Rainfall intensity grade estimation method and system based on deep learning

Abstract

The invention discloses a precipitation intensity grade estimation method and system based on deep learning, and relates to the technical field of civil aviation meteorological monitoring and warning, and the system comprises a data collection module, a precipitation prediction module, a probability weight analysis module, and a preprocessing frequency calculation module. The data acquisition module comprises a historical observation data acquisition unit and a satellite data acquisition unit, the historical observation data acquisition unit is used for acquiring historical observation data of a ground meteorological observation station in a target area, and the satellite data acquisition unit is used for acquiring multispectral remote sensing image data of the target area. The method has the beneficial effects that the distribution condition of each rainfall intensity grade in multiple time periods is analyzed, the probability weight of the rainfall intensity in the corresponding time period is calculated, the frequency of image data preprocessing is adjusted in real time, high-precision rainfall intensity prediction can be realized, the space coverage range is wide, satellite data is accurately and rapidly extracted, and the accuracy of the rainfall intensity is improved. The model prediction accuracy is improved, and the use cost is saved.

Description

Technical Field

[0001] The present invention relates to the technical field of civil aviation meteorological monitoring and warning, and particularly to a method and system for estimating precipitation intensity levels based on deep learning. Background Art

[0002] Precipitation phenomenon is one of the most common weather factors affecting civil aviation flight activities. The magnitude of its intensity directly determines whether a flight can take off and land normally. Different from the precipitation level classification standard proposed by the China Meteorological Administration, precipitation in civil aviation is divided into three levels: small (weak), medium, and large (strong) according to intensity. Currently, the judgment of precipitation intensity in civil aviation mainly relies on manual observation supplemented by instrumental measurement data. However, due to the large area of the airport and the small coverage of manual visual and automatic weather observation systems (AWOS), there are large blind spots, resulting in an incomplete determination of the precipitation intensity across the whole field and an inability to meet the urgent needs of aviation users for refined meteorological services. The currently widely used C-band Doppler radar has a wide spatial coverage, but its spatio-temporal resolution is low, resulting in untimely radar data updates and difficulty in capturing rapid changes in precipitation intensity. Summary of the Invention

[0003] This part aims to provide a method and system for estimating precipitation intensity levels based on deep learning, which can improve the accuracy of precipitation intensity prediction by using satellite remote sensing image data.

[0004] To solve the above technical problems, the present invention provides the following technical solution: A method for estimating precipitation intensity levels based on deep learning, comprising the following steps: S100, collecting historical observation data of ground meteorological observation stations in the target area and multi-spectral remote sensing image data, and setting precipitation intensity levels according to the precipitation amount; S200, preprocessing the multi-spectral remote sensing image data, extracting key features and analyzing the relationship between the key features and precipitation intensity, and constructing and training a prediction model; S300, analyzing the distribution of each precipitation intensity level in multiple time periods respectively, and calculating the probability weight of precipitation intensity in the corresponding time period; S400, according to the precipitation intensity level to which the current actual precipitation amount belongs, combining with the probability weight, adjusting the frequency of image data preprocessing in real time.

[0005] As a preferred solution of the method for estimating precipitation intensity levels based on deep learning according to the present invention, wherein: in S100, multi-spectral remote sensing image data is collected by satellite, and four precipitation amount intervals are set according to the magnitude of the precipitation amount in the historical observation data. The four intervals correspond to four precipitation intensity levels: no rain, light rain, moderate rain, and heavy rain in ascending order from small to large.

[0006] As a preferred embodiment of the precipitation intensity level estimation method based on deep learning according to the present invention, specifically: in S200, the specific steps are as follows: S201. Perform absolute calibration on the image data, convert the original DN value into radiance or reflectance, perform atmospheric correction on the image data using an atmospheric transmission model to remove the influence of the atmosphere on radiation; perform geometric correction on the image data using georeference points, project the image data into a standard geographic coordinate system to eliminate image distortion caused by terrain undulation; perform cloud detection on the image data using a threshold segmentation method, spectral index, and deep learning algorithm, analyze the data of each channel, identify the cloud-covered area, generate a cloud mask, and mark the cloud-covered area as invalid data; S202. Perform correlation analysis on all features of the image data through the Pearson correlation coefficient calculation formula, analyze the correlation degree between the features and precipitation, and use the features that meet the preset correlation standard as key features for constructing the prediction model; S203. Combine the use of a convolutional neural network and a recurrent neural network to train the prediction model, and perform supervised learning using historical observation data and multi-spectral remote sensing image data to update the model parameters.

[0007] As a preferred embodiment of the precipitation intensity level estimation method based on deep learning according to the present invention, specifically: in S300, the specific steps are as follows: S301. Divide 24 hours into i time periods according to the brightness of the natural light source received by the target area during a day; S302. Classify all precipitation events in the historical observation data according to whether they belong to the same time period, and respectively count the distribution of each precipitation level in different time periods; S303. Calculate the occurrence frequency of each precipitation intensity level in different time periods according to the distribution, and use the occurrence frequency as the probability weight of each precipitation intensity level in different time periods.

[0008] As a preferred embodiment of the precipitation intensity level estimation method based on deep learning according to the present invention, specifically: in S400, the specific steps are as follows: S401. Set a precipitation threshold N, obtain the current precipitation X. When X < N, set the frequency of image data preprocessing to M. When X ≥ N, substitute it into the formula to calculate the new frequency of image data preprocessing: Y i =(1 + k xi )×log a (X + a)×M, where: Y i represents the frequency of image data preprocessing corresponding to the i-th time period, k xi represents the probability weight corresponding to the precipitation intensity level to which the precipitation X belongs within the i-th time period, and a is a constant and a > 1; S402. According to the actual data of the ground meteorological observation station, substitute the current actual precipitation into the above formula, and combine the precipitation intensity level to which the actual precipitation belongs to calculate and adjust the frequency of image data preprocessing in real time.

[0009] As a preferred solution of the precipitation intensity level estimation system based on deep learning according to the present invention, it includes: a data acquisition module, a precipitation prediction module, a probability weight analysis module, and a preprocessing frequency calculation module.

[0010] As a preferred solution of the precipitation intensity level estimation system based on deep learning according to the present invention, it includes: the data acquisition module includes a historical observation data acquisition unit and a satellite data acquisition unit. The historical observation data acquisition unit is used to acquire historical observation data of the ground meteorological observation station in the target area, and the satellite data acquisition unit is used to acquire multi-spectral remote sensing image data of the target area.

[0011] As a preferred solution of the precipitation intensity level estimation system based on deep learning according to the present invention, it includes: the precipitation prediction module includes a satellite data preprocessing unit, a prediction model construction unit, and a prediction model training unit. The satellite data preprocessing unit is used to preprocess the multi-spectral remote sensing image data. The preprocessing includes radiometric correction, geometric correction, and cloud detection. The prediction model construction unit is used to perform correlation analysis on all features of the image data through the Pearson correlation coefficient, analyze the degree of correlation between the features and the precipitation, and use the features that meet the preset correlation standard as key features to construct a prediction model. The prediction model training unit is used to train the prediction model by combining a convolutional neural network and a recurrent neural network, and perform supervised learning using historical observation data and multi-spectral remote sensing image data to update the model parameters. The Pearson correlation coefficient is a statistical indicator used to measure the linear relationship or degree of association between two variables. The convolutional neural network and the recurrent neural network are algorithms in the field of deep learning.

[0012] As a preferred solution of the precipitation intensity level estimation system based on deep learning according to the present invention, it includes: the probability weight analysis module is used to analyze the distribution of each precipitation intensity level in multiple time periods respectively, and calculate the probability weight of the precipitation intensity in the corresponding time period.

[0013] As a preferred solution of the precipitation intensity level estimation system based on deep learning according to the present invention, it includes: the preprocessing frequency calculation module is used to adjust the frequency of image data preprocessing in real time according to the precipitation intensity level to which the current actual precipitation belongs, in combination with the probability weight.

[0014] The beneficial effects of the present invention:

[0015] 1. By collecting high-frequency and high-resolution multi-spectral remote sensing image data through Fengyun-4 satellite, preprocessing the image data, extracting key features and analyzing the relationship between key features and precipitation intensity, constructing a precipitation intensity prediction model, training the prediction model using deep learning algorithms, and continuously optimizing the model, it is possible to achieve high-precision prediction of precipitation intensity, with a wide spatial coverage range and timely data update, and be able to capture the rapid changes in precipitation intensity in a timely manner.

[0016] 2. By analyzing the distribution of each precipitation intensity level in multiple time periods respectively, calculating the probability weight of precipitation intensity in the corresponding time period, and adjusting the frequency of image data preprocessing in real time, it is possible to achieve accurate and rapid extraction of satellite data, improve the prediction accuracy of the model, and save usage costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. The following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative and laborious efforts. Among them:

[0018] Figure 1 It is a schematic flow chart of the method for estimating precipitation intensity level based on deep learning of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the drawings in the specification.

[0020] Many specific details are set forth in the following description in order to fully understand the present invention, but the present invention can also be implemented in other ways different from those described herein. The so-called "embodiment" herein refers to specific features, structures or characteristics that can be included in at least one implementation manner of the present invention.

[0021] Embodiment 1

[0022] Refer to Figure 1, this embodiment provides a precipitation intensity level estimation method based on deep learning, which specifically includes the following steps: S100. Collect historical observation data and multi-spectral remote sensing image data of ground meteorological observation stations in the target area, and set precipitation intensity levels according to the precipitation amount; S200. Preprocess the multi-spectral remote sensing image data, extract key features, analyze the relationship between the key features and precipitation intensity, and construct and train a prediction model; S300. Analyze the distribution of each precipitation intensity level in multiple time periods respectively, and calculate the probability weight of precipitation intensity in the corresponding time period; S400. According to the precipitation intensity level to which the current true precipitation amount belongs, combined with the probability weight, adjust the frequency of image data preprocessing in real time.

[0023] In S100, multi-spectral remote sensing image data is collected by satellite. According to the size of the precipitation amount in the historical observation data, four precipitation amount intervals are set, and the four intervals correspond to four precipitation intensity levels of no rain, light rain, moderate rain, and heavy rain in ascending order from small to large.

[0024] In S200, the specific steps are as follows: S201. Perform absolute calibration on the image data, convert the original DN value into radiance or reflectance, perform atmospheric correction on the image data using an atmospheric transfer model to remove the influence of the atmosphere on radiation; perform geometric correction on the image data using georeference points, project the image data into a standard geographic coordinate system, and eliminate image deformation caused by terrain undulation; perform cloud detection on the image data using the threshold segmentation method, spectral index, and deep learning algorithm, analyze the data of each channel, identify the cloud-covered area, generate a cloud mask, and mark the cloud-covered area as invalid data; S202. Analyze the correlation of all features of the image data through the Pearson correlation coefficient calculation formula, analyze the degree of correlation between the features and the precipitation amount, and use the features that meet the preset correlation standard as key features for the construction of the prediction model; S203. Combine the use of a convolutional neural network and a recurrent neural network to train the prediction model, and perform supervised learning using historical observation data and multi-spectral remote sensing image data to update the model parameters.

[0025] Use the calibration coefficient provided by the satellite to perform absolute calibration, convert the original DN value into a physical quantity (such as radiance or reflectance), consider atmospheric correction, and use the atmospheric transfer model - 6S model to remove the influence of the atmosphere on radiation. The 6S model divides the atmosphere into multiple layers, and each layer contains different gas components (such as oxygen, ozone, water vapor, etc.) and aerosols. The model calculates the scattering and absorption of light through the concentration and distribution of these components. The model uses the radiative transfer equation to describe the transmission process of electromagnetic waves in the atmosphere. The 6S model outputs the surface reflectance, atmospheric transmittance, path radiance, and total radiation by inputting the solar altitude angle, atmospheric model, and aerosol model.

[0026] Geometric correction is performed using Ground Control Points (GCPs) to project satellite image data into a standard geographic coordinate system. Orthorectification technology is used to eliminate image distortion caused by terrain undulation.

[0027] Cloud detection is carried out using threshold segmentation, spectral indices (such as the cloud index NDSI), and deep learning algorithms (SegNet). The data of each channel is analyzed to identify cloud-covered areas, generate a cloud mask, and mark the cloud-covered areas as invalid data.

[0028] In the construction of training data, the data of a 1-hour ground meteorological observation station is used as the ground truth. For the ground truth dataset of 1-hour cumulative precipitation of precipitation, the method of data soft label rewriting is adopted. Areas without radar reflectivity are considered to have no precipitation, and other areas are set as areas where strict identification is not required, that is, other areas adopt the L1 loss function.

[0029] The L1 loss function, also known as the L1 norm loss function or absolute value loss function, is a loss function widely used in machine learning and statistics.

[0030] In S300, the specific steps are as follows: S301, divide 24 hours into i time periods according to the brightness of the natural light source received by the target area during a day; S302, classify all precipitation events in the historical observation data according to whether they belong to the same time period, and respectively count the distribution of each precipitation level in different time periods; S303, calculate the occurrence frequency of each precipitation intensity level in different time periods according to the distribution, and use the occurrence frequency as the probability weight of each precipitation intensity level in different time periods.

[0031] The time period from 0 to 8 o'clock and from 20 to 24 o'clock in a 24-hour day is taken as one time period, the time period from 8 o'clock to 16 o'clock is taken as one time period, and the time period from 16 o'clock to 20 o'clock is taken as one time period.

[0032] In S400, the specific steps are as follows: S401, set the precipitation threshold N, obtain the current precipitation X. When X < N, set the frequency of image data preprocessing to M. When X ≥ N, substitute it into the formula to calculate the new frequency of image data preprocessing: Y i =(1 + k xi )×log a (X + a)×M, where: Y i represents the frequency of image data preprocessing corresponding to the i-th time period, k xiIt represents the probability weight corresponding to the precipitation intensity level to which the precipitation amount X belongs in the i-th time period, where a is a constant and a > 1; S402. According to the actual data of the ground meteorological observation station, substitute the current actual precipitation amount into the above formula, and combine with the precipitation intensity level to which the actual precipitation amount belongs to calculate and adjust the frequency of image data preprocessing in real time.

[0033] Embodiment 2

[0034] This embodiment provides a precipitation intensity level estimation system based on deep learning, which specifically includes a data acquisition module, a precipitation amount prediction module, a probability weight analysis module, and a preprocessing frequency calculation module.

[0035] The data acquisition module includes a historical observation data acquisition unit and a satellite data acquisition unit. The historical observation data acquisition unit is used to acquire the historical observation data of the ground meteorological observation station in the target area, and the satellite data acquisition unit is used to acquire the multi-spectral remote sensing image data of the target area.

[0036] The precipitation amount prediction module includes a satellite data preprocessing unit, a prediction model construction unit, and a prediction model training unit. The satellite data preprocessing unit is used to preprocess the multi-spectral remote sensing image data, and the preprocessing includes radiometric correction, geometric correction, and cloud detection. The prediction model construction unit is used to perform correlation analysis on all features of the image data through the Pearson correlation coefficient, analyze the correlation degree between the features and the precipitation amount, and use the features that meet the preset correlation standard as key features to construct a prediction model. The prediction model training unit is used to perform model training on the prediction model by combining a convolutional neural network and a recurrent neural network, and perform supervised learning using historical observation data and multi-spectral remote sensing image data to update the model parameters. The Pearson correlation coefficient is a statistical index used to measure the linear relationship or association degree between two variables, and the convolutional neural network and the recurrent neural network are algorithms in the field of deep learning.

[0037] For the short-term heavy precipitation actual situation, select the 1-hour precipitation data of the ground meteorological observation station. The grid forecast is interpolated to the verification site by the nearest neighbor interpolation method. For the 1-hour cumulative precipitation amount of the automatic station, convert the discrete distribution data into grid data suitable for the training of satellite data to precipitation intensity level estimation.

[0038] Compare the prediction results of the model with the actual precipitation observation data, evaluate the accuracy and reliability of the model, and perform comprehensive evaluation using various evaluation indicators such as accuracy, root mean square error (RMSE), and F1 value. Through detailed analysis of the prediction results of different precipitation intensity levels, identify the advantages and disadvantages of the model's performance under specific conditions, and then perform targeted optimization.

[0039] By using dimensionality reduction techniques such as principal component analysis (PCA) on key meteorological parameters such as the spectral characteristics of clouds, cloud top temperature, cloud water content, humidity, wind speed, etc., the indicators that have the greatest impact on precipitation intensity are screened out from a large number of features. By deeply analyzing the relationship between these features and precipitation intensity, it is ensured that the selected features can accurately reflect the precipitation process.

[0040] The probability weight analysis module is used to analyze the distribution of each precipitation intensity level in multiple time periods respectively, and calculate the probability weight of precipitation intensity in the corresponding time period.

[0041] The preprocessing frequency calculation module is used to adjust the frequency of image data preprocessing in real time according to the precipitation intensity level to which the current actual precipitation belongs, in combination with the probability weight.

[0042] Importantly, although only a few embodiments are described in detail in this disclosure, those who refer to this disclosure should easily understand that many modifications are possible without substantially departing from the subject matter described in this application. For example, the dimensions, structures, shapes, and proportions of various components, as well as temperature, pressure, installation arrangements, use of materials, color, orientation changes, etc.; for example, components shown as integrally formed can be composed of multiple parts or components, and the positions of the components can be inverted or otherwise changed; therefore, all such modifications should be included within the scope of the present invention, and other substitutions, modifications, changes, and omissions can be made in the design, operating conditions, and arrangements of the exemplary embodiments without departing from the scope of the present invention.

Claims

1. A precipitation intensity level estimation method based on deep learning, characterized in that: It includes the following steps: S100. Collect the historical observation data and multi-spectral remote sensing image data of the ground meteorological observation station in the target area, and set the precipitation intensity levels according to the precipitation amount. S200. Preprocess the multi-spectral remote sensing image data, extract the key features, analyze the relationship between the key features and the precipitation intensity, and construct and train a prediction model. S300. Analyze the distribution of each precipitation intensity level in multiple time periods respectively, and calculate the probability weight of the precipitation intensity in the corresponding time period. S400. According to the precipitation intensity level to which the current actual precipitation belongs, combined with the probability weight, adjust the frequency of image data preprocessing in real time.

2. The precipitation intensity level estimation method based on deep learning according to claim 1, characterized in that: In S100, collect the multi-spectral remote sensing image data by satellite, and set four precipitation amount intervals according to the size of the precipitation amount in the historical observation data. The four intervals correspond to four precipitation intensity levels of no rain, light rain, moderate rain, and heavy rain in ascending order from small to large.

3. The precipitation intensity level estimation method based on deep learning according to claim 1, characterized in that: In S200, the specific steps are as follows: S201. Perform absolute calibration on the image data, convert the original DN value into radiance or reflectance, and perform atmospheric correction on the image data using an atmospheric transmission model to remove the influence of the atmosphere on radiation. Perform geometric correction on the image data using georeference points, project the image data into a standard geographic coordinate system, and eliminate the image deformation caused by terrain undulation. Perform cloud detection on the image data using the threshold segmentation method, spectral index, and deep learning algorithm, analyze the data of each channel, identify the cloud-covered area, generate a cloud mask, and mark the cloud-covered area as invalid data. S202. Analyze the correlation of all features of the image data through the Pearson correlation coefficient calculation formula, analyze the correlation degree between the features and the precipitation amount, and use the features that meet the preset correlation criteria as key features for constructing the prediction model. S203. Combine the use of a convolutional neural network and a recurrent neural network to train the prediction model, and perform supervised learning using the historical observation data and multi-spectral remote sensing image data to update the model parameters.

4. The precipitation intensity level estimation method based on deep learning according to claim 2, characterized in that: In S300, the specific steps are as follows: S301. Divide 24h into i time periods according to the brightness of the natural light source received by the target area during the day. S302. Classify all precipitation events in the historical observation data according to whether they belong to the same time period, and respectively count the distribution of each precipitation level in different time periods. S303. Calculate the occurrence frequency of each precipitation intensity level in different time periods according to the distribution, and use the occurrence frequency as the probability weight of each precipitation intensity level in different time periods.

5. The precipitation intensity level estimation method based on deep learning according to claim 4, characterized in that: In S400, the specific steps are as follows: S401. Set a precipitation threshold N, obtain the current precipitation amount X. When X < N, set the frequency of image data preprocessing to M. When X ≥ N, substitute it into the formula to calculate the new frequency of image data preprocessing: AND i (1+k) xi )×log a (X+a)×M Where: Y i represents the frequency of image data preprocessing corresponding to the i-th time period, k xi represents the probability weight corresponding to the precipitation intensity level to which the precipitation amount X belongs in the i-th time period, a is a constant and a>1; S402. According to the actual data of the ground meteorological observation station, substitute the current actual precipitation amount into the above formula, combined with the precipitation intensity level to which the actual precipitation belongs, calculate and adjust the frequency of image data preprocessing in real time.

6. A precipitation intensity level estimation system based on deep learning, characterized in that: It includes data acquisition module, precipitation prediction module, probability weight analysis module and preprocessing frequency calculation module.

7. The precipitation intensity level estimation system based on deep learning according to claim 6, characterized in that: The data acquisition module includes a historical observation data acquisition unit and a satellite data acquisition unit. The historical observation data acquisition unit is used to acquire historical observation data of ground meteorological observation stations in the target area, and the satellite data acquisition unit is used to acquire multi-spectral remote sensing image data of the target area.

8. The precipitation intensity level estimation system based on deep learning according to claim 6, characterized in that: The precipitation prediction module includes a satellite data preprocessing unit, a prediction model construction unit and a prediction model training unit. The satellite data preprocessing unit is used to preprocess the multispectral remote sensing image data, and the preprocessing includes radiation correction, geometric correction and cloud detection. The prediction model construction unit is used to perform correlation analysis on all features of the image data through the Pearson correlation coefficient, analyze the degree of correlation between the features and precipitation, and use the features that meet the preset correlation standards as key features to construct a prediction model. The prediction model training unit is used to train the prediction model by combining convolutional neural networks and recursive neural networks, use historical observation data and multispectral remote sensing image data for supervised learning, and update model parameters. The Pearson correlation coefficient is a statistical indicator used to measure the linear relationship or correlation between two variables. The convolutional neural network and the recursive neural network are algorithms in the field of deep learning.

9. The precipitation intensity level estimation system based on deep learning according to claim 6, characterized in that: The probability weight analysis module is used to analyze the distribution of each precipitation intensity level in multiple time periods and calculate the probability weight of the precipitation intensity in the corresponding time period.

10. The precipitation intensity level estimation system based on deep learning according to claim 6, characterized in that: The preprocessing frequency calculation module is used to adjust the frequency of image data preprocessing in real time according to the precipitation intensity level to which the current real precipitation belongs in combination with the probability weight.

Images