CN115691049A - Convection birth early warning method based on deep learning - Google Patents
Convection birth early warning method based on deep learning Download PDFInfo
- Publication number
- CN115691049A CN115691049A CN202211076007.2A CN202211076007A CN115691049A CN 115691049 A CN115691049 A CN 115691049A CN 202211076007 A CN202211076007 A CN 202211076007A CN 115691049 A CN115691049 A CN 115691049A
- Authority
- CN
- China
- Prior art keywords
- convection
- remote sensing
- time
- image
- resolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 238000013135 deep learning Methods 0.000 title claims abstract description 28
- 238000012549 training Methods 0.000 claims abstract description 20
- 230000008859 change Effects 0.000 claims description 11
- 238000011156 evaluation Methods 0.000 claims description 8
- 238000012360 testing method Methods 0.000 claims description 6
- 238000007689 inspection Methods 0.000 claims description 5
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- IYLGZMTXKJYONK-ACLXAEORSA-N (12s,15r)-15-hydroxy-11,16-dioxo-15,20-dihydrosenecionan-12-yl acetate Chemical compound O1C(=O)[C@](CC)(O)C[C@@H](C)[C@](C)(OC(C)=O)C(=O)OCC2=CCN3[C@H]2[C@H]1CC3 IYLGZMTXKJYONK-ACLXAEORSA-N 0.000 claims description 2
- 230000015556 catabolic process Effects 0.000 claims description 2
- 238000006731 degradation reaction Methods 0.000 claims description 2
- IYLGZMTXKJYONK-UHFFFAOYSA-N ruwenine Natural products O1C(=O)C(CC)(O)CC(C)C(C)(OC(C)=O)C(=O)OCC2=CCN3C2C1CC3 IYLGZMTXKJYONK-UHFFFAOYSA-N 0.000 claims description 2
- 238000012544 monitoring process Methods 0.000 abstract description 15
- 238000013508 migration Methods 0.000 description 7
- 230000005012 migration Effects 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 6
- 238000011161 development Methods 0.000 description 5
- 230000003287 optical effect Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 230000009467 reduction Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000013213 extrapolation Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000002265 prevention Effects 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 230000002457 bidirectional effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007123 defense Effects 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000001556 precipitation Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000002310 reflectometry Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000012271 agricultural production Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000003750 conditioning effect Effects 0.000 description 1
- 230000010485 coping Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000036039 immunity Effects 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Image Processing (AREA)
Abstract
The invention relates to a convection birth early warning method based on deep learning, which comprises the following steps: establishing a depth super-resolution model to extract effective information in remote sensing images with different image resolutions and time resolutions, and training the remote sensing images to reconstruct the image resolution and the time resolution of the remote sensing images; acquiring a reconstructed remote sensing image, and performing initial identification on the convection cloud by adopting a cloud top bright temperature threshold method according to a preset convection cloud identification strategy; extracting convection clouds by using a computer vision algorithm based on an area overlapping method to perform convection cloud tracking; judging and identifying the final convection cloud remote sensing image based on a GOES-R algorithm, and judging and identifying convection onset according to a preset convection onset judgment condition; and the satellite time-space descending scale result is checked and evaluated by combining the satellite channel image with high time-space precision, and the convection nascent inversion result is checked by combining the radar data. The invention can realize strong convection weather monitoring and early warning and improve the judgment precision of convection birth.
Description
Technical Field
The invention relates to the technical field of strong convection weather early warning, in particular to a convection birth early warning method based on deep learning.
Background
The strong convection weather early warning is the first line of defense for disaster defense and is also the most important ring. However, the current countries are not comprehensive enough to grasp the weather information with strong convection. In actual business, radar reflectivity (a linear extrapolation method based on radar data) is generally applied to distinguish strong convection, but no radar covers plateau, desert, sea and the like, so that accurate prediction of time and place is difficult to make.
Because the weather system has the characteristics of complex development and rapid change, the processes of accurately judging, completing tracking and the like become weak links. With the continuous understanding of the generation conditions of the strong convection weather, the convection inception is an important development stage of the strong convection, and the convection inception refers to the first occurrence of an echo greater than 35dBZ on the weather radar comprehensive reflectivity image. Researches find that the strong convection weather can be accurately early warned by monitoring the convection nascent process, and the early warning can be early warned.
Currently, there are 4 major international convection generation algorithms, RDT (Rapid development Foundation), forTracC (monitoring and Tracking of Evolution of Cloud Cluster), GOES-R (The geographic operation Environment software R-Series Program) and UWCI (The University of Wisconsin Conditioning Initiation). In the process of convection identification, the RDT algorithm focuses more on peak detection of a vertical form, and velocity extrapolation is used for tracking convection, and the service algorithm of the RDT algorithm is combined with various data such as a numerical prediction product and a satellite inversion product; the ForTraCC algorithm pays more attention to active convection, comprehensively considers the situations of convection combination and splitting, and can realize extrapolation prediction; the GOES-R algorithm monitors convection cloud by using a multispectral identification technology, and mainly aims at a stationary meteorological satellite; the UWCI algorithm provides a block averaging (box averaging) concept, and uses a cloud type product and a single-channel image as auxiliary data, so that the technology is simple and fast.
At present, in the research aiming at the strong convection early warning, the convection nascent products developed by the China meteorological satellite center have hysteresis (usually 15-20 minutes) due to undisclosed algorithm, cannot be served in actual business, and also need to complete a great deal of work such as threshold value optimization, inspection and verification. And a radar echo extrapolation method based on a deep neural network is jointly developed by a central weather station and Qinghua university, the forecasting accuracy is improved by about 40%, but the strong convection early warning requirements of high-quality areas without radar on the sea cannot be met due to uneven radar distribution. Meanwhile, the traditional algorithm only uses a near-infrared channel, and fine early warning is difficult to realize due to the fact that the image resolution of the near-infrared channel is low.
Disclosure of Invention
In view of the above drawbacks of the prior art, an object of the present invention is to provide a method for early warning of convection onset based on deep learning, which is used to solve the problems of late early warning of convection onset and low accuracy in the prior art.
In order to achieve the above objects and other related objects, the present invention provides the following technical solutions:
a convection birth early warning method based on deep learning comprises the following steps:
establishing a depth super-resolution model to extract effective information in remote sensing images with different image resolutions and time resolutions, and training the remote sensing images to reconstruct the image resolution and the time resolution of the remote sensing images;
acquiring a reconstructed remote sensing image, and performing initial judgment of the convective cloud by adopting a cloud top brightness temperature threshold value method according to a preset convective cloud judgment strategy;
extracting convection clouds by using a computer vision algorithm based on an area overlapping method to perform convection cloud tracking;
judging and identifying the final convection cloud based on a GOES-R algorithm, and judging and identifying convection birth according to a preset convection birth judgment condition; and the number of the first and second groups,
and (3) checking and evaluating the satellite time-space descending scale result by combining the satellite channel image with high time-space precision, and checking the convection nascent inversion result by combining radar data.
As a preferred aspect of the present invention, the establishing of the depth super-resolution model to extract effective information in remote sensing images with different image resolutions and time resolutions specifically includes:
establishing a deep super-resolution model by utilizing the nonlinear mapping capability of deep learning and the information extraction capability of the raster data;
extracting effective information in remote sensing images with different image resolutions and time resolutions by using a depth super-resolution model; and the number of the first and second groups,
and reconstructing the space-time low-resolution remote sensing image into a space-time high-resolution remote sensing image by using the acquired effective information.
As a preferred aspect of the present invention, the training of the remote sensing image to reconstruct the image resolution and the time resolution of the remote sensing image specifically includes:
training remote sensing images with different image resolutions by adopting an ESRGAN model to reconstruct the remote sensing images into remote sensing images with high image resolution, wherein the remote sensing images with high image resolution retain channel information of the image resolution of the original remote sensing images; and the number of the first and second groups,
and training the remote sensing images with different time resolutions by adopting a Super SloMo model so as to reconstruct the remote sensing images into remote sensing images with high time resolutions, wherein the time resolutions of the remote sensing images with high time resolutions and radar data are kept consistent.
As a preferred scheme of the invention, the predetermined convection cloud identification policy specifically includes: the coldest temperature threshold of the convection cloud reaches-55 ℃, the warmest temperature threshold of the convection cloud reaches-10-5 ℃, the temperature isoline interval of the convection cloud in the vertical direction reaches 1 ℃, the extreme value of the convection cloud reaches 3 ℃, and the minimum area threshold of the convection judgment reaches one infrared pixel size.
As a preferred scheme of the invention, when the cloud top brightness temperature threshold method is adopted for initial identification of convection cloud, the RDT algorithm is utilized, and the adaptive threshold is adopted to select the long-wave infrared channel threshold.
As a preferred aspect of the invention, the preset convection initial determination condition is specifically: the brightness temperature of 10.7 mu m is less than or equal to 0 ℃, and the time change rate of the brightness temperature of 10.7 mu m is less than or equal to [ -4 ℃ (15 min) -1 [)]Bright temperature difference of-35 deg.C < 6.5 and 10.7 μm < -10 deg.C, -bright temperature difference of-25 deg.C < 13.3 and 10.7 μm < -5 deg.C, 6.5 and 10.7 μm bright temperature time change rate>[3℃(15min) -1 ]And the time change rates of the brightness temperature of 13.3 and 10.7 μm>[3℃(15min) -1 ]。
As a preferable scheme of the invention, the checking and evaluating the satellite time-space down scale result by combining the high-space-time-precision satellite channel image specifically comprises:
evaluating the accuracy of the time-space downscaling scale by adopting the peak signal-to-noise ratio and the sufficient similarity parameter; wherein,
the peak signal-to-noise ratio
Wherein,
a maximum value representing the color of an image point;
similarity of the structures
Wherein mu x 、μ y Is a mean value, σ x 、σ y Is the standard deviation, σ xy Is covariance, c 1 ＝(0.01L) 2 ，c 2 ＝(0.03L) 2 L takes a value of 255 for an 8bit gray scale map.
As a preferable aspect of the invention, the checking and evaluating the satellite time-space down scale result by combining the high-time-space-precision satellite channel image further includes:
judging the precision of the prediction result by adopting the root-mean-square error and the spatial correlation coefficient; wherein,
the root mean square error
The spatial correlation coefficient
Wherein n is the number of pixels, O i And S i For the observed value and the predicted value,
and
are the respective mean values.
As a preferable scheme of the invention, after the inspection and evaluation of the satellite time-space down scale result is performed by combining the high-space-time-precision satellite channel image, the method further comprises the following steps:
and optimizing parameters and threshold values of the flow initial judgment according to the test evaluation result.
As described above, the convection birth early warning method based on deep learning of the present invention has the following beneficial effects:
the method adopts a deep learning technology to perform space-time downscaling on the wind and cloud satellite images, wherein the space downscaling is to downscale the low resolution of the images into high resolution in a space range, improve the resolution capability of ground feature information and reflect more detailed ground information, and the time downscaling is to downscale the images with coarse time resolution into the images with finer time resolution. The wind and cloud satellite is mainly used in the fields of weather monitoring and forecasting, disaster prevention and reduction, climate change, ecological environment monitoring, service construction with one road and the like, so that the spatial resolution and the time resolution of the wind and cloud satellite are improved through deep learning, and more refined data can be provided for subsequent rainfall forecasting, weather monitoring and the like; the problem of mismatching of spatial and temporal resolutions of multi-source data is solved, and the method has profound significance for weather mode prediction and high-precision precipitation service with uniform spatial and temporal resolutions; and the space-time downscaling method can obtain a remote sensing image with higher space-time resolution, can improve the continuous and rapid monitoring capability of the medium and small-scale cloud cluster, and has important significance for monitoring and early warning of strong convection weather, especially for monitoring sudden medium and small-scale weather systems.
The invention adopts a space-time downscaling mode to perform data fusion, realizes effective utilization of the cloud picture detail information, solves the problems of small-scale missing report and large-scale delay, and greatly improves the accuracy of the forecast. Wherein the space scale is determined to be reduced from 4km to 1km, and the time scale is determined to be reduced from 15min to 6min. Meanwhile, in the actual convection nascent inversion process, setting of a plurality of meteorological parameters and thresholds is involved, such as identification of convection cloud, and judgment is carried out through an empirical value (241K) of cloud top temperature; meanwhile, the invention combines various cloud cluster tracking technologies such as an area overlapping method, a cross correlation method, computer vision and the like, and has unique advantages in the aspect of processing linear motion target tracking.
Drawings
Fig. 1 is a flowchart illustrating a method according to a first embodiment of the present invention.
Fig. 2 is a schematic diagram of an ESRGAN model according to an embodiment of the present invention.
Fig. 3 is a schematic structural diagram of a Super SloMo model in the first embodiment of the present invention.
Fig. 4 shows a cloud cluster identification map obtained by applying the RDT algorithm in the first embodiment of the present invention.
FIG. 5a shows a 4km × 4km low resolution image according to a second embodiment of the present invention.
Fig. 5b shows a high resolution image obtained by bicubic interpolation in the second embodiment of the present invention.
FIG. 5c is a high-resolution image based on the ESRGAN model using pre-training weights according to the second embodiment of the present invention.
Fig. 5d shows the high resolution image migrated and learned based on the ESRGAN model according to the second embodiment of the present invention.
FIG. 5e shows an actual high resolution image of 1km × 1km according to the second embodiment of the present invention.
Fig. 6a shows an image based on Super SloMo model prediction in the third embodiment of the present invention.
Fig. 6b shows a real remote sensing image in the third embodiment of the present invention.
Fig. 7 shows an image of the early warning of convection onset in the fourth embodiment of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention.
In some of the flows described in the present specification and claims and in the above figures, a number of operations are included that occur in a particular order, but it should be clearly understood that these operations may be performed out of order or in parallel as they occur herein, with the order of the operations being indicated as 101, 102, etc. merely to distinguish between the various operations, and the order of the operations by themselves does not represent any order of performance. Additionally, the flows may include more or fewer operations, and the operations may be performed sequentially or in parallel. It should be noted that, the descriptions of "first", "second", etc. in this document are used for distinguishing different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit the types of "first" and "second".
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example one
Referring to fig. 1, the present invention provides a convection birth early warning method based on deep learning, which includes:
s100, establishing a depth super-resolution model to extract effective information in remote sensing images with different image resolutions and time resolutions, and training the remote sensing images to reconstruct the image resolution and the time resolution of the remote sensing images.
Specifically, establishing a super-resolution depth model to extract effective information in remote sensing images with different image resolutions and time resolutions specifically includes:
s101, establishing a deep super-resolution model by utilizing the nonlinear mapping capability of deep learning and the information extraction capability of raster data;
s102, extracting effective information in remote sensing images with different image resolutions and time resolutions by using a depth super-resolution model; and the number of the first and second groups,
and S103, reconstructing the space-time low-resolution remote sensing image into a space-time high-resolution remote sensing image by using the acquired effective information.
Training a remote-sensing image to reconstruct an image resolution and a time resolution of the remote-sensing image specifically includes:
s104, training the remote sensing images with different image resolutions by adopting an ESRGAN model to reconstruct the remote sensing images into remote sensing images with high image resolution, wherein the remote sensing images with high image resolution retain channel information of the image resolution of the original remote sensing images.
As shown in fig. 2, fig. 2 shows a specific network model structure diagram of the ESRGAN model. In this embodiment, an ESRGAN generator is used to realize Super-Resolution reconstruction of an image, and an ESRGAN (Enhanced Super-Resolution generated adaptive Networks) model is a Super-Resolution generation countermeasure network, and can generate a realistic texture during Super-Resolution of a single image. The SRGAN improvement is mainly embodied in the following three aspects: (1) use of immunity loss and perception loss; (2) The method introduces Residual-in-Residu Dense Block (RRDB), combines a plurality of layers of Residual error networks and Dense connection, and is easier to train; (3) Using the VGG feature before activation to optimize the perceptual loss may provide greater supervision over brightness uniformity and texture recovery.
S105, training the remote sensing images with different time resolutions by adopting a Super SloMo model to reconstruct the remote sensing images into remote sensing images with high time resolution, wherein the time resolution of the remote sensing images with high time resolution is consistent with that of radar data, and in the embodiment, the time resolution of the remote sensing images is reconstructed from 15min to 6min by adopting the Super SloMo model.
As shown in fig. 3, fig. 3 shows a schematic structural diagram of a Super SloMo model. The Super SloMo model is an end-to-end convolutional neural network, can generate any number of intermediate video frames between two input images, and the whole frame of the Super SloMo model is composed of two full convolutional neural networks U-Net. Firstly, calculating bidirectional optical flow between adjacent input images by using a U-Net; linearly fitting the optical flow on each time step to further approximate the bidirectional optical flow of the intermediate frame, and optimizing the approximate optical flow by using another U-Net to solve the problem that the motion boundary has artifacts; finally, the two input images are warped and linearly fused to form an intermediate frame. In addition, the optical flow calculation network and the interpolation network of the Super SloMo model have parameters independent of the specific time step of the interpolated frame (the time step is the input of the network), so that the frame can be interpolated at any time step between two frames in parallel, thereby breaking through the limitation of a plurality of single-frame interpolation methods.
S200, acquiring the reconstructed remote sensing image, and performing initial identification on the convection cloud by adopting a cloud top bright temperature threshold method according to a preset convection cloud identification strategy.
Specifically, when a cloud top bright temperature threshold method is adopted for initial identification of convection cloud, an RDT algorithm is utilized, the height change of a convection cloud body in the vertical direction is considered, a long-wave infrared channel threshold value is selected by adopting a self-adaptive threshold value, and the RDT algorithm is developed by European meteorological satellite development organization EUMETSATSAF; the preset convection cloud identification strategy specifically comprises the following steps: the coldest temperature threshold of the convection cloud reaches-55 ℃, the warmest temperature threshold of the convection cloud reaches-10-5 ℃, the temperature contour line interval of the convection cloud in the vertical direction reaches 1 ℃, the extreme value of the convection cloud reaches 3 ℃, and the minimum area threshold of the convection judgment reaches the size of one infrared pixel, and the convection cloud can be judged if the conditions are met. As shown in fig. 4, fig. 4 shows a cloud cluster identification graph obtained by applying the RDT algorithm.
S300, extracting convection clouds to track the convection clouds by using a computer vision algorithm based on an area overlapping method, wherein the computer vision algorithm specifically adopts a Boosting algorithm, overlapping area values are selected, relevant values are considered, the convection clouds can be automatically extracted from a series of images by using a Boosting algorithm, analysis and understanding are carried out, and tracking of the convection clouds is achieved.
S400, judging the final convection cloud remote sensing image based on a GOES-R algorithm, and judging the convection onset according to a preset convection onset judgment condition.
In this embodiment, a specific table of the convection inception decision conditions is shown below by referring specifically to the criterion adopted by the CI algorithm of the current GOES satellite service and using 3 channel data of a GOES-R ABI (Advanced base Imager) instrument:
and S500, checking and evaluating the satellite time-space descending scale result by combining the satellite channel image with high time-space precision, and checking the convection nascent inversion result by combining radar data.
Specifically, the inspection and evaluation of the satellite time-space degradation scale result by combining the high-space-time-precision satellite channel image specifically comprises the following steps:
s501, evaluating the accuracy of a space-time reduction scale by adopting a Peak Signal to Noise Ratio (PSNR) and a Structural Similarity (SSIM) parameter; wherein,
the peak signal-to-noise ratio
Wherein,
the maximum value representing the color of the image point is 255 if each sampling point is represented by 8 bits;
similarity of the structures
Wherein mu x 、μ y Is a mean value, σ x 、σ y Is the standard deviation, σ xy Is covariance, c 1 ＝(0.01L) 2 ，c 2 ＝(0.03L) 2 L takes a value of 255 for an 8bit gray scale map.
Generally, a higher PSNR value indicates a closer reconstructed image to an original image. In order to evaluate the image quality more accurately, the image quality is evaluated by using Structural Similarity (SSIM), the structural similarity is the structural similarity degree of the super-resolution image and the actual high-resolution image, the value is not more than 1, and the closer the value is to 1, the more similar the structures of the super-resolution image and the actual high-resolution image are, the better the image reconstruction effect is.
S502, judging the precision of a prediction result by adopting Root Mean Square Error (RMSE) and a spatial Correlation Coefficient (CC); wherein,
the root mean square error
The spatial correlation coefficient
Wherein n is the number of pixels, O i And S i For the observed value and the predicted value,
and
are the respective mean values.
RMSE is the deviation between a predicted value and an actually measured value, and the smaller the RMSE value is, the smaller the deviation is, and the higher the accuracy of the representative model is; the closer the CC values are to 1, the higher their correlation.
After the satellite time-space down scale result is tested and evaluated by combining the satellite channel image with high time-space precision, the method further comprises the following steps:
and S503, optimizing parameters and threshold values of the flow initial judgment according to the test evaluation result.
The invention uses four scoring standards to evaluate the forecasting ability of the convection nascent inversion result, namely a fair forecasting rate (ETS), a Frequency deviation Frequency Bias, a hit rate (probability of detection, POD) and a False Alarm Rate (FAR), wherein the ETS is the rate of correctly forecasting the event of eliminating random factors, the Frequency deviation Frequency Bias is the rate of evaluating the event of forecasting the occurrence and the event of actual occurrence (more than 1 represents the Frequency of overestimating the event occurrence, less than 1 represents the underestimation), the hit rate is the rate of correctly forecasting the event of evaluating the occurrence, the FAR is the rate of evaluating the event which does not occur but is the best forecast, the POD is 1, 0, and the change range is-1/3-1, 0-infinity, and the following formulas are calculated as follows:
wherein, H represents the number of grid points correctly predicted, M represents the number of missed report grid points, F represents the number of empty report grid points, and C represents the number of grid points correctly predicted without convection nascent event, as shown in the following table.
On the basis of a traditional convection nascent inversion algorithm, algorithm research and development and improvement are carried out based on a domestic wind and cloud satellite, in order to meet the fine requirement of disaster early warning, a wind and cloud satellite data part is processed in a key mode, a deep learning technology is adopted to carry out space-time downscaling on a wind and cloud satellite image, the space downscaling is to downscale the low resolution of the image into high resolution in a space range, the resolution capability of ground feature information is improved, more detailed ground information is reflected, and the time downscaling is to downscale the image with coarse time resolution into an image with finer time resolution. The wind and cloud satellite is mainly used in the fields of weather monitoring and forecasting, disaster prevention and reduction, climate change, ecological environment monitoring and the like, so that the spatial resolution and the time resolution of the wind and cloud satellite are improved through deep learning, and more refined data can be provided for subsequent rainfall forecasting, weather monitoring and the like; the problem of mismatching of spatial-temporal resolutions of multi-source data is solved, and the method has profound significance in weather mode prediction and high-precision precipitation service with uniform spatial-temporal resolutions; and the space-time downscaling obtains a remote sensing image with higher space-time resolution, so that the continuous and rapid monitoring capability of medium and small-scale clouds can be improved, and the method has important significance for monitoring and early warning of strong-convection weather, especially for monitoring sudden medium and small-scale weather systems.
The invention adopts a space-time downscaling mode to perform data fusion, realizes effective utilization of the cloud picture detail information, solves the problems of small-scale missing report and large-scale delay, and greatly improves the accuracy of the forecast. Wherein, the space scale is determined to be reduced from 4km to 1km, and the time scale is determined to be reduced from 15min to 6min. Meanwhile, in the actual convection nascent inversion process, setting of a plurality of meteorological parameters and thresholds is involved, such as identification of convection cloud, and judgment is carried out through an empirical value (241K) of cloud top temperature; meanwhile, the invention combines various cloud cluster tracking technologies such as an area overlapping method, a cross correlation method, computer vision and the like, and has unique advantages in the aspect of processing linear motion target tracking.
The invention realizes a convection primary early warning system with high space-time precision, and meanwhile, the data used by the invention is domestic wind and cloud satellites, so that the early warning capability in the aspects of aviation, meteorology, agriculture and the like is improved.
1. The aviation direction is as follows: the method provides a convection nascent algorithm for an airline company, realizes generation of convection range and intensity information according to real-time satellite data, and provides meteorological guarantee for navigation of an airplane. Specifically, for aviation products real-time Turbulence report (WSI Total Turbulence) in The standard American Weather Company (The Weather Company), a convection-first warning product is integrated on The air.
According to statistics, the economic loss of China caused by flight delay is as high as 500 hundred million yuan, and 2/3 of loss can be avoided as long as weather conditions can be mastered and applied correctly. Therefore, the occurrence of strong convection weather is accurately monitored and forecasted, and the aviation business is guided to operate by using the strong convection weather forecasting product, so that the economic and life losses caused by weather problems can be reduced.
The method provides the weather guarantee for the flight of China airlines, particularly international flight, emphatically by providing strong convection information, and realizes convenient and efficient weather service. And by providing key weather information which is helpful for customers to control cost, enhance flight safety and improve efficiency, and combining China satellite, radar and ground station data, weather service is provided for national and even global aviation flight.
2. Weather direction: the linkage of all departments strengthens the construction of a service platform and a disaster prevention system, and properly performs disaster reduction work, which is one of the main tasks of the meteorological department. The purpose of the disastrous weather early warning based on deep learning is to solve problems from the source, early warn the development of a rainfall system earlier than a ground radar, send early warning of strong convection weather and technically improve early warning capability. The method has outstanding application effect in the aspect of meteorological service and generates good benefit.
3. In other aspects: early warning in the aspects of agriculture, logistics, shipping and the like is facilitated, logistics and shipping companies can master the early warning level of the disaster weather in time, the disaster weather coping capability is improved, measures are taken efficiently, and loss and risk are reduced to the minimum. In addition, the successful early warning of strong convection can effectively reduce the industrial and agricultural production loss caused by the weather, and ensure the life and property safety of people.
Example two
The embodiment discloses a space downscaling example, which specifically comprises the following steps: 4 pieces of image data of 7 months, 16 days and 2018 years are selected as test data (the size is 1200 multiplied by 784), the same test set is tested by different methods, and PSNR, SSIM, RMSE and CC indexes are calculated. The spatial resolution of a source data wind cloud 4A image is 4km multiplied by 4km, the target data is the bicubic interpolation of 1km multiplied by 1km resolution, and the precision index evaluation of a pre-training weight method and a migration method model based on an ESRGAN model is compared and shown in the following table:
according to the comparison result, the pre-training weight method and the migration method are respectively used based on the ESRGAN model, and the pre-training weight method and the migration method have better space scale reduction effects on PSNR, SSIM and EMSE indexes than a bicubic interpolation method, which shows that the deep learning-based method has the capability of extracting historical space correlation. Compared with a pre-training weight method, the migration method based on the ESRGAN model is improved in PSNR indexes, SSIM indexes, RMSE indexes and CC indexes by 0.827 indexes, 0.009 indexes, 0.002 indexes and 0.001 indexes, and the migration method effectively learns the spatial characteristics of the Fengyun No. 4A remote sensing images.
Fig. 5a to 5e show the selected spatial downscaling effect of 19 point images at 8 points of 16 days in 7 months and 16 months in 2018, where fig. 5a shows a low-resolution image of 4km × 4km, fig. 5b shows a high-resolution image obtained by bicubic interpolation, fig. 5c shows a high-resolution image obtained by using pre-training weights based on an ESRGAN model, fig. 5d shows a high-resolution image obtained by migration learning based on the ESRGAN model, and fig. 5e shows an actual high-resolution image of 1km × 1 km. It is obvious from fig. 5a to 5e that compared with the bicubic interpolation method, the remote sensing image with higher resolution can be more accurately reconstructed by the deep learning-based method under the scene of 4 times resolution improvement, and the high resolution image obtained by using the migration method based on the ESRGAN model has clear edges and overall effect closer to that of the real high resolution image.
EXAMPLE III
The embodiment discloses a time downscaling example, which specifically comprises: according to the images of 8 points, 15 points and 8 points, 30 points of 16 days, 7 months and 8 months of 2018A images, time step lengths are set, a model is input, and any image in the time period is tested. Fig. 6a and fig. 6b are 8-point 23-point images predicted by deep learning, wherein fig. 6a is (a. Image predicted based on Super SloMo model, fig. 6b is real remote sensing image; accuracy calculation results of the Super SloMo model prediction and actual image are shown in the following table:
as can be seen from the above table, the RMSE calculated by the remote sensing image subjected to the pre-training test based on the Super SloMo model and the real remote sensing image is 0.211, the CC is 0.9815, the PSNR is 33.508, and the SSIM is 0.7815, and the overall effect is closer to the real high-resolution image, the motion direction of the cloud is kept consistent, and the time resolution is changed from 15min to 6min.
Example four
The embodiment is described by taking a specific example as an example, and provides a Beijing convective birth warning example study, as shown in FIG. 7. The rainstorm intensity of Beijing 2018.7.16 is next to the "6.23 rainstorm" in 2011, which is an extreme rainstorm event in the Beijing city and lasts about 58 hours, and the rainstorm caused by strong convection occurs at 7.15 and reaches the maximum at 7.16. In the embodiment, the satellite data is used for real-time detection, when the convection cloud is identified at 11. Therefore, it is determined that convection is nascent at the place, i.e., strong convection is about to occur. And through the detection of radar images, the radar monitors strong convection only at 12.
It should be understood that, although the steps in the flowcharts of the figures are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and may be performed in other orders unless otherwise indicated herein. Moreover, at least a portion of the steps in the flow chart of the figure may include multiple sub-steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of execution is not necessarily sequential, but may be performed alternately or alternately with other steps or at least a portion of the sub-steps or stages of other steps.
Claims (9)
1. A convection birth early warning method based on deep learning is characterized by comprising the following steps:
establishing a depth super-resolution model to extract effective information in remote sensing images with different image resolutions and time resolutions, and training the remote sensing images to reconstruct the image resolution and the time resolution of the remote sensing images;
acquiring a reconstructed remote sensing image, and performing initial identification on the convection cloud by adopting a cloud top bright temperature threshold method according to a preset convection cloud identification strategy;
extracting convection clouds by using a computer vision algorithm based on an area overlapping method to perform convection cloud tracking;
judging and identifying the final convection cloud based on a GOES-R algorithm, and judging and identifying convection birth according to a preset convection birth judgment condition; and the number of the first and second groups,
and (3) the satellite time-space descending scale result is checked and evaluated by combining the satellite channel image with high time-space precision, and the convection nascent inversion result is checked by combining radar data.
2. The convection birth early warning method based on deep learning according to claim 1, wherein the establishing of the deep super-resolution model to extract effective information in remote sensing images with different image resolutions and time resolutions specifically comprises:
establishing a deep super-resolution model by utilizing the nonlinear mapping capability of deep learning and the information extraction capability of the raster data;
extracting effective information in remote sensing images with different image resolutions and time resolutions by using a depth super-resolution model; and the number of the first and second groups,
and reconstructing the space-time low-resolution remote sensing image into a space-time high-resolution remote sensing image by using the acquired effective information.
3. The convection birth early warning method based on deep learning according to claim 1 or 2, wherein the training of the remote sensing image to reconstruct the image resolution and the time resolution of the remote sensing image specifically comprises:
training remote sensing images with different image resolutions by adopting an ESRGAN model to reconstruct the remote sensing images into remote sensing images with high image resolution, wherein the remote sensing images with high image resolution retain channel information of the image resolution of the original remote sensing images; and the number of the first and second groups,
and training the remote sensing images with different time resolutions by adopting a Super SloMo model so as to reconstruct the remote sensing images into remote sensing images with high time resolutions, wherein the time resolutions of the remote sensing images with high time resolutions and radar data are kept consistent.
4. The convection birth early warning method based on deep learning according to claim 1, wherein the predetermined convection cloud identification strategy is specifically: the coldest temperature threshold of the convection cloud reaches-55 ℃, the warmest temperature threshold of the convection cloud reaches-10-5 ℃, the temperature isoline interval of the convection cloud in the vertical direction reaches 1 ℃, the extreme value of the convection cloud reaches 3 ℃, and the minimum area threshold of the convection judgment reaches one infrared pixel size.
5. The convection birth early warning method based on deep learning as claimed in claim 1, wherein an RDT algorithm is used when the cloud top brightness temperature threshold method is used for the initial judgment of the convection cloud, and a long-wave infrared channel threshold is selected by using an adaptive threshold.
6. The deep learning-based convection birth early warning method according to claim 1, wherein the preset convection birth decision condition is specifically: the brightness temperature of 10.7 mu m is less than or equal to 0 ℃, and the time change rate of the brightness temperature of 10.7 mu m is less than or equal to [ -4 ℃ (15 min) -1 [)]Brightness temperature difference of-35 deg.C < 6.5 and 10.7 μm < -10 deg.C, -25 deg.C < 13.3 and 10.7 μm < -5 deg.C, 6.5 and 10.7 μm brightness temperature time change rate > [3 deg.C (15 min) ] -1 ]And the time change rates of the brightness temperature of 13.3 and 10.7 μm>[3℃(15min) -1 ]。
7. The deep learning-based convective birth early warning method according to claim 1, wherein the inspection and evaluation of the satellite time-space down scale result by combining the high time-space precision satellite channel image specifically comprises:
evaluating the accuracy of the time-space downscaling by adopting a peak signal-to-noise ratio and a structural similarity parameter; wherein,
the peak signal-to-noise ratio
Wherein,
a maximum value representing the color of an image point;
similarity of the structures
Wherein mu x 、μ y Is the mean value, σ x 、σ y Is the standard deviation, σ xy Is covariance, c 1 ＝(0.01L)，c 2 = 0.03L, L taking 255 for an 8bit greyscale map.
8. The deep learning-based convective birth early warning method according to claim 7, wherein the verifying and evaluating the satellite time-space descent scale result by combining the high-time-space-precision satellite channel image further comprises:
judging the precision of the prediction result by adopting the root-mean-square error and the spatial correlation coefficient; wherein,
the root mean square error
The spatial correlation coefficient
Wherein n is the number of pixels, O i And S i In order to obtain the observed value and the predicted value,
and
are the respective mean values.
9. The deep learning-based convective birth early warning method according to claim 8, wherein after the inspection and evaluation of the satellite time-space degradation scale result are performed by combining the satellite channel image with high time-space precision, the method further comprises:
and optimizing parameters and threshold values of the flow initial judgment according to the test evaluation result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211076007.2A CN115691049A (en) | 2022-09-02 | 2022-09-02 | Convection birth early warning method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211076007.2A CN115691049A (en) | 2022-09-02 | 2022-09-02 | Convection birth early warning method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115691049A true CN115691049A (en) | 2023-02-03 |
Family
ID=85060796
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211076007.2A Pending CN115691049A (en) | 2022-09-02 | 2022-09-02 | Convection birth early warning method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115691049A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117518299A (en) * | 2024-01-05 | 2024-02-06 | 南京大学 | Classified strong convection proximity probability forecasting method, system, equipment and terminal |
| CN117592002A (en) * | 2024-01-18 | 2024-02-23 | 国家卫星气象中心(国家空间天气监测预警中心) | Primary convection identification method and device |
-
2022
- 2022-09-02 CN CN202211076007.2A patent/CN115691049A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117518299A (en) * | 2024-01-05 | 2024-02-06 | 南京大学 | Classified strong convection proximity probability forecasting method, system, equipment and terminal |
| CN117518299B (en) * | 2024-01-05 | 2024-03-22 | 南京大学 | Classified strong convection proximity probability forecasting method, system, equipment and terminal |
| CN117592002A (en) * | 2024-01-18 | 2024-02-23 | 国家卫星气象中心(国家空间天气监测预警中心) | Primary convection identification method and device |
| CN117592002B (en) * | 2024-01-18 | 2024-04-26 | 国家卫星气象中心(国家空间天气监测预警中心) | Primary convection identification method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20220043182A1 (en) | Spatial autocorrelation machine learning-based downscaling method and system of satellite precipitation data | |
| CN115691049A (en) | Convection birth early warning method based on deep learning | |
| KR102006847B1 (en) | System and Method for radar based nowcasting using optical flow with a multi scale strategy | |
| CN107728234B (en) | Thunder and lightning strength value prediction method based on atmospheric electric field data | |
| CN109814175B (en) | Strong convection monitoring method based on satellite and application thereof | |
| CN111983732B (en) | Rainfall intensity estimation method based on deep learning | |
| CN108646319B (en) | Short-time strong rainfall forecasting method and system | |
| CN113496104A (en) | Rainfall forecast correction method and system based on deep learning | |
| Liang et al. | A composite approach of radar echo extrapolation based on TREC vectors in combination with model-predicted winds | |
| CN110942111B (en) | Method and device for identifying strong convection cloud cluster | |
| CN113469278B (en) | Strong weather target identification method based on deep convolutional neural network | |
| CN109407177B (en) | Machine learning and conventional meteorological observation-based fog identification system and application method | |
| CN113298295B (en) | Meteorological forecast system for power production | |
| CN108254750B (en) | Down-blast intelligent identification early warning method based on radar data | |
| CN115907574A (en) | Drainage basin type rainstorm flood disaster-bearing body resetting cost remote sensing simulation method | |
| CN114048944A (en) | Estimation method for people to be evacuated and houses to be damaged under rainstorm induced geological disaster | |
| CN116384733A (en) | Power transmission line risk early warning method based on weather radar in strong convection weather | |
| CN114325879A (en) | Quantitative precipitation correction method based on classification probability | |
| CN114742206A (en) | Rainfall intensity estimation method for comprehensive multi-space-time scale Doppler radar data | |
| CN117078045B (en) | Holographic inspection and monitoring method and system for traffic infrastructure | |
| CN114386654A (en) | Multi-scale numerical weather forecasting mode fusion weather forecasting method and device | |
| CN113821895A (en) | Construction method and device of power transmission line icing thickness prediction model and storage medium | |
| CN117710508A (en) | Near-surface temperature inversion method and device for generating countermeasure network based on improved condition | |
| CN117849902A (en) | Method for forecasting and releasing precipitation through space-time double-correction high-resolution sub-seasonal grid set | |
| Zuo et al. | Determination of bus crowding coefficient based on passenger flow forecasting |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |