmental observation data and ground precipitation real-time data, perform data quality control and space-time registration, and generate a multi-dimensional feature input field with space-time coordination; In step S2, a precipitation type is dynamically identified based on multi-feature fusion, and a Z-R relationship is adaptively selected: Multi-dimensional microphysical features of radar echoes at each grid point are extracted, an integrated learning classification model is used to dynamically identify the precipitation type, and corresponding Z-R relationship parameters are adaptively selected from a dynamic Z-R relationship knowledge base according to the identification result to generate an initial precipitation rate field. In step S3, an estimated result is adaptively revised based on environmental physical processes: Evaporation effect revision based on low-level humidity conditions and raindrop spectrum revision based on mid-level humidity stratification are introduced to physically and consistently correct the initial precipitation rate field. In step S4, an estimated sequence is optimized and systematic deviation is corrected based on a sliding time window and real-time feedback: A sliding time window filtering technique is used to time-series smooth the estimated results of consecutive time to suppress oscillation, and a comparison result of recent prediction and real-time observation is used to correct the location and intensity of the current prediction field. In step S5, a multi-source information optimal fusion is performed and a refined heavy precipitation product is generated: The revised and corrected radar estimation field and ground rain station observation data are data assimilated and fused to generate a final high-spatial and temporal resolution short-time heavy precipitation estimation product.

[0008] Further, in step S1, the multi-dimensional environmental observation data at least includes temperature, relative humidity and melting layer height information from sounding, ground automatic station and numerical model reanalysis data; the space-time registration adopts a bilinear interpolation method or a Kriging interpolation method to unify all data to a space-time grid system consistent with the radar extrapolation prediction field.

[0009] Further, in step S2, the multi-dimensional microphysical features of the radar echoes include instantaneous reflectivity factor, horizontal gradient module, local texture uniformity and reflectivity vertical profile structure features; the integrated learning classification model is a random forest, a gradient boosting decision tree or a deep neural network model.

[0010] Further, in step S2, the dynamic Z-R relationship knowledge base stores a plurality of Z-R relationship formulas and Z-R relationship parameters corresponding to different precipitation types, and the Z-R relationship formulas and Z-R relationship parameters are obtained by type statistical analysis and fitting optimization on historical observation data.

[0011] Further, in the step S3, the evaporation effect correction specifically includes diagnosing the near-surface layer average relative humidity of the grid point, determining that there is a significant evaporation effect when the humidity is lower than a preset threshold, and down-regulating or filtering the initial precipitation rate through a weakening algorithm based on the humidity function; and the raindrop spectrum correction specifically includes dynamically adjusting the coefficient value in the selected Z-R relationship through a nonlinear correction coefficient related to the environmental humidity in the middle and low layers, and calculating the nonlinear correction coefficient through an empirical model trained by historical observation data.

[0012] Further, in the step S4, the sliding time window filtering technique adopts a median filtering or an exponential weighted average algorithm; and the systematic bias correction is achieved by calculating the average displacement vector and the regional average intensity bias between the historical prediction field and the corresponding real-time field, and applying the bias to the prediction field at the current time to realize spatial translation and intensity scaling.

[0013] Further, in the step S5, the data assimilation fusion adopts an optimal interpolation method or a geographic weighted regression method to fuse the processed radar estimation field with sparse but accurate ground automatic station precipitation observation, to finally locally calibrate and improve the accuracy of the radar estimation field by using the station data, and to finally generate a high spatiotemporal resolution short-time heavy precipitation estimation product, which includes a precipitation rate field, a precipitation type identifier, and a probability distribution map of a strong convective region.

[0014] In summary, due to the adoption of the above technical solutions, the present application has the following advantages: 1. Breaking the limitation of fixed Z-R relationship: dynamically identifying the precipitation type and matching the optimal Z-R relationship through machine learning, fundamentally solving the estimation bias problem caused by the spatial heterogeneity of precipitation microphysical processes, and test results show that the intensity accuracy and drop zone consistency of 1-kilometer high-resolution grid prediction are outstanding, and the TS skill evaluation is obviously positive skill compared with other objective prediction products, showing stable prediction skill.

[0015] 2. Deeply integrating environmental physical constraints: innovatively introducing key physical processes such as low-layer evaporation and environmental humidity into the correction link, making the estimation results more consistent with the precipitation formation mechanism under actual atmospheric conditions, and enhancing the physical rationality and interpretability of the method.

[0016] 3. Building a closed-loop feedback optimization system: through sliding time window smoothing and real-time guided bias correction, random errors are effectively suppressed, systematic bias is corrected, and the stability in time sequence and the accuracy in space of the estimation results are significantly improved.

[0017] 4. The optimal fusion of multi-source information is realized: the spatial continuity advantage of radar is combined with the point precision advantage of rain gauge, and the final product which maintains details and has calibrated accuracy is generated through a scientific fusion algorithm, thereby improving the reliability of business application.

[0018] 5. Strong business potential is possessed: the method has clear process, high modularization degree and strong automation capability, and can be seamlessly connected to the existing meteorological business system, thereby providing high-quality basic data for short-term weather forecast, hydrological simulation and disaster warning. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is the overall architecture diagram of the application.

[0020] Figure 2 It is the process schematic diagram of the application.

[0021] Figure 3 It is the query and matching schematic diagram of the dynamic Z-R relationship knowledge base of the application.

[0022] Figure 4 It is the environment physical process subscription schematic diagram of the application.

[0023] Figure 5 It is the process schematic diagram of the sliding time window optimization and systematic deviation correction of the application.

[0024] Figure 6 It is the short-term heavy rain contrast chart of the estimation effect of the traditional method and the application on August 21, 2025 02:00-22, 05:00. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical scheme and advantages of the application more clear and obvious, the technical scheme in the embodiments of the application is described clearly and completely below in combination with embodiments. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Meanwhile, the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0026] A radar short-term heavy rain estimation method based on classified echo and environmental physical constraint comprises the following steps: Step 1, multi-source data integration fusion and collaborative gridding preprocessing: A multi-source data input layer is constructed, and radar echo extrapolation forecast field, multi-dimensional environmental observation field and real-time observation data of ground rain gauge are synchronously accessed. Quality control including noise filtering, non-meteorological echo identification and elimination is performed on the radar data; the environmental parameters and rain gauge data are uniformly registered to the time-space grid system consistent with the radar forecast field by using bilinear interpolation method and Kriging spatial analysis technology, thereby forming a three-dimensional feature input field with time-space alignment and collaborative resolution.

[0027] Step 2, Dynamic precipitation type identification based on multi-feature fusion and Z-R relationship adaptive initial selection Extract the multi-dimensional features of radar echo at each grid point, including instantaneous reflectivity intensity, horizontal gradient module, local texture uniformity feature, and reflectivity vertical profile structure feature. Input the feature vector into the pre-trained ensemble learning classification model, output the precipitation type probability distribution of each grid point, and determine it as stratiform cloud precipitation, mixed precipitation or convective precipitation according to the maximum a posteriori probability criterion. Based on the built-in dynamic Z-R relationship knowledge base, appropriate Z-R relationship parameter pairs are assigned to each grid point according to the real-time classification result, realizing the spatial typing and dynamic initial selection of Z-R relationship, and generating the initial precipitation rate field.

[0028] Step 3, Adaptive correction of estimation results driven by environmental physical processes Introduce dynamic correction modules based on environmental physical parameters, including: Evaporation effect correction of low-level humidity conditions: Diagnose the relative humidity of the grid point near the ground layer. When the humidity is lower than the preset threshold, it is determined that there is significant evaporation effect, and the evaporation weakening algorithm based on humidity function is started to physically consistent down or filter out the initial precipitation rate.

[0029] Humidity layering raindrop spectrum correction: Introduce a correction coefficient related to the total atmospheric precipitable water or the average humidity in the low and middle layers to dynamically adjust the coefficient in the Z-R relationship. The correction coefficient is calculated by a nonlinear empirical model trained from historical observation data to correct the influence of raindrop spectrum changes caused by environmental humidity differences on radar estimation.

[0030] Step 4, Estimation sequence optimization and systematic bias correction based on sliding time window and real-time feedback Use sliding time window filtering technology to optimize the estimation results of consecutive times in time sequence. By applying median filtering or exponential weighted average to the precipitation rate sequence in the window, the estimation value oscillation caused by instantaneous jump of echo type is suppressed. At the same time, a "recent live guided" bias correction mechanism is established: by comparing the recent forecast with the corresponding live observation at the moment, the systematic displacement vector and intensity bias field are calculated, and this bias model is applied to the current forecast field to realize the adaptive correction of the spatial position and intensity of the future short-time estimation results.

[0031] Step 5, Optimal fusion of multi-source information and generation of refined heavy precipitation products The radar estimation field processed by the above steps is fused with sparse but accurate ground automatic station precipitation observations using optimal interpolation or geographically weighted regression algorithm. The station data is used to perform the final local calibration and precision improvement of the radar estimation field. Finally, a high spatial and temporal resolution, physically consistent short-term heavy precipitation grid estimation product is output, which includes precipitation rate field, precipitation type identification and strong convective area probability distribution map.

[0032] Embodiment This embodiment shows a business short-term heavy precipitation forecasting system integrating the method of the present application.

[0033] Step 1, multi-source data integration and collaborative gridding preprocessing The system synchronously reads the following data sources through the data interface: Radar data: from the AI extrapolation model based on U-Net architecture, output the reflectivity factor prediction grid field of 0-3 hours in the future, time interval 6 minutes, spatial resolution 1 km.

[0034] Environmental data: extract temperature, relative humidity, potential height and other parameters from numerical model reanalysis field to obtain two-dimensional / three-dimensional environmental field; obtain real-time temperature, pressure, humidity, wind observation from ground automatic station.

[0035] Precipitation live: access to the hourly cumulative precipitation data of thousands of automatic rain stations in the region.

[0036] In the preprocessing process, the clutter recognition algorithm based on spectral texture analysis is applied to the radar data to effectively eliminate ground objects and non-precipitation echoes. Then, all environmental data and rain station data are strictly registered to the 1 km x 1 km latitude and longitude grid consistent with the radar prediction field, and the time reference is unified to the world time, and the time step is unified to 6 minutes.

[0037] Step 2, dynamic recognition of precipitation type based on multi-feature fusion and adaptive initial selection of Z-R relationship For each 1 km x 1 km grid point, the following feature vectors are calculated at each prediction time: F1: reflectivity factor Z of the grid point.

[0038] F2: Calculate the horizontal gradient module of Z value within the 3x3 window centered on the grid point Z|.

[0039] F3: Calculate the standard deviation σ_Z of Z value within the 5x5 local window as the texture uniformity index.

[0040] F4: Extract the reflectivity vertical profile above the point from the three-dimensional radar mosaic and calculate its maximum gradient value.

[0041] The feature vector [F1, F2, F3, F4] is input into a pre-trained random forest classifier. This classifier has been trained on historical radar and precipitation type labeled data, and can output the probabilities [P_strat, P_mix, P_conv] that the point belongs to

stratiform precipitation, mixed precipitation, convective precipitation

[0042] The system determines the final type according to the maximum probability principle. Subsequently, the dynamic Z-R relationship knowledge base is queried, for example: If it is determined to be stratiform precipitation, then Z = 200 * R^1.6 is applied.

[0043] If it is determined to be mixed precipitation, then Z = 250 * R^1.5 is applied.

[0044] If it is determined to be convective precipitation, then Z = 300 * R^1.4 is applied. In this way, an initial precipitation rate estimate R_initial is generated for each grid point.

[0045] Step 3, adaptive correction of estimation results driven by environmental physical processes Low-level evaporation effect correction: read the average relative humidity RH_low of the grid point below the 1000-meter height layer. Set the threshold to 70%. If RH_low < 70%, it is considered that the precipitation experiences significant evaporation during its descent. The correction algorithm is: R_evap = R_initial * (RH_low / 70%)^k, where k is an empirical index, or when RH_low is very low, directly set R_evap to zero.

[0046] Humidity layering raindrop spectrum correction: read the average relative humidity RH_mid of the 850hPa to 500hPa layer of the grid point. Introduce a correction coefficient γ, and the corrected Z-R relationship is Z = γ * a * R^b. The coefficient γ is calculated by a quadratic polynomial model fitted by historical data: γ = 1.0 + 0.015 * (RH_mid - 80) + 0.0002 * (RH_mid - 80)^2. When RH_mid > 80%, γ > 1, appropriately increase the estimated precipitation rate to correct the raindrop spectrum dominated by collisional growth; when RH_mid is low, γ ≤ 1, which plays a role in suppressing the estimated value.

[0047] Step 4, optimization of estimation sequence and systematic bias correction based on sliding time window and live feedback Sliding time window optimization: The system maintains a sliding window of 30 minutes (5 times) to store the precipitation rate sequence of each grid point after environmental correction [R_t-4, R_t-3,..., R_t]. The median filter is performed on the sequence of each grid point, and the median R_median is used as the optimization estimate value at time t, effectively filtering out isolated abnormal jumps.

[0048] Real-time correction of drop zone deviation: Let T0 be the current time. The system retrieves the precipitation field Fcst_T0 forecast at T0 time made at T0-30 minutes, and compares it with the radar inversion / living analysis field Obs_T0 at T0 time. The average displacement vector (Δx, Δy) of Fcst_T0 relative to Obs_T0 is calculated by cross-correlation algorithm, and the regional average intensity deviation ΔI = mean(Fcst_T0) / mean(Obs_T0) is calculated. Assuming that the deviation remains stable in the future 1-2 hours, the newly produced forecast field is shifted in space (-Δx, -Δy), and the entire field is multiplied by the correction factor 1 / ΔI.

[0049] Step 5, optimal fusion of multi-source information and generation of refined heavy precipitation product The radar estimation field R_radar after step 4 optimization and correction is fused with the automatic station observation R_gauge at T0 time. Optimal interpolation method is adopted: R_final = R_radar + W * (R_gauge - R_radar). Wherein, W is the weight matrix, which is determined by the radar estimation error covariance and the station observation error, to ensure that in the vicinity of the station, the fusion result approaches the accurate station observation, and in the area without station, the radar estimation after multiple corrections is trusted.

[0050] Finally, the system outputs the short-term heavy precipitation grid prediction product covering the predetermined area (such as Northeast China), with spatial resolution of 1 km, 0-3 hours in the future every 6 minutes. The product is stored in standard NetCDF or GRIB2 format, containing precipitation rate, precipitation type, estimation uncertainty and other data layers, and is displayed through the visualization platform, such as Figure 6 As shown in the figure, the present application significantly improves the capture of heavy precipitation center intensity and drop zone compared with the traditional fixed Z-R relationship method. Among them, TS_LN represents the TS score of the product of the present method, TS_GJ is the TS score of the traditional method, STS is the relative skill score, BIAS_LN represents the BIAS score of the product of the present method, BIAS_GJ represents the BIAS score of the product of the traditional method, and SBI is the relative skill score The above is the embodiment of the present application. The foregoing is the various preferred embodiments of the present application, and the preferred embodiments in the various preferred embodiments can be arbitrarily superimposed and combined if not obviously contradictory or with a certain preferred embodiment as a prerequisite. The embodiments and specific parameters in the embodiments are only for clearly describing the verification process of the application and are not used to limit the patent protection scope of the application. The patent protection scope of the application is still subject to its claims. Any equivalent structural changes made by using the content of the specification and drawings of the application should also be included in the protection scope of the application.

Claims

1. A radar short-time heavy precipitation estimation method based on classification echoes and environmental physical constraints, characterized in that, The method comprises the following steps: Step S1, multi-source data integration and fusion and collaborative gridding preprocessing: Construct a multi-source data input layer, synchronously access radar echo extrapolation prediction data, multi-dimensional environmental observation data and ground precipitation real-time data, perform data quality control and space-time registration, and generate a space-time collaborative multi-dimensional feature input field; Step S2, dynamic identification of precipitation type based on multi-feature fusion and self-adaptive initial selection of Z-R relationship: Extract the multi-dimensional microphysical features of radar echo at each grid point, use a pre-trained integrated learning classification model to dynamically identify the precipitation type, and self-adaptively select the corresponding Z-R relationship parameters from the dynamic Z-R relationship knowledge base according to the identification result to generate an initial precipitation rate field; Step S3, adaptive correction of estimation results driven by environmental physical processes: Introduce evaporation effect correction based on low-level humidity conditions and raindrop spectrum correction based on mid-level humidity stratification to physically correct the initial precipitation rate field; Step S4, estimation sequence optimization and systematic bias correction based on sliding time window and real-time feedback: Use a sliding time window filtering technique to perform time sequence smoothing on the estimation results of consecutive time to suppress oscillation, and use the comparison results of recent prediction and real-time observation to perform systematic bias correction on the landing area and intensity of the current prediction field; Step S5, multi-source information optimal fusion and fine heavy precipitation product generation: Fuse the radar estimation field and ground rain station observation data after correction and correction to generate a final high-spatial and temporal resolution short-time heavy precipitation estimation product.

2. The method for radar short-time heavy rain estimation based on classification echo and environmental physical constraint according to claim 1, characterized in that, In step S1, the multi-dimensional environmental observation data at least includes temperature, relative humidity and melting layer height information from sounding, ground automatic station and numerical model reanalysis data; the space-time registration adopts a bilinear interpolation method or a Kriging interpolation method to unify all data to a space-time grid system consistent with the radar extrapolation prediction field.

3. The method for radar short-time heavy rain estimation based on classification echo and environmental physical constraint according to claim 1, characterized in that, In step S2, the multi-dimensional microphysical features of the radar echo include instantaneous reflectivity factor, horizontal gradient module, local texture uniformity and reflectivity vertical profile structure features; the integrated learning classification model is a random forest, a gradient boosting decision tree or a deep neural network model.

4. The method for radar short-time heavy precipitation estimation based on classification echoes and environmental physical constraints according to claim 1, characterized in that, In step S2, the dynamic Z-R relationship knowledge base stores a plurality of Z-R relationship formulas and Z-R relationship parameters corresponding to different precipitation types, and the Z-R relationship formulas and Z-R relationship parameters are obtained by type statistical fitting and optimization on historical observation data.

5. The method for radar short-time heavy precipitation estimation based on classification echoes and environmental physical constraints according to claim 1, characterized in that, In step S3, the evaporation effect correction is specifically: diagnosing the average relative humidity of the grid point near the ground layer, when the humidity is lower than a preset threshold, it is determined that there is a significant evaporation effect, and the initial precipitation rate is adjusted downward or filtered out through a weakening algorithm based on the humidity function; the raindrop spectrum correction is specifically: dynamically adjusting the coefficient value in the selected Z-R relationship through a nonlinear correction coefficient related to the environmental humidity of the middle and low layers, and the nonlinear correction coefficient is calculated by an empirical model trained from historical observation data.

6. The method for radar short-time heavy precipitation estimation based on classification echoes and environmental physical constraints according to claim 1, characterized in that, In the step S4, the sliding time window filtering technique adopts median filtering or exponential weighted average algorithm; the systematic bias correction is achieved by calculating the average displacement vector and regional average intensity bias between the historical forecast field and the corresponding real-time field, and applying the bias to the forecast field at the current time to realize spatial translation and intensity scaling.

7. The method for radar short-time heavy precipitation estimation based on classification echoes and environmental physical constraints according to claim 1, characterized in that, In the step S5, the data assimilation fusion adopts optimal interpolation method or geographic weighted regression method to fuse the processed radar estimation field with sparse but accurate ground automatic station precipitation observation, to use the station data to perform final local calibration and precision improvement on the radar estimation field, and finally generate a high spatiotemporal resolution short-time heavy precipitation estimation product, which includes a precipitation rate field, a precipitation type identification, and a probability distribution map of a strong convective region.

Images