Quantitative rainfall forecasting method and system fusing high-precision geographic information

Abstract

The invention discloses a quantitative rainfall forecasting method and system fused with high-precision geographic information. The method comprises the following steps: acquiring high-precision geographic information data, meteorological radar data and historical meteorological knowledge data; encoding and extracting to obtain a geographic information vector data set and a radar vector data set, and fusing historical meteorological knowledge data to form multi-source fusion vector data; taking a plurality of groups of multi-source fusion vector data sets with different spatio-temporal features as input, removing related spatio-temporal features from the multi-source fusion vector database through a feature extraction network, and outputting to obtain a non-spatio-temporal dependent recombined vector data set; and taking the non-space-time dependent recombinant vector data set as input, decoding through a feature recognition network, and outputting a non-time-sequence dependent quantitative rainfall forecast result. According to the invention, the accuracy of quantitative rainfall forecast of short temporary rainfall can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of near-term quantitative precipitation forecasting, and in particular to a quantitative precipitation forecasting method and system integrating high-precision geographic information. Background Art

[0002] Precipitation is a common and important natural phenomenon. It is indispensable to the daily life, agricultural production, aerospace and other fields of human society. Precipitation involves many fields such as physics, remote sensing, hydrology, atmospheric circulation, etc. Appropriate precipitation can bring convenience and benefits to human production and life, but excessive precipitation may cause floods and secondary disasters such as landslides and urban waterlogging, which seriously affect human production and life, cause great property losses, and even threaten human life safety.

[0003] Nowcasting of quantitative precipitation has always been an important research field in meteorology at home and abroad, and plays an important role in the aerospace field, agricultural system, transportation field, and flood prevention and early warning. Accurate precipitation forecast can not only warn of natural disasters such as heavy rain and floods in advance, effectively reduce economic losses, protect life and property safety, but also benefit all aspects of people's lives.

[0004] At present, the prediction of short-term precipitation is a hot spot and difficulty in domestic disaster prevention and reduction. Meteorological forecast is based on its complex physical mechanism of formation, and it needs to be comprehensively studied and judged according to multiple factors. Among them, the extrapolation algorithm based on detection data is not only simpler, less computational, and faster than other existing algorithms, but also can give full play to the characteristics of high spatiotemporal resolution of detection data, and can better meet the needs of real-time forecasting. However, there are the following deficiencies in the current quantitative precipitation forecasting work: In the current practical application process, the satellite image data obtained still mainly rely on the operator's historical experience for subjective evaluation and judgment. Radar echo extrapolation is difficult to predict the changes in the formation and disappearance of convective clouds. At the same time, various types of precipitation are affected by the complex physical processes of multiple factors such as the type of precipitation particles, local weather changes, and underlying surface and terrain. The precipitation forecast is outputted by radar echo signals alone, and the original information is missing too much, which limits the upper limit of this method. In addition, no targeted research on quantitative precipitation forecasting has been carried out, and the business pain point of insufficient accuracy of quantitative precipitation forecasting cannot be solved.

[0005] In addition, the triggering and evolution process of severe convective systems are affected by the background weather system, the configuration of regional environmental conditions and their changes over time, and are closely related to many factors such as local topography and landforms. How to process the complex observation data in an orderly manner and then combine it with conventional data to improve the accuracy of quantitative precipitation forecasts is another important difficulty. Summary of the invention

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a quantitative precipitation forecasting method and system integrating high-precision geographic information, which solves the technical problem of insufficient accuracy of quantitative precipitation forecasting.

[0007] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0008] A quantitative precipitation forecasting method integrating high-precision geographic information comprises the following steps:

[0009] Obtain high-precision geographic information data, weather radar data, and historical meteorological knowledge data;

[0010] A geographic information vector data set and a radar vector data set are respectively extracted from the high-precision geographic information data and the weather radar data through corresponding encoders;

[0011] Fuse geographic information vector data sets, radar vector data sets and historical meteorological knowledge data to form multi-source fused vector data;

[0012] Taking multiple sets of multi-source fused vector data sets with different spatiotemporal characteristics as input, the features of the sequence data are extracted through a feature extraction network and then compressed and reorganized to eliminate the time order influence of the sequence data, and the direct relationship of the data relative space is removed by calculating the absolute geographic location information, so as to remove the relevant spatiotemporal characteristics from the multi-source fused vector database and output a reorganized vector data set that is not spatiotemporally dependent.

[0013] The non-temporal and spatial dependent reorganized vector dataset is taken as input, decoded through the feature recognition network, and the non-time-series dependent quantitative precipitation forecast results are output.

[0014] As a further improvement of the method of the present invention:

[0015] Optionally, the high-precision geographic information data includes: national topographic data, land cover data, hydrological data, and historical meteorological data;

[0016] Weather radar data include: weather radar image information matrix;

[0017] Historical meteorological knowledge data includes: meteorological experience data representation and meteorological prior knowledge representation.

[0018] Optionally, extracting a geographic information vector dataset from the high-precision geographic information data by a corresponding encoder comprises the following steps:

[0019] Based on the national terrain data, surface cover data, hydrological data and historical meteorological data, the absolute geographical location, underlying surface environment and characteristic climate conditions related to precipitation are extracted as important characteristic factors. The autoencoder network is used to perform embedded vector encoding on the geographical location to obtain a geographic information vector dataset containing effective semantic information.

[0020] Optionally, the underlying surface environment includes: vegetation conditions, water and soil conditions, surface morphology and surface temperature.

[0021] Optionally, the characteristic climate conditions include: precipitation data and extreme weather data.

[0022] Optionally, fusing the geographic information vector data set, the radar vector data set, and the historical meteorological knowledge data to form multi-source fused vector data includes the following steps:

[0023] The geographic information vector data set and the radar vector data set are subjected to self-environment enhancement processing, and the historical meteorological knowledge data after self-environment enhancement processing are cross-feature enhanced, and then fused and output as multi-source fused vector data.

[0024] Optionally, features of the sequence data are extracted through a feature extraction network and then compressed and reorganized to eliminate the influence of the time order of the sequence data. This is achieved through feature reorganization, including: extracting common features through a gradient reversal layer, and extracting features through a feature mapping layer and projecting them to an orthogonal direction of the common features output by the gradient reversal layer, so as to map the features to a feature space orthogonal to the common features to extract pure features, and then reorganizing and fusing the common features and pure features through a multi-scale channel attention mechanism, and outputting the reorganized and fused vector data.

[0025] Optionally, a non-time-space dependent recombinant vector dataset is used as input, decoded through a feature recognition network, and a non-time-series dependent quantitative precipitation forecast result is output, including: taking a non-time-space dependent recombinant vector dataset as input, decoding the features through a Swin-transformer decoder, and rolling outputting future precipitation values ​​in the falling area; when it is necessary to output the total amount of precipitation forecast within a specified time scale in the future, a corresponding section of the rolling output of the future precipitation values ​​in the falling area is selected for statistical output.

[0026] Optionally, the future specified time scale ranges from 1 to 3 hours in the future.

[0027] As a general technical concept, the present invention also provides a quantitative precipitation forecast system integrating high-precision geographic information, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of any of the above methods when executing the computer program.

[0028] The present invention has the following beneficial effects:

[0029] The quantitative precipitation forecasting method and system integrating high-precision geographic information of the present invention fully considers factors such as the underlying surface and characteristic climate by integrating high-precision geographic information data, and integrates meteorological radar data. Feature extraction and recognition are performed based on multi-source fused vector data, wherein feature extraction adopts a feature extraction algorithm that is not time-space dependent, thereby improving the efficiency of the prediction algorithm, thereby improving the accuracy of the quantitative precipitation forecast for short-term precipitation.

[0030] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0032] Figure 1 It is a flow chart of a quantitative precipitation forecasting method integrating high-precision geographic information according to a preferred embodiment of the present invention;

[0033] Figure 2 is a schematic diagram of interpolation processing of geographic data in a preferred embodiment of the present invention;

[0034] Figure 3 It is a schematic diagram of the encoding process of the geographic information vector data set of the preferred embodiment of the present invention;

[0035] Figure 4 is a schematic diagram of the fusion process of multi-source fusion vector data in a preferred embodiment of the present invention;

[0036] Figure 5 Schematic diagram of the processing flow of the feature extraction network of the preferred embodiment of the present invention;

[0037] Figure 6 It is a schematic diagram of the processing flow of the feature recognition network of the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0039] The quantitative precipitation forecasting method and system integrating high-precision geographic information proposed in the embodiment of the present invention fully considers factors such as the underlying surface and characteristic climate by integrating high-precision geographic information data, and integrates meteorological radar data. Feature extraction and recognition are performed based on multi-source fused vector data, wherein the feature extraction adopts a feature extraction algorithm that is not time-space dependent, thereby improving the efficiency of the prediction algorithm, thereby improving the accuracy of the quantitative precipitation forecast for short-term precipitation.

[0040] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0041] Figure 1 The flowchart of the quantitative precipitation forecasting method integrating high-precision geographic information in this embodiment is shown in FIG. The quantitative precipitation forecasting method integrating high-precision geographic information in this embodiment of the present invention comprises the following steps:

[0042] S1. Obtain high-precision geographic information data, weather radar data, and historical meteorological knowledge data.

[0043] During implementation, high-precision geographic information data may include: national terrain data, land cover data, hydrological data, and historical meteorological data. Meteorological radar data include: meteorological radar image information matrix. Historical meteorological knowledge data include: meteorological experience data representation and meteorological prior knowledge representation.

[0044] Among them, during implementation, the absolute geographic location coordinates are clearly defined, and the coding model for the absolute geographic location is determined through literature research, domain knowledge and experimental optimization. The high-precision geographic information dataset of this embodiment mainly obtains terrain elevation, slope, slope direction and other information of the study area through satellite remote sensing technology (such as MODIS, Landsat, etc.) and digital terrain model (DTM). And collect and obtain terrain DEM data, meteorological underlying surface classification data, soil, hydrological data and other data.

[0045] Then, the collected data is processed. Geographic information data processing: The acquired high-precision geographic information data is cleaned. At the same time, the geographic data is interpolated according to the grid density of the weather radar data. The specific weighting method is inverse distance interpolation ( Figure 2): It estimates the pixel value by taking the average of the sample data points in the area of ​​each pixel to be processed. It has been widely used in spatial interpolation in GIS in recent years. This interpolation method assumes that each sampling point has a certain local influence, which decreases with the increase of distance. The points closer to the pixel being processed have a greater weight than the points farther away. A certain number of points or points within a certain radius can be used to determine the value of each output pixel.

[0046] 1) First calculate the distance from the location point to all points:

[0047]

[0048] Among them, d i is the distance from the position point to the i-th point, (x, y) is the coordinate of the position point, (x i ,y i ) is the coordinate of the i-th point.

[0049] 2) Calculate weights:

[0050]

[0051] Among them, w i is the weight of the i-th point, and n is the number of points in the range.

[0052] 3) Calculate the value of the interpolation point:

[0053]

[0054] Among them, Z 0 is the value of the interpolation point, Z(X i ,Y i ) is the average value within the range of the interpolation point i.

[0055] S2. Extracting a geographic information vector data set and a radar vector data set from the high-precision geographic information data and the weather radar data through corresponding encoders.

[0056] Among them, see Figure 3 , extracting a geographic information vector dataset from the high-precision geographic information data through a corresponding encoder, including the following steps:

[0057] The resource integration is based on the metafile of the meteorological bureau's geographic information data. It includes national terrain data, surface cover data, hydrological data, and historical meteorological data (meteorological record data of ground meteorological stations, and climate analysis data). The correspondence between meteorological radar data and absolute geographical location is fully considered. The self-encoding network is used to perform embedded vector encoding on the geographical location, and the absolute geographical location, underlying surface environment (the underlying surface environment in this embodiment includes: vegetation conditions, water and soil conditions, surface morphology and surface temperature) and characteristic climate conditions (the characteristic climate conditions in this embodiment include: precipitation data and extreme weather data) related to precipitation are extracted as important feature factors to obtain efficient semantic information, form a high-dimensional geographic information vector data set, and construct a reusable geographic information vector data set containing effective semantic information.

[0058] On the other hand, the radar vector data set is extracted from the weather radar data by a corresponding encoder, including the following steps: the weather radar image data is encoded to form radar vector data, and then interpolated to form a radar vector data set that is consistent with the grid points and dimensions of the geographic information vector data.

[0059] S3, integrate geographic information vector data set, radar vector data set and historical meteorological knowledge data to form multi-source fusion vector data. Figure 4 , integrating geographic information vector data set, radar vector data set and historical meteorological knowledge data to form multi-source fused vector data, including the following steps:

[0060] Using the multimodal feature fusion method, the geographic information vector dataset (high-dimensional vector matrix data) and the radar vector dataset are processed by self-environment enhancement (ECA, Ego-Context Augment) and the historical meteorological knowledge data after self-environment enhancement (ECA, Ego-Context Augment) are cross-feature augmented (CFA, Cross-Feature Augment) processing. The data are uniformly mapped to high dimensions and the fused output is multi-source fused vector data.

[0061] S4. Multiple groups of multi-source fused vector data sets with different spatiotemporal characteristics are taken as input. The features of the sequence data are extracted through a feature extraction network and then compressed and reorganized to eliminate the influence of the time order of the sequence data. The direct relative spatial relationship of the data is removed by calculating the absolute geographic location information, so as to remove the relevant spatiotemporal features from the multi-source fused vector database and output a recombined vector data set that is not spatiotemporally dependent.

[0062] When implementing, see Figure 5, the features of the sequence data are extracted through the feature extraction network and then compressed and reorganized, and the data is processed by the differential transformation method to eliminate the time dependence of the sequence data, and then completed through feature reorganization, including: extracting common features through the gradient reversal layer, and extracting features through the feature mapping layer and projecting them to the orthogonal direction of the common features output by the gradient reversal layer, so as to map the features to the feature space orthogonal to the common features to extract pure features, and then reorganize and fuse the common features and pure features through the multi-scale channel attention mechanism, and output the reorganized and fused vector data.

[0063] Feature extraction is to obtain the characteristic information of data, which is tensor data that does not depend on time and space. The present invention uses a feature extraction algorithm that is not time-space dependent. It is to extract the characteristics of sequence data and then compress and reorganize them ( Figure 5 ), eliminating the time order of the sequence data. And by calculating the absolute geographic location information, the direct relationship of the data relative to space is removed.

[0064] Existing feature recombination usually uses feature summation (add) or feature combination (concat) to simply piece together different features. Such a simple processing method cannot solve the problem of spatiotemporal dependency of features. Therefore, the present invention proposes a new feature recombination module. The feature recombination fusion module is composed of a gradient reversal layer (GRL), a feature mapping layer (OPL) and an attention mechanism. The gradient reversal layer is used to extract common features, the feature mapping layer is used to extract pure features, and the multi-scale channel attention mechanism is used to recombine and fuse common features and pure features.

[0065] GRL reverses the gradient during the gradient back propagation process and adds a negative sign to the derivative value so that it does not move closer to the optimal solution. The reason for this is that common features are usually features that the classifier cannot distinguish.

[0066] OPL is used for feature extraction. The main step is to project the extracted features to the orthogonal direction of the common features output by the gradient reversal layer, and map the features to a feature space orthogonal to the common features. This space is completely pure. Assume that the common feature vector is H 1 , the feature H 2 In H 1 Projection is done on it, that is:

[0067] H 1 ′ =Proj(H 1 ,H 2 )

[0068] Among them, Proj is the projection formula:

[0069]

[0070] Among them, (x, y) is the coordinate position.

[0071] The multi-scale channel attention mechanism reorganizes and fuses the extracted common features with the pure features, and outputs the reorganized and fused vector data. This method is suitable for most attention features, including feature fusion caused by short connections and long connections. It can better solve the problems arising from the fusion of features of different scales.

[0072] S5. The non-temporal and spatial dependent reorganized vector dataset is taken as input, decoded through the feature recognition network, and the non-time-series dependent quantitative precipitation forecast results are output.

[0073] During implementation, the non-time-space dependent reorganized vector dataset is used as input, decoded through a feature recognition network, and a non-time-series dependent quantitative precipitation forecast result is output, including: the non-time-space dependent reorganized vector dataset is used as input, decoded through a Swin-transformer decoder (see Figure 6 ) to perform feature decoding and output the future precipitation value of the falling area in a rolling manner; when it is necessary to output the total amount of precipitation forecast within the range of a specified time scale in the future, a corresponding section of the future precipitation value of the falling area that is rolled out is selected for statistical output. In this embodiment, the range of the future specified time scale is 1 to 3 hours in the future. That is, the present invention is suitable for quantitative forecasting of short-term precipitation.

[0074] An embodiment of the present invention also provides a quantitative precipitation forecasting system integrating high-precision geographic information, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the computer program.

[0075] In summary, the present invention conducts targeted research on quantitative precipitation forecasting to solve the business pain point of insufficient accuracy of quantitative precipitation forecasting. The present invention adopts a high-precision geographic information coding algorithm, a multi-source data fusion algorithm, and an efficient feature extraction algorithm, first obtains high-precision geographic information data, then processes and encodes the data, then fuses the meteorological radar coded data, and extracts the features of the fused data, and calculates the features using a feature recognition algorithm. It is a quantitative precipitation forecasting algorithm that is not time-space dependent, and can improve the accuracy of quantitative precipitation forecasting of short-term precipitation.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A quantitative precipitation forecasting method integrating high-precision geographic information, characterized in that: The following steps are involved: Obtain high-precision geographic information data, weather radar data, and historical meteorological knowledge data; A geographic information vector data set and a radar vector data set are respectively extracted from the high-precision geographic information data and the weather radar data through corresponding encoders; Fusing the geographic information vector data set, radar vector data set and historical meteorological knowledge data to form multi-source fused vector data; Taking multiple groups of multi-source fused vector data sets with different spatiotemporal characteristics as input, extracting the features of the sequence data through a feature extraction network and then compressing and reorganizing them to eliminate the influence of the time sequence of the sequence data, and removing the direct relationship of the data relative to space through the calculation of absolute geographic location information, so as to remove the relevant spatiotemporal characteristics from the multi-source fused vector database, and outputting a reorganized vector data set that is not spatiotemporally dependent; The non-temporal and spatial dependent reorganized vector dataset is taken as input, decoded through the feature recognition network, and the non-time-series dependent quantitative precipitation forecast results are output.

2. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 1 is characterized in that: The high-precision geographic information data includes: national topographic data, surface coverage data, hydrological data and historical meteorological data; The weather radar data includes: an image information matrix of the weather radar; The historical meteorological knowledge data includes: meteorological experience data representation and meteorological prior knowledge representation.

3. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 2 is characterized in that: Extracting a geographic information vector dataset from high-precision geographic information data through a corresponding encoder includes the following steps: Based on the national terrain data, surface cover data, hydrological data and historical meteorological data, the absolute geographical location, underlying surface environment and characteristic climate conditions related to precipitation are extracted as important characteristic factors. The autoencoder network is used to perform embedded vector encoding on the geographical location to obtain a geographic information vector dataset containing effective semantic information.

4. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 2 is characterized in that: The underlying surface environment includes: vegetation conditions, water and soil conditions, surface morphology and surface temperature.

5. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 2 is characterized in that: The characteristic climate conditions include: precipitation data and extreme weather data.

6. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 2 is characterized in that: The geographic information vector data set, radar vector data set and historical meteorological knowledge data are integrated to form multi-source fused vector data, including the following steps: The geographic information vector data set and the radar vector data set are subjected to self-environment enhancement processing, and the historical meteorological knowledge data subjected to the self-environment enhancement processing are subjected to cross-feature enhancement processing, and then fused and output as multi-source fused vector data.

7. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 1, characterized in that: The feature extraction network extracts features from the sequence data and then compresses and reorganizes them to eliminate the influence of the time order of the sequence data. This is accomplished through feature reorganization, including: extracting common features through a gradient reversal layer, and extracting features through a feature mapping layer and projecting them to an orthogonal direction of the common features output by the gradient reversal layer, so as to map the features to a feature space orthogonal to the common features to extract pure features, and then reorganizing and fusing the common features and pure features through a multi-scale channel attention mechanism, and outputting the reorganized and fused vector data.

8. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 1, characterized in that: The method takes a non-time-space dependent reorganized vector data set as input, decodes it through a feature recognition network, and outputs a non-time-series dependent quantitative precipitation forecast result, including: taking a non-time-space dependent reorganized vector data set as input, decoding the features through a Swin-transformer decoder, and rolling outputting future precipitation values ​​in the falling area; when it is necessary to output the total amount of precipitation forecast within a specified time scale in the future, select a corresponding section of the rolling output of the future precipitation values ​​in the falling area for statistical output.

9. The quantitative precipitation forecasting method integrating high-precision geographic information according to claim 6, characterized in that: The future specified time scale ranges from 1 to 3 hours in the future.

10. A quantitative precipitation forecasting system integrating high-precision geographic information, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 9 are implemented.

Images