A method, device and system for predicting the incipient convection based on geostationary meteorological satellites

Abstract

The present invention provides a method, device and system for predicting the incipient convection based on a geostationary meteorological satellite, and relates to the field of meteorological monitoring. The present invention is based on the L1 infrared spectrum data of the Geostationary Orbital Radiation Imager (AGRI) carried by the Fengyun-4B geostationary meteorological satellite (FY‑4B), uses a neighborhood matching technology to match and track cloud targets, does not preset the critical value of the multi-spectral features of the satellite cloud image, but automatically extracts the temporal and spatial features of the incipient convection spectrum through a deep learning model based on a 3D spatiotemporal attention mechanism, thereby achieving more comprehensive and accurate incipient convection prediction, and providing important technical support for severe convective weather prediction and disaster prevention and mitigation.

Description

Technical Field

[0001] The invention relates to a method, device and system for predicting the incipient convection based on a geostationary meteorological satellite, and belongs to the technical field of meteorology. Background Art

[0002] Convective initiation is defined as the first time that the Doppler weather radar detects a reflectivity factor ≥ 35 dBZ, which marks the beginning of severe convective activity. Grasping the process of convection initiation is the basis for successfully forecasting severe convective systems and the disastrous weather they cause. Geostationary meteorological satellites have high temporal resolution and a wide imaging range, and can capture the physical characteristics of clouds over a large area and in real time. Therefore, they are widely used in the development of operational convection initiation prediction algorithms.

[0003] In the current mainstream business at home and abroad, the possibility of convective cloud development is judged based on the empirical threshold method by carefully selecting satellite spectral channels, channel combinations and time evolution laws that are sensitive to convection; 8 spectral factors are often used as the criterion for the initiation of convection in the existing technology (SATCAST algorithm for short), and the algorithm is applied to GOES business forecasts, with a forecast time of 30-45 minutes, which improves the early convection initiation warning that only uses radar data; these indicators have also been applied in some regions in China, but the disadvantage is that the cloud displacement vector needs to be accurately calculated during the indicator calculation process, which leads to high calculation cost and long time consumption. The wide threshold range leads to a high false alarm rate and poor timeliness.

[0004] In order to solve this problem, a widely used convective initiation prediction technology, namely SATCASTv2, has been developed. In the improved technology, satellite pixels are no longer used for spectral calculations, but the cloud cluster as a whole. In the SATCASTv2 algorithm, the cloud body must meet the corresponding criteria to indicate that the cloud object has sufficiently strong cloud top growth and position signals, and a convective initiation event will occur. For the same cloud object, a series of temporal predictions may change from not meeting the conditions to meeting the conditions, which indicates that the cloud top is slowly growing (i.e., cooling) at a rate very close to the critical threshold of the algorithm. However, it still has limitations in application because the characteristics of convective clouds are not fixed. For example, the thickness of the cloud body varies significantly with latitude. The thickness of convective initiation clouds in mid-latitudes can reach 5~6 km, while it is only 1~2 km in tropical regions. This will lead to large fluctuations in the reliability of the threshold-dependent method in different regions and a high false alarm rate. At the same time, due to the large temporal resolution of the geostationary satellite itself, which usually scans the entire disk in 15 minutes, and the low spatial resolution of 4KM, the same cloud target cannot be matched and tracked at adjacent times, which cannot meet the needs of actual meteorological support work. Summary of the invention

[0005] The purpose of the present invention is to provide a method, device and system for predicting the incipient convection based on a geostationary meteorological satellite. The method does not preset the critical value of the multi-spectral characteristics of the satellite cloud image, but automatically extracts the temporal and spatial characteristics of the incipient convection spectrum through a convection incipient prediction model that combines a 3D spatiotemporal attention mechanism with high-dimensional hidden features, thereby achieving a more comprehensive and accurate incipient convection prediction and providing important technical support for severe convective weather prediction and disaster prevention and mitigation.

[0006] In order to achieve the above-mentioned purpose / solve the above-mentioned technical problem, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting the incipient convection based on a geostationary meteorological satellite, comprising:

[0008] Obtain typical factor cloud diagrams of potential convective primary cloud clusters at successive moments;

[0009] The typical factor cloud images of the potential convective primary cloud cluster at consecutive moments are divided into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center, so as to construct a representation vector of the minimum image of the potential convective primary cloud cluster;

[0010] Inputting the representation vector of the minimum image of the potential convective primary cloud cluster into a pre-trained convective primary prediction model, extracting 3D spatiotemporal features from the representation vector of the minimum image of the potential convective primary cloud cluster, and processing the extracted spatiotemporal features with a multi-layer perceptron to obtain a determination result of whether the potential convective primary cloud cluster can develop into a convective primary cloud cluster;

[0011] The method of obtaining typical factor cloud diagrams of potential convective primary cloud clusters at consecutive moments specifically includes:

[0012] Obtain the L1-level spectral data cloud map of the continuous time of the geostationary orbit radiation imager carried by the FY-4B satellite;

[0013] Based on the neighborhood matching method, the cloud top brightness temperature of the long-wave infrared channel set in the L1-level spectral data cloud map at consecutive moments is combined to obtain the potential convective primary cloud clusters;

[0014] Obtain L1-level spectral data cloud maps of the water vapor channel and long-wave infrared channel of potential convective primary clouds at consecutive moments;

[0015] The typical factor cloud diagram of the potential convective primary cloud cluster at continuous moments is calculated according to the L1-level spectral data cloud diagram of the water vapor channel and the long-wave infrared channel at continuous moments.

[0016] Furthermore, the method of obtaining potential convective primary cloud clusters based on the neighborhood matching method combined with the long-wave infrared cloud top brightness temperature of a set channel in the L1-level spectral data cloud map at consecutive moments specifically includes:

[0017] The pixel values ​​of the pixel points in the L1-level spectral data cloud map at continuous moments are converted into brightness temperature values, and the pixel points whose cloud top brightness temperature of the set bandwidth long-wave infrared channel in the L1-level spectral data cloud map at continuous moments is lower than the set temperature are retained, and the L1-level spectral data cloud map at continuous moments containing several cloud clusters is constructed based on the eight-connected method according to the retained pixel points, and the cloud clusters whose number of pixel points in the L1-level spectral data cloud map at continuous moments is lower than the set number are retained as initial cloud clusters; the set temperature is 283K, and the set number is 50;

[0018] If the initial cloud cluster at the current moment intersects with the initial cloud cluster at the next moment, or the neighborhood of the initial cloud cluster at the current moment intersects with the initial cloud cluster at the next moment, or the initial cloud cluster at the current moment intersects with the neighborhood of the initial cloud cluster at the next moment, then the initial cloud cluster at the current moment and the initial cloud cluster at the next moment are the same initial cloud cluster;

[0019] If the initial cloud cluster of the same initial cloud cluster at the next moment cools down, it is determined to be a potential convective primary cloud cluster;

[0020] The neighborhood matching method avoids the situation where clouds that move faster or change shape greatly cannot be matched and tracked at adjacent moments, and tracks clouds more scientifically.

[0021] Furthermore, the shape attribute expression of the initial cloud is:

[0022] ;

[0023] in: is the shape attribute of the initial cloud, is the characteristic length of the initial cloud major axis, is the characteristic length of the initial cloud short axis, and D is the set neighborhood scale;

[0024] The method for obtaining the neighborhood of the initial cloud specifically includes:

[0025] Calculate the distance from the point outside the initial cloud to each pixel point inside the initial cloud;

[0026] Determine the distance from the point outside the initial cloud cluster to the initial cloud cluster according to the distance from the point outside the initial cloud cluster to each pixel point inside the initial cloud cluster d ;

[0027] All initial cloud groups are satisfied The area formed by the points is the neighborhood range;

[0028] The calculation formula for the distance from the point outside the initial cloud to each pixel point inside the initial cloud is:

[0029] ;

[0030] The calculation formula for determining the distance d from the point outside the initial cloud cluster to the initial cloud cluster according to the distance from the point outside the initial cloud cluster to each pixel point inside the initial cloud cluster is:

[0031] ;

[0032] in: is a point outside the initial cloud To the initial cloud Pixels The distance is the first pixels, is a point outside the initial cloud, is the characteristic direction of the long axis of the cloud, is the characteristic direction of the short axis of the cloud, is a point outside the initial cloud The distance to the initial cloud, is the number of points outside the initial cloud cluster, and the neighborhood range is all points outside the initial cloud cluster that satisfy The area formed by the points.

[0033] Furthermore, the typical factor cloud images of the potential convective primary cloud cluster at consecutive moments are divided into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center to construct a representation vector of the minimum image of the potential convective primary cloud cluster, which specifically includes:

[0034] ;

[0035] , ;

[0036] ;

[0037] in, is the potential convective primary cloud sample after segmentation, is the field of real numbers, is the radius of the potential convective primary cloud along the radial direction, is the angle of the potential convective primary cloud along the tangential direction, is the number of canonical factors, is the number of sampling moments, for dimensional real number space; for exist Location The smallest image at a moment, is the block location, It's time. is the number of blocks, is the minimum image size of the convective primary cloud, is a learnable linear parameter matrix, is a learnable positional encoding matrix, yes Location The representation vector of the minimum image of the potential convective incipient cloud at time .

[0038] Furthermore, the step of extracting 3D spatiotemporal features from the representation vector of the minimum image of the potential convective primary cloud cluster and performing multi-layer perceptron processing on the extracted spatiotemporal features specifically includes:

[0039] The spatiotemporal attention of the representation vector of the minimum image of the primary convective cloud cluster in 3D spatiotemporal feature extraction is calculated to automatically extract the temporal and spatial features of the primary convective spectrum, avoiding the problem that the actual threshold critical value of the existing empirical threshold method will be affected by the season, underlying surface, daily changes, etc., extracting more accurate cloud cluster features, improving the accuracy of model prediction, and effectively reducing the false alarm rate;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] in, , , Respectively layer Attention Head Location The query vector, key vector, and value vector at the moment; For the -1st floor The representation vector of the minimum image of the potential convective primary cloud cluster at position t, For the layer of spatiotemporal attention, For the A head of attention, , , Respectively Tier The trainable parameter matrix of the attention heads, is the layer normalization process, D h is the dimension of each attention head; For the layer Attention Head Location Time and attention, For the layer Attention Head Location The spatial attention at each moment, Softmax is the activation function; for and The key vector for the first position of For the layer Attention Head Location The key vector at time, ; For the layer Attention Head Always The key vector of the position, ;

[0046] After merging the attention of each head, we get The representation vector after layer encoding , then enter The layer performs attention calculation:

[0047] ;

[0048] in, For the layer Location The representation vector of the minimum image of the potential convective primary cloud at time , To merge the attention of each head of each spatiotemporal attention layer, It is a multi-layer perceptron, Resid is a residual connection;

[0049] The determination result calculation expression is:

[0050] ;

[0051] in, For the The representation vector of the minimum image of the potential convective primary cloud cluster at layer 0 position 0 time,

[0052] y It is the result of judging whether a potential convective primary cloud cluster can develop into a convective primary cloud cluster.

[0053] Furthermore, the training method of the convection incipient prediction model includes:

[0054] Obtain the typical factors of historical convective primary cloud clusters and the corresponding convective primary labels;

[0055] Construct a dataset based on typical factors of historical convective primary clouds and corresponding convective primary labels;

[0056] The dataset generates training sets and validation sets in time sequence;

[0057] The training set is input into the convection nascent prediction model for training. The training batch size is set to 512, the maximum number of training rounds is set to 100, the AdamW optimizer and cross entropy are used as the loss function, and the initial learning rate is set to 10 -3 , divided by 10 in the 21st, 41st, and 61st rounds until a preliminary convection inception prediction model is generated;

[0058] The preliminary convection incipient prediction model is adjusted based on the best weight with the highest preservation accuracy in the validation set to determine the final trained convection incipient prediction model.

[0059] Furthermore, the method for obtaining typical factors of historical convective primary cloud clusters specifically includes:

[0060] The L1-level spectral data of the water vapor channel and long-wave infrared channel of the geostationary radiation imager carried by the FY-4B satellite of historical convective primary clouds are used as potential influencing factors of convective primary. The importance of features is calculated through random forest, and several influencing factors with the highest importance ranking are selected as typical factors. By screening the most important influencing factors for predicting convective primary clouds in advance, the model convergence speed is accelerated and the prediction accuracy is improved.

[0061] Input the convection initiation factor X to the random forest v , the importance score of each impact factor is calculated as follows:

[0062] Impact Factor X v Importance of node q in the uth tree in a random forest for:

[0063] ;

[0064] ;

[0065] Impact Factor X v The importance of the u-th tree for:

[0066] ;

[0067] ;

[0068] in, is the Gini index of the u-th tree node q, is the proportion of category c in node q, u is the u-th decision tree, and C is the classification category;

[0069] is the Gini index of the u-th tree node o after branching, is the Gini index of the uth tree node r after branching; Q is the set of nodes where the influencing factor appears in the decision tree u;

[0070] The impact factor X v Importance rating;

[0071] U is the total number of decision trees, V is the total number of influencing factors, X v is the vth impact factor.

[0072] In a second aspect, the present invention provides a device for predicting the incipient convection based on a geostationary meteorological satellite, comprising:

[0073] Acquisition module: used to obtain typical factor cloud images of potential convective primary cloud clusters at consecutive moments;

[0074] A construction module is used to divide the typical factor cloud images of the potential convective primary cloud cluster at consecutive moments into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center, so as to construct a representation vector of the minimum image of the potential convective primary cloud cluster;

[0075] A processing module, used for inputting the representation vector of the minimum image of the potential convective primary cloud cluster into a pre-trained convective primary cloud prediction model, performing 3D spatiotemporal feature extraction on the representation vector of the minimum image of the potential convective primary cloud cluster, performing multi-layer perceptron processing on the extracted spatiotemporal features, and obtaining a determination result of whether the potential convective primary cloud cluster can develop into a convective primary cloud cluster;

[0076] The method of obtaining typical factor cloud diagrams of potential convective primary cloud clusters at consecutive moments specifically includes:

[0077] Obtain the L1-level spectral data cloud map of the continuous time of the geostationary orbit radiation imager carried by the FY-4B satellite;

[0078] Based on the neighborhood matching method, the cloud top brightness temperature of the long-wave infrared channel set in the L1-level spectral data cloud map at consecutive moments is combined to obtain the potential convective primary cloud clusters;

[0079] Obtain L1-level spectral data cloud maps of the water vapor channel and long-wave infrared channel of potential convective primary clouds at consecutive moments;

[0080] The typical factor cloud diagram of the potential convective primary cloud cluster at continuous moments is calculated according to the L1-level spectral data cloud diagram of the water vapor channel and the long-wave infrared channel at continuous moments.

[0081] In a third aspect, the present invention provides a convection incipient prediction system based on a geostationary meteorological satellite, comprising:

[0082] Memory, for storing computer programs / instructions;

[0083] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned method for predicting the incipient convection based on a geostationary meteorological satellite.

[0084] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: based on the L1 infrared spectrum data of the Geostationary Radiation Imager (AGRI) carried by the Fengyun-4B geostationary meteorological satellite (FY-4B), the present invention adopts a new neighborhood matching technology in the acquisition part of potential convective primary clouds. Compared with the current method of directly using the overlapping areas of cloud clusters, the present invention can track clouds more scientifically in the case where cloud targets that move faster or have larger shape changes cannot be matched and tracked at adjacent moments due to the low temporal and spatial resolution of geostationary meteorological satellites; in the feature extraction part of potential convective primary clouds, a method based on the 3D spatiotemporal attention mechanism is adopted to automatically extract the temporal and spatial features of the convective primary spectrum. Compared with the current empirical threshold method, this technology no longer sets a fixed critical value for the multi-spectral channels of the satellite cloud image, and can avoid the problem that the actual threshold critical value will be affected by the season, underlying surface, daily changes, etc., extract more accurate cloud features, improve the accuracy of model prediction, and effectively reduce the false alarm rate.

[0085] The convective initiation prediction method provided by the present invention can achieve more comprehensive convective initiation prediction, provide important technical support for severe convective weather prediction and disaster prevention and mitigation, and has a lower false alarm rate and higher accuracy than the existing empirical threshold method. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a schematic diagram of the process of the present invention;

[0087] Figure 2 A schematic diagram of the neighborhood matching method of the present invention;

[0088] Figure 3 Schematic diagram of spatiotemporal attention of the present invention. DETAILED DESCRIPTION

[0089] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0090] The term "and / or" is only a way to describe the association relationship of associated objects. There can be three kinds of relationships. For example, A and / or B can be: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally means that the associated objects before and after are in an "or" relationship.

[0091] It should be noted that the Geostationary Radiation Imager (AGRI) carried by FY-4B has added a water vapor detection channel compared to FY-4A, and has adjusted the spectra of some channels to improve the level of refined observation. It can currently provide a full-disk satellite image every 15 minutes. The high-frequency ground-atmosphere target observation data has a good monitoring effect on the initial convection process, and plays an important role in forecasters' prediction of weather system development and improving meteorological business service capabilities. The performance parameters of the instrument are shown in Table 1 below;

[0092] Table 1. Performance parameters of the Geostationary Radiation Imager (AGRI) carried by FY-4B

[0093]

[0094] Example 1

[0095] like Figure 1 An embodiment shown in the figure provides a method for predicting the initial convection based on a geostationary meteorological satellite, comprising:

[0096] Step 1: Training of the convection initial prediction model, including:

[0097] Step 1.1: Obtain the typical factors of historical convective incipient clouds and the corresponding convective incipient labels, and use the 9-15 infrared single channels of the geostationary orbit radiation imager carried by the FY-4B satellite of historical convective incipient clouds and the multi-channel criteria used by mainstream businesses at home and abroad (10.7µm temperature, 10.7µm temperature time trend, 6.5-10.7µm temperature difference, 13.3-10.7µm temperature difference, 6.5-10.7µm change trend, 13.3-10.7µm change trend) as potential influencing factors of convective incipient clouds, calculate the feature importance through random forest, and select the top 5 influencing factors with importance ranking greater than 0.08 as typical factors;

[0098] Input the convection initiation factor X to the random forest v, the importance score of each impact factor is calculated as follows:

[0099] Impact Factor X v Importance of node q in the uth tree in a random forest for:

[0100] ;

[0101] ;

[0102] Impact Factor X v The importance of the u-th tree for:

[0103] ;

[0104] ;

[0105] in, is the Gini index of the u-th tree node q, is the proportion of category c in node q, u is the u-th decision tree, and C is the classification category;

[0106] is the Gini index of the u-th tree node o after branching, is the Gini index of the uth tree node r after branching; Q is the set of nodes where the influencing factor appears in the decision tree u;

[0107] The impact factor X v Importance rating;

[0108] U is the total number of decision trees, V is the total number of influencing factors, X v is the vth impact factor.

[0109] After calculation, the feature importance scores are shown in Table 2. The five typical factors with the highest importance scores are retained, including: long-wave infrared channel 12 (8.55µm), indicating cloud thickness; 10.7µm temperature time trend, indicating cloud development; 6.5-10.7µm temperature difference, indicating cloud thickness; 13.3-10.7µm temperature difference, indicating cloud thickness; 13.3-10.7µm change trend, indicating cloud thickness change.

[0110] Table 2 Feature importance scores

[0111]

[0112] Step 1.2: Construct a dataset based on typical factors and corresponding convection primary labels;

[0113] Step 1.3: Generate training set and validation set according to the time sequence of the data set;

[0114] Step 1.4: Based on the typical factors selected above, perform hierarchical clustering on the extremely unbalanced convective primary samples (typical factors and corresponding convective primary labels) in the training set labels, and resample the training set;

[0115] Specifically, the similarity matrix (Euclidean distance matrix) of the initial samples of the convection is calculated, clusters with similar distances are merged, and then the similarity matrix is ​​updated. This process is iterated to divide into 6-10 clusters.

[0116] The similarity distance between sample points of different categories is calculated using Euclidean distance, and the sample points with the smallest distance value are combined. The calculation formula is as follows:

[0117] ;

[0118] in, and are the first and second coordinates of the initial sample point X of the convection, and are the first and second coordinates of the initial sample point Y of the convection, is the distance between two convection primary sample points in Euclidean space:

[0119] Calculate the silhouette coefficient of each of the 6-10 clusters, and select the cluster classification result with the highest silhouette coefficient as the final classification. The silhouette coefficient of a single sample is calculated as follows:

[0120] ;

[0121] in, is the average distance to other samples in the cluster to which a single sample belongs, is the average distance between a single sample and the nearest other cluster samples, is the average distance between a single sample and samples within the cluster;

[0122] The overall silhouette coefficient is:

[0123] ;

[0124] in, is the average value of the silhouette coefficient of all samples, and Num is the total number of samples;

[0125] Based on the median of the number of primary samples of all flows, clusters with a number lower than the median are considered minority classes, otherwise they are considered majority classes. All primary samples of minority flows are retained, and 10% of primary samples of majority flows are randomly selected to construct the resampled training set.

[0126] Step 1.5: The resampled training set is input into the convection nascent prediction model for training. The training batch size is set to 512, the maximum number of training rounds is set to 100, the AdamW optimizer and cross entropy are used as the loss function, and the initial learning rate is set to 10 -3 , divided by 10 in the 21st, 41st, and 61st rounds until a preliminary convection inception prediction model is generated;

[0127] The preliminary convection incipient prediction model is adjusted based on the best weight with the highest preservation accuracy in the validation set to determine the final trained convection incipient prediction model.

[0128] Step 2: Obtain typical factor cloud diagrams of potential convective primary cloud clusters at successive moments, including:

[0129] Step 2.1: Obtain the L1-level spectral data cloud map of the continuous time of the geostationary orbit radiation imager carried by the FY-4B satellite;

[0130] Step 2.2: Based on the neighborhood matching method, the long-wave infrared cloud top brightness temperature of the set channel in the L1-level spectral data cloud map at continuous moments is combined to obtain the potential convective primary cloud cluster, which specifically includes: converting the pixel values ​​of the pixel points in the L1-level spectral data cloud map at continuous moments into brightness temperature values, and retaining the 10.8 set in the L1-level spectral data cloud map at continuous moments. The cloud top brightness temperature of the long-wave infrared channel and the pixels below 283K (Kelvin temperature unit) are used to construct the L1-level spectral data cloud map containing several cloud clusters at continuous moments based on the eight-connectivity method according to the retained pixels, and the cloud clusters with less than 50 pixels in the L1-level spectral data cloud map at continuous moments are retained as the initial cloud clusters;

[0131] The shape attribute expression of the initial cloud is:

[0132] ;

[0133] in: is the shape attribute of the initial cloud, is the characteristic length of the initial cloud major axis, is the characteristic length of the initial cloud short axis, and D is the set neighborhood scale;

[0134] like Figure 2 As shown, the method for obtaining the neighborhood of the initial cloud group specifically includes:

[0135] Calculate the distance from the point outside the initial cloud to each pixel point inside the initial cloud;

[0136] Determine the distance from the point outside the initial cloud cluster to the initial cloud cluster according to the distance from the point outside the initial cloud cluster to each pixel point inside the initial cloud cluster d ;

[0137] All initial cloud groups are satisfied The area formed by the points is the neighborhood range;

[0138] The calculation formula for the distance from the point outside the initial cloud to each pixel point inside the initial cloud is:

[0139] ;

[0140] The calculation formula for determining the distance d from the point outside the initial cloud cluster to the initial cloud cluster according to the distance from the point outside the initial cloud cluster to each pixel point inside the initial cloud cluster is:

[0141] ;

[0142] in: is a point outside the initial cloud To the initial cloud Pixels The distance is the first pixels, is a point outside the initial cloud, is the characteristic direction of the long axis of the cloud, is the characteristic direction of the short axis of the cloud, is a point outside the initial cloud The distance to the initial cloud, is the number of points outside the initial cloud cluster, and the neighborhood range is all points outside the initial cloud cluster that satisfy The area formed by the points;

[0143] When the cloud cluster size is smaller than D, the neighborhood range extends outward uniformly; when the cloud cluster size is larger than D and is cluster-shaped, the neighborhood range extends outward uniformly; when the cloud cluster size is larger than D and is band-shaped, the neighborhood range is longer in the long axis direction, up to twice D. Considering that a pixel point of FY-4B / AGRI is 4km, D is set to 8km here;

[0144] If the initial cloud cluster A at the current moment intersects with the initial cloud cluster B at the next moment, or the neighborhood of the initial cloud cluster A at the current moment intersects with the initial cloud cluster B at the next moment, or the initial cloud cluster A at the current moment intersects with the neighborhood of the initial cloud cluster B at the next moment, then the initial cloud cluster A at the current moment and the initial cloud cluster B at the next moment are the same initial cloud cluster; if the initial cloud cluster at the next moment of the same initial cloud cluster experiences a drop in temperature, it is determined to be a potential convective primary cloud cluster.

[0145] The L1-level spectral data cloud map of continuous moments containing several cloud clusters based on the eight-connectivity method is specifically as follows:

[0146] Select pixel M whose cloud top brightness temperature is lower than 283K in the 10.8µm long-wave infrared channel, and traverse the 8 adjacent pixels around it. If one or more of these 8 pixels meet the condition that the brightness temperature is lower than 283K, then pixel M and the adjacent pixels that meet the condition belong to the same eight-connected cloud cluster. Until the adjacent pixels around any pixel in the eight-connected range do not meet the condition, the pixels within the same eight-connected range are judged to be the same cloud cluster.

[0147] Step 2.3: Obtain the L1-level spectral data cloud map of the water vapor channel and the long-wave infrared channel (channels 9-15) of the potential primary convective cloud cluster at continuous moments; calculate the typical factor cloud map of the potential primary convective cloud cluster at continuous moments based on the L1-level spectral data cloud map of the water vapor channel and the long-wave infrared channel (channels 9-15) at continuous moments.

[0148] Step 3: The typical factor cloud images of the potential convective primary cloud cluster at consecutive moments are divided into blocks along the radial direction and the tangential direction with the potential convective primary cloud cluster as the center, and a representation vector of the minimum image of the potential convective primary cloud cluster is constructed, which specifically includes:

[0149] ;

[0150] , ;

[0151] ;

[0152] in, is the potential convective primary cloud sample after segmentation, R is the real number domain, is the radius of the potential convective primary cloud along the radial direction, is the angle of the potential convective primary cloud along the tangential direction, is the number of canonical factors, is the number of sampling moments, for dimensional real number space;

[0153] for exist The minimum image at position t, is the block position, t is the time, N is the number of blocks, is the minimum image size of the convective primary cloud, is a learnable linear parameter matrix, is a learnable positional encoding matrix, yes The representation vector of the minimum image of the potential convective incipient cloud at position t.

[0154] Step 4: Input the representation vector of the minimum image of the potential primary convective cloud cluster into the pre-trained primary convective cloud prediction model for prediction processing, perform 3D spatiotemporal feature extraction and feature vector random fusion processing on the representation vector of the minimum image of the potential primary convective cloud cluster, and obtain the judgment result of whether the potential primary convective cloud cluster can develop into the primary convective cloud cluster, which specifically includes:

[0155] Step 4.1: Figure 3 As shown, the spatiotemporal attention of the representation vector of the minimum image of the nascent cloud cluster in the 3D spatiotemporal feature extraction is calculated:

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] ;

[0161] in, , , Respectively layer Attention Head The query vector, key vector, and value vector at position t; For the -1st floor The representation vector of the minimum image of the potential convective primary cloud cluster at position t, It is layer of spatiotemporal attention, For the A head of attention, , , Respectively Tier The trainable parameter matrix of the attention heads, is the layer normalization process, D h is the dimension of each attention head; It's time attention. is spatial attention, Softmax is the activation function; for and The key vector for the first position of For the layer Attention Head Location The key vector at time, ; For the layer Attention Head Always The key vector of the position, ;

[0162] After merging the attention of each head, we get The representation vector after layer encoding , then enter +1 layer for attention calculation:

[0163] ;

[0164] in, For the layer The representation vector of position at time t, To merge the attention of each head of each spatiotemporal attention layer, It is a multi-layer perceptron, Resid is a residual connection;

[0165] The determination result calculation expression:

[0166] ;

[0167] in, For the The representation vector of the minimum image of the potential convective primary cloud cluster at layer 0 position 0 time, y It is the result of judging whether a potential convective primary cloud cluster can develop into a convective primary cloud cluster.

[0168] In order to verify the effectiveness of the model method in this invention, three evaluation indicators, threat score (TS), hit rate (POD), and false alarm rate (FAR), are used to evaluate the model method, as follows:

[0169] ;

[0170] ;

[0171] ;

[0172] in , , Represents hit, missed alarm, and false alarm respectively.

[0173] The results based on the above model on the test set are shown in Table 3 below:

[0174] Table 3 Comparison of convection inception prediction model results

[0175]

[0176] Compared with the traditional empirical threshold method, the present invention significantly improves TS and reduces FAR, which shows that the machine learning algorithm has significant advantages over the threshold algorithm in the prediction of convection initiation. This is because the threshold method relies on critical values, but these predetermined values ​​are highly sensitive to environmental factors, such as cross-regional, seasonal and instrument changes. The present invention overcomes the limitations of the threshold method; specifically, the TS and FAR of the threshold method are 0.214 and 0.774, respectively, and the present invention improves them by 59.8% and 38.5%. The present invention proposes a convection initiation prediction method, which has higher accuracy than the traditional empirical threshold method and can provide key technical support for the forecast of convection initiation and severe convective weather.

[0177] Example 2

[0178] This embodiment provides a convection initiation prediction device based on a geostationary meteorological satellite, comprising:

[0179] Acquisition module: used to obtain typical factor cloud images of potential convective primary cloud clusters at consecutive moments;

[0180] A construction module is used to divide the typical factor cloud images of the potential convective primary cloud cluster at consecutive moments into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center, so as to construct a representation vector of the minimum image of the potential convective primary cloud cluster;

[0181] A processing module, used for inputting the representation vector of the minimum image of the potential convective primary cloud cluster into a pre-trained convective primary prediction model for prediction processing, and obtaining a determination result of whether the potential convective primary cloud cluster can develop into a convective primary cloud cluster;

[0182] The step of inputting the representation vector of the minimum image of the potential primary convective cloud cluster into a pre-trained convective primary prediction model for prediction processing includes: extracting 3D spatiotemporal features of the representation vector of the minimum image of the potential primary convective cloud cluster and performing random fusion processing of feature vectors to obtain a determination result;

[0183] The method of obtaining typical factor cloud diagrams of potential convective primary cloud clusters at consecutive moments specifically includes:

[0184] Obtain the L1-level spectral data cloud map of the continuous time of the geostationary orbit radiation imager carried by the FY-4B satellite;

[0185] Based on the neighborhood matching method, the cloud top brightness temperature of the long-wave infrared channel set in the L1-level spectral data cloud map at consecutive moments is combined to obtain the potential convective primary cloud clusters;

[0186] Obtain L1-level spectral data cloud maps of the water vapor channel and long-wave infrared channel of potential convective primary clouds at consecutive moments;

[0187] The typical factor cloud diagram of the potential convective primary cloud cluster at continuous moments is calculated according to the L1-level spectral data cloud diagram of the water vapor channel and the long-wave infrared channel at continuous moments.

[0188] Example 3

[0189] This embodiment provides a convection incipient prediction system based on a geostationary meteorological satellite, comprising:

[0190] Memory, for storing computer programs / instructions;

[0191] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned method for predicting the incipient convection based on a geostationary meteorological satellite.

[0192] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0193] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0194] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0195] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0196] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A method for predicting the incipient convection based on a geostationary meteorological satellite, characterized in that: include: Obtain typical factor cloud diagrams of potential convective primary cloud clusters at successive moments; The typical factor cloud images of the potential convective primary cloud cluster at consecutive moments are divided into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center, so as to construct a representation vector of the minimum image of the potential convective primary cloud cluster; The representation vector of the minimum image of the potential convective primary cloud cluster is input into a pre-trained convective primary prediction model, 3D spatiotemporal features are extracted from the representation vector of the minimum image of the potential convective primary cloud cluster, and the extracted spatiotemporal features are processed by a multi-layer perceptron to obtain a determination result of whether the potential convective primary cloud cluster can develop into a convective primary cloud cluster. Wherein: the acquisition of typical factor cloud diagrams of potential convective primary cloud clusters at consecutive moments specifically includes: Obtain the L1-level spectral data cloud map of the continuous time of the geostationary orbit radiation imager carried by the FY-4B satellite; Based on the neighborhood matching method, the cloud top brightness temperature of the long-wave infrared channel set in the L1-level spectral data cloud map at consecutive moments is combined to obtain the potential convective primary cloud clusters; Obtain L1-level spectral data cloud maps of the water vapor channel and long-wave infrared channel of potential convective primary clouds at consecutive moments; The typical factor cloud diagram of the potential convective primary cloud cluster at continuous moments is calculated according to the L1-level spectral data cloud diagram of the water vapor channel and the long-wave infrared channel at continuous moments.

2. The method for predicting the incipient convection based on a geostationary meteorological satellite according to claim 1, characterized in that: The method of obtaining potential convective primary cloud clusters based on the neighborhood matching method combined with the long-wave infrared cloud top brightness temperature of a set channel in the L1-level spectral data cloud map at consecutive moments specifically includes: The pixel values ​​of the pixels in the L1-level spectral data cloud map at continuous moments are converted into brightness temperature values, and the pixels whose cloud top brightness temperature of the set bandwidth long-wave infrared channel in the L1-level spectral data cloud map at continuous moments is lower than the set temperature are retained. The L1-level spectral data cloud map at continuous moments containing several cloud clusters is constructed based on the eight-connected method according to the retained pixels, and the cloud clusters whose number of pixels in the L1-level spectral data cloud map at continuous moments is lower than the set number are retained as initial cloud clusters; If the initial cloud cluster at the current moment intersects with the initial cloud cluster at the next moment, or the neighborhood of the initial cloud cluster at the current moment intersects with the initial cloud cluster at the next moment, or the initial cloud cluster at the current moment intersects with the neighborhood of the initial cloud cluster at the next moment, then the initial cloud cluster at the current moment and the initial cloud cluster at the next moment are the same initial cloud cluster; If the initial cloud cluster of the same initial cloud cluster cools down at the next moment, it is judged to be a potential convective primary cloud cluster.

3. The method for predicting the incipient convection based on a geostationary meteorological satellite according to claim 2, characterized in that: The shape attribute expression of the initial cloud is: ； in: is the shape attribute of the initial cloud, is the characteristic length of the initial cloud major axis, is the characteristic length of the initial cloud short axis, and D is the set neighborhood scale; The method for obtaining the neighborhood of the initial cloud specifically includes: Calculate the distance from the point outside the initial cloud to each pixel point inside the initial cloud; Determine the distance from the point outside the initial cloud cluster to the initial cloud cluster according to the distance from the point outside the initial cloud cluster to each pixel point inside the initial cloud cluster d ; All initial cloud groups are satisfied The area formed by the points is the neighborhood range; The calculation formula for the distance from the point outside the initial cloud to each pixel point inside the initial cloud is: ； The calculation formula for determining the distance d from the point outside the initial cloud cluster to the initial cloud cluster according to the distance from the point outside the initial cloud cluster to each pixel point inside the initial cloud cluster is: ； in: is a point outside the initial cloud To the initial cloud Pixels The distance is the first pixels, is a point outside the initial cloud, is the characteristic direction of the long axis of the cloud, is the characteristic direction of the short axis of the cloud, is a point outside the initial cloud The distance to the initial cloud, is the number of points outside the initial cloud cluster, and the neighborhood range is all points outside the initial cloud cluster that satisfy The area formed by the points.

4. The method for predicting the incipient convection based on a geostationary meteorological satellite according to claim 1, characterized in that: The typical factor cloud images of the potential convective primary cloud cluster at consecutive moments are divided into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center to construct a representation vector of the minimum image of the potential convective primary cloud cluster specifically including: ； ； ； ； ； ； in, is the potential convective primary cloud sample after segmentation, is the field of real numbers, is the radius of the potential convective primary cloud along the radial direction, is the angle of the potential convective primary cloud along the tangential direction, is the number of canonical factors, is the number of sampling moments, for dimensional real number space; for exist Location The smallest image at a moment, is the block location, It's time. is the number of blocks, is the minimum image size of the convective primary cloud, is a learnable linear parameter matrix, is a learnable positional encoding matrix, yes Location The representation vector of the minimum image of the potential convective incipient cloud at time .

5. The method for predicting the incipient convection based on a geostationary meteorological satellite according to claim 4, characterized in that: The step of extracting 3D spatiotemporal features from the representation vector of the minimum image of the potential convective primary cloud cluster and performing multi-layer perceptron processing on the extracted spatiotemporal features specifically includes: Calculate the spatiotemporal attention of the representation vector of the minimum image of the nascent cloud in 3D spatiotemporal feature extraction: ； ； ； ； ； in, , , Respectively layer Attention Head Location The query vector, key vector, and value vector at the moment; For the -1st floor The representation vector of the minimum image of the potential convective primary cloud cluster at position t, For the layer of spatiotemporal attention, For the A head of attention, , , Respectively Tier The trainable parameter matrix of the attention heads, is the layer normalization process, D h is the dimension of each attention head; For the layer Attention Head Location Time and attention, For the layer Attention Head Location The spatial attention at each moment, Softmax is the activation function; for and The key vector for the first position of For the layer Attention Head Location The key vector at time, ; For the layer Attention Head Always The key vector of the position, ; After merging the attention of each head, we get The representation vector after layer encoding , then enter The layer performs attention calculation: ； in, For the layer Location The representation vector of the minimum image of the potential convective primary cloud at time , To merge the attention of each head of each spatiotemporal attention layer, It is a multi-layer perceptron, Resid is a residual connection; The determination result calculation expression: ； in, For the The representation vector of the minimum image of the potential convective primary cloud cluster at layer 0 position 0 time, y It is the result of judging whether a potential convective primary cloud cluster can develop into a convective primary cloud cluster.

6. The method for predicting the incipient convection based on a geostationary meteorological satellite according to claim 1, characterized in that: The training method of the convection incipient prediction model comprises: Obtain the typical factors of historical convective primary cloud clusters and the corresponding convective primary labels; Construct a dataset based on typical factors of historical convective primary clouds and corresponding convective primary labels; The dataset generates training sets and validation sets in time sequence; The training set is input into the convection nascent prediction model for training. The training batch size is set to 512, the maximum number of training rounds is set to 100, the AdamW optimizer and cross entropy are used as the loss function, and the initial learning rate is set to 10 -3 , divided by 10 in the 21st, 41st, and 61st rounds until a preliminary convection inception prediction model is generated; The preliminary convection incipient prediction model is adjusted based on the best weight with the highest preservation accuracy in the validation set to determine the final trained convection incipient prediction model.

7. The method for predicting the incipient convection based on a geostationary meteorological satellite according to claim 6, characterized in that: The method for obtaining typical factors of historical convective primary cloud clusters specifically includes: The L1-level spectral data of the water vapor channel and long-wave infrared channel of the geostationary radiation imager carried by the FY-4B satellite of the historical convective primary cloud clusters are used as potential influencing factors of convective primary formation. The feature importance is calculated by random forest, and several influencing factors with the highest importance ranking are selected as typical factors. Impact Factor X v Importance of node q in the uth tree in a random forest for: ； ； Impact Factor X v The importance of the u-th tree for: ； ； in, The impact factor X v Importance rating; is the Gini index of the u-th tree node q, is the proportion of category c in node q, u is the u-th decision tree, and C is the classification category; is the Gini index of the u-th tree node o after branching, is the Gini index of the uth tree node r after branching; Q is the set of nodes where the influencing factor appears in the decision tree u; U is the total number of decision trees, V is the total number of influencing factors, X v is the vth impact factor.

8. A device for predicting the onset of convection based on a geostationary meteorological satellite, characterized in that: include: Acquisition module: used to obtain typical factor cloud images of potential convective primary cloud clusters at consecutive moments; A construction module is used to divide the typical factor cloud images of the potential convective primary cloud cluster at consecutive moments into blocks along radial and tangential directions with the potential convective primary cloud cluster as the center, so as to construct a representation vector of the minimum image of the potential convective primary cloud cluster; A processing module, used for inputting the representation vector of the minimum image of the potential convective primary cloud cluster into a pre-trained convective primary cloud prediction model, performing 3D spatiotemporal feature extraction on the representation vector of the minimum image of the potential convective primary cloud cluster, performing multi-layer perceptron processing on the extracted spatiotemporal features, and obtaining a determination result of whether the potential convective primary cloud cluster can develop into a convective primary cloud cluster; Wherein: the acquisition of typical factor cloud diagrams of potential convective primary cloud clusters at consecutive moments specifically includes: Obtain the L1-level spectral data cloud map of the continuous time of the geostationary orbit radiation imager carried by the FY-4B satellite; Based on the neighborhood matching method, the cloud top brightness temperature of the long-wave infrared channel set in the L1-level spectral data cloud map at consecutive moments is combined to obtain the potential convective primary cloud clusters; Obtain L1-level spectral data cloud maps of the water vapor channel and long-wave infrared channel of potential convective primary clouds at consecutive moments; The typical factor cloud diagram of the potential convective primary cloud cluster at continuous moments is calculated according to the L1-level spectral data cloud diagram of the water vapor channel and the long-wave infrared channel at continuous moments.

9. A convection incipient prediction system based on a geostationary meteorological satellite, characterized in that: include: Memory, for storing computer programs / instructions; A processor, configured to execute the computer program / instructions to implement the steps of the method for predicting the incipient convection based on a geostationary meteorological satellite according to any one of claims 1 to 7.

Images