Precipitation target identification and revision method based on adaptive threshold
By dynamically identifying precipitation targets using an adaptive threshold method and employing connected component algorithms and neighborhood range adjustments, the pattern recognition problem caused by fixed thresholds is solved, resulting in more accurate precipitation forecasts that can adapt to complex weather systems and improve the accuracy and reliability of forecasts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU PLATEAU METEOROLOGICAL INST OF CHINA METEOROLOGICAL ADMINISTRATION
- Filing Date
- 2024-10-22
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies rely on fixed thresholds in precipitation forecasting, leading to incomplete pattern recognition or false positives, making it impossible to accurately identify weak precipitation systems. Furthermore, the lack of objective evaluation criteria affects forecast accuracy and reliability.
An adaptive threshold method is adopted to dynamically identify precipitation targets through connected component algorithm and neighborhood range adjustment, and target matching and correction are performed based on area ratio and matching degree to optimize the spatial distribution of precipitation forecast.
It improves the spatial verification accuracy and reliability of precipitation forecasts, adapts to changes in complex weather systems, reduces location bias and intensity mismatch, and enhances the accuracy and reliability of forecasts.
Smart Images
Figure CN119439317B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of weather verification and correction, specifically a precipitation target identification and correction method based on adaptive thresholds. Background Technology
[0002] MODE (Method for Object-Based Diagnostic Evaluation) is an effective means of evaluating weather forecast performance. While focusing on traditional skill scoring, it places greater emphasis on the spatial differences between forecasts and actual conditions at different scales. The MODE method defines and calculates different attributes of objects, sets weight coefficients for these attributes, and uses a fuzzy logic algorithm to calculate the total benefit function of forecast performance, thereby judging the overall forecast performance. Object differentiation and identification have clear boundaries. Object identification uses convolution operators to determine the objects to be processed, making the original data field smoother and more continuous. By controlling the range of meteorological element thresholds, targets are filtered for verification, and key severe weather targets of concern are selected. After precipitation objects are identified, a series of spatial attributes are assigned. Spatial attribute parameters describing the target individual are calculated, as well as the attributes of the geometric and spatial relationships between any target individual and any target individual in the actual field. Once all object attributes are determined, the similarity between any two objects (from different fields) is calculated using the attribute diagnostic parameters of the object pairs. The similarity calculation uses a fuzzy logic algorithm, inputting all object attributes into it to calculate the total benefit function of forecast performance.
[0003] The MODE method relies on a fixed threshold to identify objects. Under different weather systems, the observation field and the model respond differently to the threshold, exhibiting significant differences in intensity and morphology. For weak precipitation systems or weak model forecasts, a single threshold may not be able to correctly identify these objects, leading to incomplete model recognition or missed detections, or too many false positives in strong models, causing object shape distortion and size deviation. Summary of the Invention
[0004] To improve the overall accuracy and reliability of forecast verification, this application provides a precipitation target identification and correction method based on adaptive threshold.
[0005] The technical solution adopted by the present invention to solve the above problems is:
[0006] A precipitation target identification and correction method based on adaptive thresholds includes:
[0007] Step 1: After preprocessing the original field, a gridded spatial field is obtained. The original field includes the observation field and the prediction field.
[0008] Step 2: Select the study area in the space field and calculate the precipitation thresholds for the observation field and the forecast field respectively;
[0009] Step 3: Based on the calculated precipitation threshold, filter the precipitation values in the observation field and the forecast field to obtain the mask field. Specifically, precipitation values less than the precipitation threshold are marked as 0, and precipitation values greater than or equal to the precipitation threshold are marked as 1.
[0010] Step 4: Extract connected targets from the filtered observation and forecast fields using a connected component algorithm and merge the targets:
[0011] Identify connected targets in the observation and forecast fields that exceed their respective precipitation thresholds;
[0012] The neighborhood range (near_dis) and the neighborhood ratio (near_rate) between connected targets are determined based on the size of the connected target. If the neighborhood ratio (near_rate) is 0, the connected targets are not merged. If the neighborhood ratio (near_rate) is greater than or equal to the neighborhood ratio threshold, the two connected targets are merged to form a new connected target. This process is repeated until the neighborhood ratio (near_rate) between all connected targets is less than the neighborhood ratio threshold; at this point, the connected targets are considered precipitation targets. , and These represent the number of grid points for the precipitation target in the forecast field and the precipitation target in the observation field, respectively. and It is the number of grid points within the neighbor range near_dis of each precipitation target;
[0013] Step 5: Set the neighborhood range cover_dis and matching degree cover_rate based on the precipitation target, and match the precipitation target in the observation field with the precipitation target in the forecast field based on the neighborhood range cover_dis and matching degree cover_rate respectively;
[0014] Step 6: Calculate the area ratio of precipitation targets in the forecast field based on the matching results. The area ratio is the number of grid points with a 1-fold margin divided by the area of the rectangle formed by the major and minor axes of the precipitation target.
[0015] Step 7: If the area ratio is less than the set area ratio threshold, adjust the neighborhood range near_dis and the neighborhood ratio threshold and repeat steps 4-6 until the area ratio is greater than the area ratio threshold.
[0016] Step 8: Calculate the spatial attribute error of the target pair based on the matching results, including centroid position, intensity, and area error;
[0017] Step 9: Optimize and adjust based on spatial attribute errors.
[0018] Furthermore, step 1 specifically includes:
[0019] Step 11: Preprocess the observation field and forecast field, including filling in missing data and interpolating them into a grid of the same resolution;
[0020] Step 12: Determine the smoothing coefficient and perform filtering on the observation field and forecast field based on the smoothing coefficient.
[0021] Furthermore, the calculation method for the precipitation threshold in step 2 is as follows: , In the formula, Let x be the precipitation threshold, n be the number of data points in the dataset, and p be the percentile, ranging from 0 to 1. This indicates taking the integer part of np. This represents the precipitation amount after taking the last digit of np.
[0022] Furthermore, for mesoscale precipitation systems, the neighborhood range near_dis is 30-100km.
[0023] Furthermore, the neighborhood range near_dis is set to 50km, and the neighborhood ratio threshold is set to 0.2.
[0024] Furthermore, step 5 specifically includes:
[0025] Based on the merged precipitation target, set the neighborhood range (cover_dis) and the matching degree (cover_rate) settings.
[0026] Calculate the matching degree cover_rate between the precipitation targets to be matched. If the matching degree cover_rate is greater than the set value of the matching degree cover_rate, the match is successful.
[0027] Match ,in and These represent the number of grid points for the precipitation target in the forecast field and the precipitation target in the observation field, respectively. and It represents the number of grid points within the neighboring range cover_dis of each precipitation target.
[0028] Furthermore, the neighborhood range cover_dis is set to 50km, and the matching degree cover_rate is set to 0.7.
[0029] Furthermore, in step 7, the area ratio threshold is set to 0.5.
[0030] Furthermore, step 9 specifically includes:
[0031] Based on the latitude and longitude error of the centroid position, the forecast field is shifted in the longitude and / or latitude directions;
[0032] Add or reduce the corresponding intensity error in the forecast field, and set the grid points with negative adjusted precipitation intensity to zero;
[0033] The area error is determined as a scaling factor. Based on the scaling factor, the forecast field is spatially stretched or shrunk to obtain the corrected precipitation field.
[0034] The advantages of this invention compared to existing technologies are: it provides more accurate spatial verification of forecast fields by dynamically adjusting thresholds based on the distribution of adaptive thresholds to capture precipitation systems of different intensities and scales; simultaneously, it uses the area ratio as an objective quantitative indicator of the reasonableness of target identification; after identifying different target pairs, it corrects the forecast field using the error information of the target pairs, optimizing the intensity and spatial distribution of precipitation forecasts. The method of this application not only improves the overall accuracy and reliability of forecast verification but also enables the forecast model to adapt to changes in complex weather systems. Attached Figure Description
[0035] Figure 1 The flowchart shows the precipitation target identification and correction method based on adaptive threshold.
[0036] Figure 2 Schematic diagram of different percentile precipitation thresholds for different individual observation fields and forecast fields;
[0037] Figure 3 This is a schematic diagram of the neighborhood range;
[0038] Figure 4 This is a schematic diagram of a heavy rainfall event in a slope area of Sichuan Province. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0040] This will be illustrated using a heavy rainfall event in a slope area of Sichuan Province as an example. Figure 1 As shown, the precipitation target identification and correction method based on adaptive threshold includes:
[0041] Step 1: After preprocessing the original field f(x,y), a gridded spatial field C(x,y) is obtained. The original field includes the observation field and the prediction field.
[0042] Preprocessing includes filling in missing data in the observation and forecast fields and interpolating them into a grid of the same resolution; determining the smoothing coefficient; and filtering the observation and forecast fields based on the smoothing coefficient. The smoothing coefficient can be selected according to the actual process, for example, a smoothing coefficient of 4 times the number of grid points, and there is no restriction here.
[0043] Step 2: Select the study area in the space field C(x,y) and calculate the precipitation thresholds for the observation field and the forecast field respectively. The calculation method for the precipitation threshold is as follows: , In the formula, Here, x is the sorted precipitation dataset, n is the number of data points in the dataset, and p is the percentile, ranging from 0 to 1. The larger the p value, the more weak precipitation values are filtered out. The target is usually heavy precipitation, in which case the p value can generally be 75%. This indicates taking the integer part of np. This represents the precipitation amount after taking the last digit of np. The precipitation threshold is dynamically adjusted based on the precipitation amount, resulting in captured precipitation systems that better reflect actual needs.
[0044] Let the sorted precipitation dataset be x = [1, 2, 3, 4, 5]. If p = 0.8, n If p=4, then m= =0.5 (4+5)=4.5; if p is 0.7, n If p = 3.5, then m = =4. Figure 2 Threshold results for different percentiles in different cases are given.
[0045] Step 3: Based on the calculated precipitation threshold, filter the precipitation values of the observation field and the forecast field to obtain the mask field M(X,Y). Specifically, precipitation values less than the precipitation threshold are marked as 0, and precipitation values greater than or equal to the precipitation threshold are marked as 1.
[0046] If connected targets are directly extracted and merged on the mask field, some precipitation information from the original field will be lost. To reduce information loss, the original observation field and the original forecast field can be assigned values by the obtained mask field to obtain the reconstructed field F(x,y)=M(x,y)f(x,y). Connected targets are then extracted and merged on the reconstructed field.
[0047] Step 4: Extract connected targets from the filtered observation and forecast fields using a connected component algorithm and merge the targets:
[0048] Identify connected targets in the observation and forecast fields that exceed their respective precipitation thresholds;
[0049] The neighborhood range (near_dis) and the neighborhood ratio (near_rate) between connected targets are determined based on the size of the connected targets. The neighborhood range (near_dis) refers to the range formed by extending the outer contour of the connected target outward by a distance D1. Figure 3 The dashed line indicates the neighborhood ratio. , and These represent the number of grid points for the precipitation target in the forecast field and the precipitation target in the observation field, respectively. and It is the number of grid points within the neighbor range (near_dis) of each precipitation target.
[0050] If the neighborhood ratio near_rate is 0, then connected targets are not merged; if the neighborhood ratio near_rate is greater than or equal to the neighborhood ratio threshold, then two connected targets are merged to form a new connected target; repeat the above steps until the neighborhood ratio near_rate between all connected targets is less than the neighborhood ratio threshold, at which point the connected targets are precipitation targets.
[0051] The specific values of distance D1 and the neighborhood ratio threshold can be set according to actual needs. In this embodiment, distance D1 is set to 50km and the neighborhood ratio threshold is set to 0.2. According to previous statistics, for mesoscale targets (20 to 200km), distance D1 can be set to 30-100km; the smaller the neighborhood ratio threshold, the lower the merging probability. The selection can be made based on the effect through continuous debugging.
[0052] Step 5: Set the neighborhood range cover_dis and the matching degree cover_rate, and match the precipitation targets in the observation field with the precipitation targets in the forecast field based on the neighborhood range cover_dis and the matching degree cover_rate.
[0053] Based on the merged precipitation target, the neighborhood range cover_dis and the matching degree cover_rate are set. The neighborhood range cover_dis is the allowable distance deviation when calculating the overlap of observed and predicted targets, that is, the range formed by extending the outer contour of the target outward by a distance D2. In this study, the target is mesoscale, and D2 is set to 50km. The matching degree cover_rate is set to 0.7. It is generally considered that a value above 0.7 is a successful match, but it can also be set to other values according to actual needs.
[0054] Calculate the matching degree cover_rate between the precipitation targets to be matched. If the matching degree cover_rate is greater than the set value of the matching degree cover_rate, the match is successful.
[0055] Match ,in and These represent the number of grid points for the precipitation target in the forecast field and the precipitation target in the observation field, respectively. and It represents the number of grid points within the neighboring range cover_dis of each precipitation target.
[0056] Step 6: Calculate the area ratio of precipitation targets in the forecast field based on the matching results. The area ratio is the number of grid points with a 1-fold margin divided by the area of the rectangle formed by the major and minor axes of the precipitation target.
[0057] Step 7: If the area ratio is less than the set area ratio threshold, adjust the neighborhood range near_dis and the neighborhood ratio threshold, and repeat steps 4-6 until the area ratio is greater than the area ratio threshold to complete the precipitation target identification. In this embodiment, the area ratio threshold is set to 0.5, which is considered to be good at identifying the target. Other values can also be set, and there are no restrictions here.
[0058] Traditional identification methods often rely on subjective judgment, while this application provides a more quantitative evaluation standard through area ratio, which can effectively measure whether the object of pattern recognition matches the actual observation, more objectively evaluate forecast performance, and avoid the bias caused by relying solely on subjective judgment.
[0059] Step 8: Calculate the spatial attribute error of the target pair based on the matching results, including the centroid position, intensity, and area error, where the area error is the ratio of the precipitation target area in the forecast field to that in the observation field.
[0060] Step 9: Optimize and adjust based on spatial attribute errors:
[0061] Based on the latitude and longitude error of the centroid position, the forecast field is shifted in the longitude and / or latitude directions;
[0062] Add or reduce the corresponding intensity error in the forecast field, and set the grid points with negative adjusted precipitation intensity to zero;
[0063] The area error is determined as a scaling factor. Based on the scaling factor, the forecast field is spatially stretched or shrunk to obtain the corrected precipitation field.
[0064] The comprehensive corrections to the precipitation field significantly improve the consistency between the forecast and actual observation fields in terms of spatial location, precipitation intensity, and coverage. The corrected results more accurately reflect the actual precipitation situation, especially during heavy precipitation events, effectively reducing issues such as location bias, intensity mismatch, and overestimation or underestimation of the area in the forecast. This method is not only applicable to the precise correction of individual precipitation events but can also be extended to the same type of heavy precipitation events, systematically improving the accuracy and reliability of forecasts and providing more valuable forecast products for disaster early warning.
[0065] Figure 4 A schematic diagram of a case of heavy rainfall in a slope area of Sichuan Province: (a) is the observation field, (b) is the FY4B satellite precipitation product, and (c) is the corrected precipitation field. Unit: millimeters (mm).
Claims
1. A precipitation target identification and correction method based on adaptive threshold, characterized in that, include: Step 1: After preprocessing the original field, a gridded spatial field is obtained. The original field includes the observation field and the prediction field. Step 2: Select the study area in the space field and calculate the precipitation thresholds for the observation field and the forecast field respectively; The calculation method for the precipitation threshold is as follows: ， In the formula, Let x be the precipitation threshold, n be the number of data points in the dataset, and p be the percentile, ranging from 0 to 1. This indicates taking the integer part of np. This represents the precipitation amount after taking the last digit of np. Step 3: Based on the calculated precipitation threshold, filter the precipitation values in the observation field and the forecast field to obtain the mask field. Specifically, precipitation values less than the precipitation threshold are marked as 0, and precipitation values greater than or equal to the precipitation threshold are marked as 1. Step 4: Extract connected targets from the filtered observation and forecast fields using a connected component algorithm and merge the targets: Identify connected targets in the observation and forecast fields that exceed their respective precipitation thresholds; The neighborhood range (near_dis) and the neighborhood ratio (near_rate) between connected targets are determined based on the size of the connected target. If the neighborhood ratio (near_rate) is 0, the connected targets are not merged. If the neighborhood ratio (near_rate) is greater than or equal to the neighborhood ratio threshold, the two connected targets are merged to form a new connected target. This process is repeated until the neighborhood ratio (near_rate) between all connected targets is less than the neighborhood ratio threshold; at this point, the connected targets are considered precipitation targets. , and These represent the number of grid points for the precipitation target in the forecast field and the precipitation target in the observation field, respectively. and It is the number of grid points within the neighbor range near_dis of each precipitation target; Step 5: Set the neighborhood range cover_dis and matching degree cover_rate based on the precipitation target, and match the precipitation target in the observation field with the precipitation target in the forecast field based on the neighborhood range cover_dis and matching degree cover_rate respectively; Step 6: Calculate the area ratio of precipitation targets in the forecast field based on the matching results. The area ratio is the number of grid points with a mask of 1 divided by the area of the rectangle formed by the major and minor axes of the precipitation target. Step 7: If the area ratio is less than the set area ratio threshold, adjust the neighborhood range near_dis and the neighborhood ratio threshold and repeat steps 4-6 until the area ratio is greater than the area ratio threshold. Step 8: Calculate the spatial attribute error of the target pair based on the matching results, including centroid position, intensity, and area error; Step 9: Optimize and adjust based on spatial attribute errors: Based on the latitude and longitude errors of the centroid position, shift the forecast field in the longitude and / or latitude directions; Add or reduce the corresponding intensity error in the forecast field, and set the grid points with negative adjusted precipitation intensity to zero; The area error is determined as a scaling factor. Based on the scaling factor, the forecast field is spatially stretched or shrunk to obtain the corrected precipitation field.
2. The precipitation target identification and correction method based on adaptive threshold according to claim 1, characterized in that, Step 1 specifically includes: Step 11: Preprocess the observation field and forecast field, including filling in missing data and interpolating them into a grid of the same resolution; Step 12: Determine the smoothing coefficient and perform filtering on the observation field and forecast field based on the smoothing coefficient.
3. The precipitation target identification and correction method based on adaptive threshold according to claim 1, characterized in that, For mesoscale precipitation systems, the neighborhood range of near_dis is 30-100km.
4. The precipitation target identification and correction method based on adaptive threshold according to claim 3, characterized in that, The neighborhood range near_dis is set to 50km, and the neighborhood ratio threshold is set to 0.
2.
5. The precipitation target identification and correction method based on adaptive threshold according to claim 1, characterized in that, Step 5 specifically involves: Based on the merged precipitation target, set the neighborhood range (cover_dis) and the matching degree (cover_rate) settings. Calculate the matching degree cover_rate between the precipitation targets to be matched. If the matching degree cover_rate is greater than the set value of the matching degree cover_rate, the match is successful. Match ,in and These represent the number of grid points for the precipitation target in the forecast field and the precipitation target in the observation field, respectively. and It represents the number of grid points within the neighboring range cover_dis of each precipitation target.
6. The precipitation target identification and correction method based on adaptive threshold according to claim 5, characterized in that, The neighborhood range cover_dis is set to 50km, and the matching degree cover_rate is set to 0.
7.
7. The precipitation target identification and correction method based on adaptive threshold according to claim 1, characterized in that, In step 7, the area ratio threshold is set to 0.5.