A Multi-Layer Cloud Detection Method Based on a 3D Lookup Table

Abstract

This invention discloses a multi-layer cloud detection method based on a 3D lookup table. The method constructs and trains a 3D lookup table in the model training phase, and constructs a confidence score based on the confidence parameters retrieved from the trained 3D lookup table in the model invocation phase. The final confidence score determines the probability of a multi-layer cloud. The model training phase includes a data feature creation stage, a feature combination stage, a 3D lookup table creation stage, and an optimal segmentation threshold determination stage. The model invocation phase includes a 3D lookup table reading stage, a threshold segmentation stage, and a weight combination stage. This invention solves the problems of low accuracy and poor linearity between the final confidence score and detection accuracy in existing technologies. The method of this invention can be used 24 / 7, both day and night. Detection results show good linearity between the final confidence score and detection accuracy, with high accuracy at a final confidence score of 1, and a maximum detection accuracy of approximately 90%.

Description

Technical Field

[0001] This invention belongs to the field of meteorological science, specifically relating to a multi-layer cloud detection method based on a three-dimensional lookup table. Background Technology

[0002] Accurate detection of multi-layered clouds is crucial. For example, when using radiometric imager data to invert cloud properties, it's often assumed that each observed pixel is uniform. However, this assumption can lead to decreased accuracy when multi-layered cloud structures exist. Particularly when the upper layer of a multi-layered cloud is a transmissive cloud while the lower layer is a non-transmissive cloud, using CO2 slicing to derive cloud top pressure can produce inaccurate results, typically resulting in an underestimation of the cloud top pressure. Furthermore, the presence of multi-layered cloud structures can also cause errors in cloud physical property inversion. For instance, in scenarios with an upper layer of ice clouds and a lower layer of water clouds, failing to detect the lower water cloud and misclassifying it as a single-layer ice cloud will result in a significantly smaller effective radius inverted for the ice cloud. Conversely, when the optical thickness of the ice cloud is very thin and the rising radiation of the lower water cloud is not significantly affected, misclassifying it as a single-layer water cloud often leads to unrealistically large water droplet radii in cloud inversion results. Additionally, failing to consider the multi-layered cloud structure can also lead to incorrect identification of the cloud facies at the cloud top. Therefore, some research is based on the first step of correctly determining whether a cloud layer is a multi-layered cloud structure, for example, by simultaneously inverting the optical thickness and effective radius of the upper ice cloud and the lower liquid water cloud. This requires correctly determining the existence of a multi-layered cloud structure.

[0003] A multi-layer cloud detection method was designed for MODIS. Existing literature 1 (Joiner J, Vasilkov AP, Bhartia P K, et al. Detection of multi-layer and vertically-extended clouds using A-train sensors[J]. Atmospheric Measurement Techniques, 2010, 3(1):233-247) developed a multi-layer cloud detection method for MODIS Volume 5. This method uses the detection results of multi-layer clouds as an important quality indicator of cloud inversion products. It uses confidence levels of 2, 3, 4, 5, 6, 7, 8, and 9 to determine whether a cloud is multi-layered; higher confidence levels indicate a higher likelihood of multi-layer clouds, while lower confidence levels indicate a higher likelihood of single-layer clouds. However, the validation of this method is limited by the simulation data using the forward radiative transfer model, and there is no good linear relationship between confidence level and detection accuracy. Reference 2 (Baum BA, Menzel WP, Frey RA, et al. MODIS cloud-top property refinements for collection 6[J]. Journal of Applied Meteorology and Climatology, 2012, 51(6): 1145-1163) discloses a cloud phase difference detection method, which combines the results of infrared cloud phase detection method and short-wave infrared combined with optical characteristics cloud phase detection method. When there is a difference between the two phases, it is determined that there are multiple clouds; Reference 3 (Marchant B, Platnick S, Meyer K, et al. MODIS Collection 6 shortwave-derived cloud phase classification algorithm and comparisons with CALIOP[J]. Atmospheric Measurement Techniques, 2016, 9(4):1587-1599) discloses a method for testing the difference in precipitable cloud cover. This method uses carbon dioxide slicing and cloud top pressure to calculate the precipitable cloud cover value, and compares it with the difference in total precipitation obtained from the 0.94 μm channel. When the difference is greater than 8%, it is determined that multiple clouds exist. Another method, using cloud parameters set at a cloud height of 900 MPa, again applies the method from reference 3 to determine multiple clouds.The method also employs the PH method proposed in reference 4 (Pavolonis MJ, Heidinger AK. Daytime cloud overlap detection from AVHRR and VIIRS[J]. Journal of Applied Meteorology and Climatology, 2004, 43(5):762-778). This method utilizes the observed brightness temperatures of the 0.65μm, 1.6μm, 1.38μm, 11μm, and 12μm channels, as well as the brightness temperature difference data between the 11μm and 12μm channels. To aid in discrimination, the brightness temperature of the 13.6μm channel is introduced as an additional constraint, and two sets of discriminant equations are set based on the 30-degree latitude boundary. The constraints of the PH method help MODIS 6 correctly identify thicker single-layer clouds in the eye of a typhoon. The aforementioned references 1-4 have the following major flaws: the secondary products used may have errors, indirectly weakening the effectiveness of the detection results; they were not designed for multi-layered cloud structures with ice clouds on the upper and lower layers, or water clouds on the upper and lower layers, leading to missed detections of multi-layered clouds; because the visible light channel (0.65μm) and near-infrared channel (0.94μm) observations are used as auxiliary judgments, the multi-layered cloud detection method is only applicable during the day and cannot be used at night; the linearity between the final confidence level and the detection accuracy is poor. When the final confidence level is 1, the accuracy of detecting multi-layered clouds is 55%, while when the final confidence level is 0.85, the accuracy of detecting multi-layered clouds is 68%. Therefore, the relative probability of multi-layered clouds cannot be quickly determined based on the calculated final confidence level. For multi-layer cloud detection using AGRI, reference 5 (Yu Z, Ma S, Han D, et al. A cloud classification method based on random forest for FY-4A[J]. International Journal of Remote Sensing, 2021, 42(9): 3353-3379) used observation data from all channels of FY-4A and inverted cloud top optical thickness, effective cloud radius, and cloud height data. Combined with cloud type labels from 2B-CLDCLASS matched with AGRI, this was used as the overall dataset, which was then divided into training and test sets using stratified sampling. Using AGRI data from September to December 2017, and through experimental comparison, the results showed that for multi-layer cloud detection, random forest demonstrated better performance than K-nearest neighbor and backpropagation neural networks, achieving a maximum multi-layer cloud detection rate of 50.12%.Furthermore, the method in Reference 5 also requires the use of secondary cloud products, which themselves have product errors. Using these products as input may further reduce the accuracy of multi-layer cloud detection. Moreover, since this discrimination method requires waiting for the aforementioned secondary cloud products to be prepared before further determining whether multi-layer cloud exists, Reference 6 has a slow detection speed, low detection accuracy, and its multi-layer cloud detection results are not effectively applied. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-layer cloud detection method based on a three-dimensional lookup table, which solves the problems of low accuracy and poor linearity between the final confidence level and the detection accuracy in existing technologies. This invention uses cloud layer number information detected by orthogonally polarized cloud aerosol lidar as labels for single-layer and multi-layer clouds. A three-dimensional lookup table is created, and the confidence parameters of nodes in the three-dimensional lookup table are used to obtain the judgment weights for whether a cloud is multi-layered. The confidence level is obtained through a weight combination stage, and the final confidence level is obtained through normalization.

[0005] To achieve the above objectives, this invention provides a multi-layer cloud detection method based on a three-dimensional lookup table. This method constructs and creates a three-dimensional lookup table during model training, and constructs a confidence score by looking up confidence parameters based on the created three-dimensional lookup table during model invocation. The final confidence score is then used to determine the likelihood of a multi-layer cloud. The method includes the following steps:

[0006] (S100) The model training part includes the data feature creation stage, the feature combination stage, the 3D lookup table creation stage, and the optimal segmentation threshold determination stage, as detailed below:

[0007] (S110) The data feature creation stage includes:

[0008] The mid-wave infrared and long-wave infrared channels in the training dataset are used as data features, or the mid-wave infrared and long-wave infrared channels, the channel difference of mid-wave infrared and the channel difference of long-wave infrared in the training dataset are used as data features.

[0009] (S120) The feature combination stage includes: using permutations and combinations to obtain data feature combinations from the data features obtained in the feature creation stage, and randomly selecting 3 features from the total number of data features as the three dimensions in each data feature combination, as follows:

[0010]

[0011] In equation (1), N1 is the total number of data features, and K is the total number of combinations of data features;

[0012] (S130) The stage of creating a three-dimensional lookup table includes the following steps:

[0013] (S131) ​​Construct a three-dimensional lookup table

[0014] The K data feature combinations obtained in the feature combination stage are used to construct three-dimensional lookup tables. The three dimensions of each data feature combination are used as the three dimensions of the corresponding three-dimensional lookup table, as follows:

[0015] For the k-th 3D lookup table in [1,K], the j-th dimension in [1,3] is divided into its three dimensions using nodes based on the distribution range of the observations in the j-th dimension. This yields the resolution in the j-th dimension. The relationship between the resolution and the distribution range of the observations is as follows:

[0016]

[0017] In equation (2), r j For the resolution of the j-th dimension, S j Let n be the distribution range of the observations in the j-th dimension. j Let be the number of nodes in the j-th dimension.

[0018] After determining the resolution in the j-th dimension, for any combination of data features, the corresponding three-dimensional lookup table can be created using the resolution in the j-th dimension and the number of nodes in the corresponding three-dimensional lookup table.

[0019] (S132) Calculate the parameters of the three-dimensional lookup table

[0020] Assume there are N in the training dataset a observation point P o It is represented as:

[0021] P = {P} o}, o∈[0,N a (3)

[0022] In equation (3), o is the o-th sample point, and P is the total number of observation points. o A set;

[0023] Let P be any node in the k-th three-dimensional lookup table. t Let the coordinates be (x, y, z), where x is the first dimension, y is the second dimension, and z is the third dimension. Let P be any observation point in the training dataset. o The coordinates are (a1, b1, c1), where a1 is the observation value in the first dimension, b1 is the observation value in the second dimension, and c1 is the observation value in the third dimension. The coordinates are used to determine the observation point P. o Does it belong to node P? t Three conditions must be met simultaneously, as follows:

[0024]

[0025]

[0026]

[0027] In equations (4), (5) and (6), r1 is the resolution of the first dimension in the k-th three-dimensional lookup table; r2 is the resolution of the second dimension in the k-th three-dimensional lookup table; and r3 is the resolution of the third dimension in the k-th three-dimensional lookup table.

[0028] If the observation points P in the training dataset o If all three conditions (4), (5), and (6) are met, then the observation point P... o It belongs to node P in the three-dimensional lookup table. t of;

[0029] Suppose there are N in the training dataset. t observation point P o It belongs to node P t In N t observation point P o The CPC has N t1 observation point P o It belongs to multi-layered clouds; if N t If the value is 0, then the node P t There is no corresponding observation point P in the training dataset. o Set η t It is an empty value; if N t If not equal to 0, then according to N t With N t1 Calculate η t :

[0030]

[0031] In equation (7), η t For the k-th node P in the lookup table t Confidence parameter;

[0032] By calculating step by step using the above formula (7), all nodes P in any k-th three-dimensional lookup table can be obtained. t Confidence parameter;

[0033] (S140) The stage of determining the optimal segmentation threshold includes the following steps:

[0034] (S141) Set a threshold, and determine each node P in the k-th three-dimensional lookup table based on the set threshold. t Does it belong to a multi-layered cloud?

[0035] Take a threshold T in the k-th 3D lookup table k This threshold and each node P in any k-th three-dimensional lookup table obtained during the three-dimensional lookup table creation stage. t η t The relationship is as follows:

[0036] T k ≤η t ≤1 (8)

[0037] 0≤η t <T k (9)

[0038] In equations (8) and (9), T k As the threshold, T k ∈[0,1],η t ∈[0,1];

[0039] If η t If formula (8) is satisfied, then node P t Classified as belonging to the multi-layered cloud category; if η t If formula (9) is satisfied, then node P t The category does not belong to the multi-layered cloud part;

[0040] (S142) Determine the optimal segmentation threshold using the threshold evaluation index.

[0041] The evaluation metric for the threshold is the absolute value of the difference between the true positive rate and the true negative rate, as shown below:

[0042]

[0043] In equation (10), T k As the threshold, T k ∈[0,1],η t Let P be the kth node in the lookup table. t The confidence parameter; N is the number of multi-layer clouds in P; P belongs to a multi-layered cloud, and its confidence parameter satisfies T. k ≤η t The number of elements ≤1; n is the number of elements in P that do not belong to multi-layered clouds; P is not a multi-layer cloud, and its confidence parameter satisfies T. k ≤η t The number of items ≤1; For true yang rate, The true negative rate;

[0044] When the calculation result of the formula (10) is the maximum value, the corresponding threshold T k The optimal segmentation threshold is denoted as T.k ', i.e. T k ' is the optimal segmentation threshold for the k-th 3D lookup table;

[0045] (S200) The model invocation section includes the three-dimensional lookup table reading stage, the threshold segmentation stage, and the weight combination stage, as detailed below:

[0046] (S210) The three-dimensional lookup table reading stage includes:

[0047] Point P in the observation dataset that needs to be classified c If the observation value in the first dimension of the k-th 3D lookup table is a2, the observation value in the second dimension is b2, and the observation value in the third dimension is c2, then the point P to be classified is... c Given coordinates (a2, b2, c2), determine which point P needs to be classified. c Does it belong to node P? t The determination criteria require that three conditions be met simultaneously. These three conditions are as follows:

[0048]

[0049]

[0050]

[0051] In equations (11) and (13), r1 is the resolution of the first feature dimension in the k-th three-dimensional lookup table; r2 is the resolution of the second feature dimension in the k-th three-dimensional lookup table; and r3 is the resolution of the third feature dimension in the k-th three-dimensional lookup table.

[0052] If point P needs to be classified c If both conditions are met, then point P... c It belongs to node P in the lookup table. t , will node P t η t As point P c Confidence parameter in the k-th three-dimensional lookup table:

[0053] η kt =η t (14)

[0054] In equation (14), η kt Let P be the point c The confidence parameter is obtained by querying the k-th three-dimensional lookup table;

[0055] The threshold segmentation stage (S220) includes:

[0056] Using the confidence parameters obtained in the 3D lookup table reading stage and the optimal segmentation threshold obtained in the optimal segmentation threshold determination stage, the point P to be classified is determined. c The specific decision weights obtained from the k-th three-dimensional lookup table are as follows:

[0057] If η kt ≥T k (15)

[0058] Then τ k It equals 1;

[0059] If η kt <T k '(16)'

[0060] Then τ k Equals 0;

[0061] In equations (15) and (16), T k ' is the optimal segmentation threshold for the k-th 3D lookup table, τ k Let P be the point c Based on the decision weights obtained from the k-th 3D lookup table; repeat the 3D lookup table reading stage and the threshold segmentation stage, and obtain K decision weights for the K 3D lookup tables;

[0062] The weight combination stage (S230) includes:

[0063] The K decision weights of the K three-dimensional lookup tables obtained in the threshold segmentation stage are set, and the decision coefficients of the decision weights are obtained by linear combination, as follows:

[0064]

[0065] In equation (17), μ k τ is the determination coefficient, which can take any real number; k The weights are used to determine the confidence level; D is the confidence level.

[0066] The confidence scores are normalized as follows:

[0067]

[0068] In equation (18), D' is the final confidence level, D'∈[0,1]; D is the confidence level, D min D is the minimum confidence level. min =0; D max This represents the maximum confidence level.

[0069] The final confidence level is the point P that needs to be classified. cTo determine whether a point P belongs to a multi-layered cloud, the closer the final confidence score is to 0, the better the point P needs to be classified. c The more likely a point P is to contain a single layer of cloud, the closer its final confidence level is to 1, thus determining whether it needs to be classified. c It is more likely to contain multiple layers of clouds.

[0070] Preferably, in the stage of creating the three-dimensional lookup table, different numbers of nodes result in different resolutions, and the corresponding three-dimensional lookup table has different precision.

[0071] More preferably, in the stage of creating the three-dimensional lookup table, if n j If n ≥ 50, the resulting three-dimensional lookup table is a high-precision three-dimensional lookup table; if n j If the value is less than 50, the resulting 3D lookup table will be a low-precision 3D lookup table.

[0072] More preferably, in the stage of reading the three-dimensional lookup table, all three-dimensional lookup tables are read to obtain the point P that needs to be classified. c The decision weights in all three-dimensional lookup tables, where if the point belongs to η in the high-precision three-dimensional lookup table t If the value is empty, a low-precision 3D lookup table is used to find the result and replace it, or the nearest neighbor method is used to find the nearest neighboring point in the high-precision 3D lookup table.

[0073] Preferably, in the stage of creating the three-dimensional lookup table, the creation of the three-dimensional lookup table is updated according to the creation time of different training datasets.

[0074] Preferably, in the stage of creating the three-dimensional lookup table, the three-dimensional lookup table is created on the ground, and the created three-dimensional lookup table can be used for a long time.

[0075] Preferably, the three-dimensional lookup table is uploaded to the satellite to directly determine whether it is a multi-layered cloud; or the satellite observation data is transmitted to the ground observation center, and the three-dimensional lookup table is used to determine whether it is a multi-layered cloud.

[0076] Preferably, the final confidence level is linearly related to the detection accuracy of the final confidence level.

[0077] Preferably, the method can be used both during the day and at night.

[0078] The present invention also provides an application of the method described above in detecting multi-layered clouds.

[0079] The multi-layer cloud detection method based on a three-dimensional lookup table of the present invention solves the problems of low detection accuracy and poor linearity between final confidence and detection accuracy in existing technologies, and has the following advantages:

[0080] 1. This invention directly uses primary observation data from a radiation imager, without using secondary cloud inversion products, thus reducing the deviation in multi-layer cloud detection results caused by the inherent inaccuracies of secondary cloud products. Furthermore, this invention uses observation data from long-wave infrared channels, enabling all-weather use both day and night.

[0081] 2. The method of this invention can effectively determine whether the cloud structure is an ice cloud on the upper layer and a water cloud on the lower layer, or a multi-layered cloud structure where the upper layer is an ice cloud and the lower layer is also an ice cloud, or the upper layer is a water cloud and the lower layer is also a water cloud.

[0082] 3. This invention uses the final confidence level to make a more refined determination of whether it is a multi-layer cloud. It can determine different final confidence levels for different cloud scenarios, which is conducive to improving the inversion quality of cloud secondary products more precisely.

[0083] 4. The present invention uses an independent set of multiple judgment weights, which can quickly provide the final confidence level and determine whether it is a multi-layer cloud structure.

[0084] 5. The detection results of this invention show that the linearity between the final confidence level and its accuracy is good. After verification, the linearity between the final confidence level and its accuracy of the multi-layer cloud detection method is good. The final confidence level determined by the detection method can be used directly to compare whether there are multi-layer clouds between different observation points.

[0085] 6. The final confidence level of this invention is used as the predicted probability for determining multi-layered clouds. The detection results show that the closer the final confidence level is to 0, the more likely it is to be a single-layered cloud, and the closer the final confidence level is to 1, the more likely it is to be a multi-layered cloud. The detection accuracy is relatively high when the final confidence level is 1, and the highest detection accuracy is about 90%. Attached Figure Description

[0086] Figure 1 This is a flowchart illustrating the creation of multiple sets of 3D lookup tables for a multi-layer cloud detection method based on 3D lookup tables.

[0087] Figure 2 This is a flowchart illustrating the 3D lookup table invocation process for a multi-layer cloud detection method based on a 3D lookup table.

[0088] Figure 3 These are the accuracy curve, quantity curve, and perfect calibration curve of the method of this invention (compared to machine learning methods).

[0089] Figure 4 Accuracy curves, quantity curves, and perfect calibration curves for the Hist Gradient Boosting Classifier.

[0090] Figure 5For XGBoost, the accuracy curve, quantity curve, and perfect calibration curve.

[0091] Figure 6 Accuracy curves, quantity curves, and perfect calibration curves for Voting Classifier.

[0092] Figure 7 Accuracy curves, quantity curves, and perfect calibration curves for Random Forest Classifier.

[0093] Figure 8 Accuracy curves, quantity curves, and perfect calibration curves for the Multi-layer Perception Classifier.

[0094] Figure 9 Accuracy curves, quantity curves, and perfect calibration curves for the Gradient Boosting Classifier.

[0095] Figure 10 Accuracy curves, quantity curves, and perfect calibration curves for Calibrated Classifier CV.

[0096] Figure 11 Accuracy curves, quantity curves, and perfect calibration curves for the Bagging Classifier.

[0097] Figure 12 Accuracy curves, quantity curves, and perfect calibration curves for Logistic Regression.

[0098] Figure 13 Accuracy curves, quantity curves, and perfect calibration curves for Linear Discriminant Analysis.

[0099] Figure 14 For MODIS accuracy curves, quantity curves, and perfect calibration curves.

[0100] Figure 15 The accuracy curve, quantity curve, and perfect calibration curve of the method of this invention (compared to MODIS) are shown. Detailed Implementation

[0101] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0102] The abbreviations, key term definitions, and references used in the following embodiments and experimental examples are as follows:

[0103] Multilayered clouds: Clouds with a layered structure at a single pixel observation point.

[0104] The Moderate Resolution Imaging Spectroradiometer (MODIS) is a sensor technology jointly developed by NASA and Earth Observing System (EOS).

[0105] Feng-Yun-4A Satellite, FY-4A, 4.2m × 2.0m × 3.8m (length × width × height), manufactured by China Aerospace Science and Technology Corporation (China Aerospace Science and Technology Corporation First Academy);

[0106] The Advanced Geostationary Radiation Imager (AGRI) is a payload instrument on the FY-4A satellite, developed and manufactured by the China Aerospace Science and Technology Corporation (China Academy of Space Technology).

[0107] The Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) is a satellite observation project using lidar technology, developed in collaboration between NASA and the French National Centre for Research and Development (CNRS).

[0108] The Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) is a lidar device carried on the CALIPSO satellite, jointly developed and manufactured by NASA and the French National Centre for Research and Development (CNRS).

[0109] Confidence level: The detection method provides a numerical value to determine whether a point is a multi-layered cloud. The higher the value, the more likely the method is to determine that the point is a multi-layered cloud.

[0110] Final confidence level: The value after normalizing the confidence level.

[0111] Detection accuracy: The accuracy rate at which multi-layered clouds are detected.

[0112] The full name of Hist Gradient Boosting Classifier is Histogram-based Gradient Boosting Classification Tree. Source of the literature: Guryanov, Aleksei. "Histogram-based algorithm for building gradient boosting ensembles of piecewise linear decision trees." Analysis of Images, Social Networks and Texts: 8th International Conference, AIST 2019, Kazan, Russia, July 17–19, 2019, Revised Selected Papers 8. Springer International Publishing, 2019.

[0113] Source of the literature of XGBoost: Chen, Tianqi, et al. "Xgboost: extreme gradient boosting." R package version 0.4-21.4 (2015): 1-4.

[0114] Source of the literature of Voting Classifier: Ruta, Dymitr, and Bogdan Gabrys. "Classifier selection for majority voting." Information fusion 6.1 (2005): 63-81.

[0115] Source of the literature of Random Forest Classifier: Pal, Mahesh. "Random forest classifier for remote sensing classification." International journal of remote sensing 26.1 (2005): 217-222.

[0116] Literature source of Multi-layer Perception Classifier: Hampshire II, John B., and Barak Pearlmutter. "Equivalence proofs for multi-layer perceptron classifiers and the Bayesian discriminant function." Connectionist Models. Morgan Kaufmann, 1991. 159-172. Literature source of Gradient Boosting Classifier: Chakrabarty, Navoneel, et al. "Flight arrival delay prediction using gradient boosting classifier." Emerging Technologies in Data Mining and Information Security: Proceedings of IEMIS 2018, Volume 2. Springer Singapore, 2019. Natekin, Alexey, and Alois Knoll. "Gradient boosting machines, a tutorial." Frontiers in neurorobotics 7(2013):21.

[0117] The full name of Calibrated Classifier CV is Probability calibration with isotonic regression or logistic regression. Literature source: Leathart, Tim, et al. "Probability calibration trees." Asian conference on machine learning. PMLR, 2017.

[0118] Document source of Bagging Classifier: Quinlan, J. Ross. "Bagging, boosting, and C4.5." Aaai / Iaai, vol.1.1996. Zareapoor, Masoumeh, and Pourya Shamsolmoali. "Application of credit card fraud detection: Based on bagging ensemble classifier." Procedia computerscience 48.2015(2015):679-685.

[0119] Literature sources for Logistic Regression: LaValley, Michael P. "Logisticregression." Circulation 117.18 (2008): 2395-2399. Wright, Raymond E. "Logisticregression." (1995).

[0120] Document source for Linear Discriminant Analysis: Izenman, Alan Julian. "Lineardiscriminant analysis." Modern multivariate statistical techniques: regression, classification, and manifold learning. New York, NY: Springer New York, 2013.237-280.

[0121] Example 1

[0122] A multi-layer cloud detection method based on a 3D lookup table, such as Figure 1 As shown, the flowchart illustrates the creation process of multiple sets of 3D lookup tables for the multi-layer cloud detection method based on 3D lookup tables. Figure 2 The diagram shows the flowchart of the 3D lookup table call in the multi-layer cloud detection method based on 3D lookup tables. This method includes:

[0123] The model training part and the model calling part are as follows: the model training part constructs and creates a 3D lookup table, and the model calling part uses the created 3D lookup table to find the confidence parameters and construct the confidence score. After the confidence score is normalized, the final confidence score is obtained.

[0124] (S100) The model training part includes the data feature creation stage, the feature combination stage, the 3D lookup table creation stage, and the optimal segmentation threshold determination stage, as detailed below:

[0125] (S110) Data Feature Creation Stage

[0126] Specific features are created for different application observation instruments. For example, for multi-layer cloud detection designed for AGRI, in order to achieve all-weather use, mid-wave infrared and long-wave infrared channels are selected, specifically 3.75μm, 6.25μm, 7.1μm, 8.5μm, 10.7μm and 12μm. Since the observations of the 7.1μm and 6.25μm channels, the 8.5μm and 10.7μm channels, and the 10.7μm and 12μm channels are relatively close to the cloud observations, in order to improve the sensitivity of similar observations to multi-layer cloud observations, the 6.25μm, 8.5μm, and 12μm channels in the 3.75μm, 6.25μm, 7.1μm, 8.5μm, 10.7μm, and 12μm channels are replaced with BTD(7.1-6.25), BTD(8.5-10.7), and BTD(10.7-12) to characterize the observations of channels adjacent to the observations. Channels of 3.75μm, 7.1μm, and 10.7μm were used, along with brightness temperature difference (BTD), BTD(7.1-6.25), BTD(8.5-10.7), and BTD(10.7-12) as the final data features, for a total of 6 data features (N1 = 6).

[0127] (S120) Feature Combination Stage

[0128] Data features are combined using a permutation and combination method (three-by-three combination). Each data feature combination contains three data features; that is, three features are randomly selected from the total number of data features as the three dimensions of each data feature combination. The specific representation method is as follows:

[0129]

[0130] In equation (1), N1 is the total number of data features, and K (K = 20) is the total number of combinations of data features.

[0131] It is evident that a maximum of K combinations of data features can be constructed from the data features. This invention uses combinations of data features to represent the data, and employs combinations of three-dimensional lookup tables that differ in at least one dimension to determine whether a cloud is multi-layered. This allows the detection results to be used directly to determine the probability of different observation points forming multi-layered clouds based on the detection confidence level.

[0132] (S130) The stage of creating a three-dimensional lookup table includes the following steps:

[0133] (S131) ​​Construct a three-dimensional lookup table

[0134] The K data feature combinations obtained in the feature combination stage are used to construct three-dimensional lookup tables. The three dimensions of each data feature combination are used as the three dimensions of the corresponding three-dimensional lookup table, as follows: For the k-th (k∈[1,K]) three-dimensional lookup table, the j-th dimension (j∈[1,3]) is divided into three dimensions based on the distribution range of the observations in the j-th dimension. The resolution in the j-th dimension is obtained by using nodes to divide the distribution range of the observations in the j-th dimension. The relationship between resolution and distribution range is as follows:

[0135]

[0136] In equation (2), r j For the resolution of the j-th dimension, S j Let n be the distribution range of the observations in the j-th dimension. j The number of nodes in the j-th dimension;

[0137] The distribution range of observed values ​​for the corresponding data features was statistically determined, and the corresponding resolution was calculated based on the number of nodes, as detailed in Table 1. These channels are all brightness temperatures; therefore, the units for distribution range and resolution are Kelvin (K).

[0138] Table 1. Distribution range of observation values ​​for each channel in this embodiment.

[0139]

[0140] After determining the resolution in the j-th dimension, any combination of data features can be represented using the resolution and number of nodes in the corresponding three-dimensional lookup table. That is, for the k-th three-dimensional lookup table, the first feature dimension represents the number of nodes in the first dimension as n1, with a resolution of r1; the second feature dimension represents the number of nodes in the second dimension as n2, with a resolution of r2; and the third feature dimension represents the number of nodes in the third dimension as n3, with a resolution of r3. Setting different numbers of nodes results in different resolutions and different precisions for the corresponding three-dimensional lookup tables. If the number of nodes is n... j ≥50, resulting in a high-precision 3D lookup table; if node n j If the value is less than 50, the resulting three-dimensional lookup table will be a low-precision three-dimensional lookup table. In this embodiment, 20 sets of n were created respectively. j A high-precision three-dimensional lookup table of 50 and 20 sets of n jA low-precision three-dimensional lookup table of 25 is obtained, resulting in a total of 40 three-dimensional lookup tables, j∈[1,3].

[0141] The three-dimensional lookup table method of this invention allows the distribution characteristics of observed data for each channel (data feature) to be stored in the three-dimensional lookup table as confidence parameters. The high-precision three-dimensional lookup table is created to ensure higher resolution of the data distribution within the lookup table, resulting in more accurate judgments in the model invocation process. The low-precision lookup table compensates for the range of missing observation values ​​in the high-precision lookup table, thus supplementing the corresponding prediction results. This invention uses high- and low-precision three-dimensional lookup tables to represent data, reducing the computational load and complexity, accelerating the model training process, and significantly reducing the computational cost of data classification.

[0142] (S132) Calculate the parameters of the three-dimensional lookup table

[0143] Because the observational characteristics of clouds are more consistent over the ocean surface, data from the ocean surface were used as the training dataset. The training dataset contains 97,456 data sets, accounting for 29.087% of the total dataset of 335,053 sets (land + ocean). Therefore, the training dataset contains N data sets. a (N a =97456) observation points P o It is represented as:

[0144] P = {P} o}, o∈[0,N a (3)

[0145] In equation (3), o is the o-th sample point, and P is the total number of observation points. o A set;

[0146] Let P be any node in the k-th three-dimensional lookup table. t Let the coordinates be (x, y, z), where x is the first dimension, y is the second dimension, and z is the third dimension. Let P be any observation point in the training dataset. o The coordinates of the observation point are (a1, b1, c1), where a1 is the observation value in the first dimension, b1 is the observation value in the second dimension, and c1 is the observation value in the third dimension. The observation point P is determined. o Does it belong to node P? t Three conditions must be met simultaneously, as follows:

[0147]

[0148]

[0149]

[0150] In equations (4), (5) and (6), r1 is the resolution of the first feature dimension in the k-th three-dimensional lookup table; r2 is the resolution of the second feature dimension in the k-th three-dimensional lookup table; and r3 is the resolution of the third feature dimension in the k-th three-dimensional lookup table.

[0151] If the observation points P in the training dataset o If all three conditions (4), (5), and (6) are met, then the observation point P... o It belongs to node P in the three-dimensional lookup table. t .

[0152] Suppose there are N in the training dataset. t observation point P o It belongs to node P t In N t observation point P o The CPC has N t1 observation point P o It belongs to multi-layered clouds; if N t If the value is 0, then the node P t There is no corresponding observation point P in the training dataset. o Set η t It is an empty value; if N t If not equal to 0, then according to N t With N t1 Calculate η t :

[0153]

[0154] In equation (7), η t For the k-th node P in the lookup table t Confidence parameter;

[0155] By calculating step by step using the above formula (7), all nodes P in any k-th three-dimensional lookup table can be obtained. t The confidence level parameter.

[0156] The invented 3D lookup table possesses advantages such as low algorithm complexity, ease of implementation, and strong generalization ability, making the method operable at different resolutions. For the same feature combination, high- and low-precision 3D lookup tables can be used to compensate for each other, avoiding the problem of missing parameters due to the absence of corresponding nodes in the training dataset. Furthermore, the creation of different lookup tables is an independent process; therefore, the creation process of the 3D lookup table for the multi-layer cloud detection method of this invention can also be performed in parallel, improving computation speed and reducing computation time.

[0157] The creation of 3D lookup tables is typically conducted on the ground, and a single 3D lookup table can be used long-term. The 3D lookup table is updated appropriately based on the creation time of different training datasets. The 3D lookup table, created on the ground, is uploaded to a satellite to directly determine whether a cloud layer is multi-layered; alternatively, satellite observation data is transmitted to a ground observation center, and the 3D lookup table is used in conjunction with the data to determine whether a cloud layer is multi-layered.

[0158] (S140) The stage of determining the optimal segmentation threshold includes the following steps:

[0159] (S141) Set a threshold, and determine each node P in the k-th three-dimensional lookup table based on the set threshold. t Does it belong to a multi-layered cloud?

[0160] Take a classification threshold in the k-th 3D lookup table. For each node P in any k-th 3D lookup table obtained during the 3D lookup table creation stage... t η t The relationship with the threshold is as follows:

[0161] T k ≤η t ≤1 (8)

[0162] 0≤η t <T k (9)

[0163] In equations (8) and (9), T k As the threshold, T k ∈[0,1],η t ∈[0,1].

[0164] If η t If formula (8) is satisfied, then node P t Classified as belonging to the multi-layered cloud category; if η t If formula (9) is satisfied, then node P t This category does not belong to the multi-layered cloud category.

[0165] (S142) Determine the optimal segmentation threshold using the threshold evaluation index.

[0166] The evaluation metric for the threshold is the absolute value of the difference between the true positive rate and the true negative rate, as shown below:

[0167]

[0168] In equation (10), T k As the threshold, T k ∈[0,1],η t For the k-th node P in the lookup table tThe confidence parameter; N is P (P is the total number of observation points P). o The number of multi-layered clouds in the set); P belongs to a multi-layered cloud, and its confidence parameter satisfies T. k ≤η t The number of elements ≤1; n is the number of elements in P that do not belong to multi-layered clouds; P is not a multi-layer cloud, and its confidence parameter satisfies T. k ≤η t The number of items ≤1; For true yang rate, The true negative rate;

[0169] The threshold T corresponding to the maximum value calculated by formula (10) k The optimal segmentation threshold is denoted as T. k ', T k As the optimal segmentation threshold for the k-th 3D lookup table, the optimal segmentation threshold T for each lookup table in this embodiment is calculated. k 'All are equal to 0.4.'

[0170] The determination of the optimal segmentation threshold is to combine the search results of multiple three-dimensional lookup tables, and also to find the threshold for the optimal classification of a single three-dimensional lookup table.

[0171] This invention provides a method for determining the optimal segmentation threshold of a three-dimensional lookup table. This method is used to effectively classify the thresholds of various three-dimensional lookup tables, and to combine the classification effects of multiple three-dimensional lookup tables to achieve a higher accuracy classification result.

[0172] (S200) Model Call Section

[0173] To evaluate the effectiveness of this invention, the dataset used in the model invocation phase is a test dataset. Clouds detected by CALIOP with two or more layers are labeled as multi-layer clouds, while those detected with only one layer are labeled as single-layer clouds. This embodiment compares this invention with two other multi-layer cloud detection methods: First, it compares with conventional machine learning methods, using observation data from AGRI and CALIOP at the top of the hour throughout 2018, totaling 335,053 data sets. Approximately 30% of these data sets were randomly sampled as the test dataset, i.e., approximately 100,516 test dataset sets. Second, it compares with the multi-layer cloud detection results (Cloud_Multi_Layer_Flag) from the on-orbit MODIS cloud detection product MYD06_L2—MODIS / Aqua Clouds 5-Min L2 Swath 1km and 5km. After data matching, a total of 40,406 data sets were used as the test dataset. For each test data point in the two types of test datasets above, the model calls the following steps in sequence: reading the lookup table, threshold segmentation, and weight combination.

[0174] The process includes three stages: reading the 3D lookup table, threshold segmentation, and weight combination, as detailed below:

[0175] (S210) The stage of reading the three-dimensional lookup table includes: Let P be any node in the k-th three-dimensional lookup table. t Its coordinates are (x, y, z), where x is the first dimension, y is the second dimension, and z is the third dimension. For the point P that needs to be classified... c If the observation value in the first dimension of the k-th three-dimensional lookup table is a2, the observation value in the second dimension is b2, and the observation value in the third dimension is c2, then the point P to be classified is... c Given coordinates (a2, b2, c2), determine which point P needs to be classified. c Does it belong to node P? t The determination criteria require that three conditions be met simultaneously. These three conditions are as follows:

[0176]

[0177]

[0178]

[0179] In equations (11) and (13), r1 is the resolution of the first feature dimension in the k-th three-dimensional lookup table; r2 is the resolution of the second feature dimension in the k-th three-dimensional lookup table; and r3 is the resolution of the third feature dimension in the k-th three-dimensional lookup table.

[0180] If point P needs to be classifiedc If both conditions are met, then point P... c It belongs to node P in the lookup table. t , will node P t η t As point P c Confidence parameter in the k-th three-dimensional lookup table:

[0181] η kt =η t (14)

[0182] In equation (14), η kt Let P be the point c The confidence level parameter.

[0183] Read all the 3D lookup tables to obtain the point P that needs to be classified. c The decision weights in all three-dimensional lookup tables, where if the point belongs to η in the high-precision three-dimensional lookup table t If the value is null, then η is obtained by using a low-precision three-dimensional lookup table. t Alternatively, you can use the nearest neighbor method to find the nearest point in a high-precision 3D lookup table.

[0184] The lookup table reading stage performs classification operations on the dataset to be classified, and the use of different three-dimensional lookup tables is independent of each other. Therefore, the method of this invention has low complexity, low computational resource consumption, and high computational efficiency, and can be used on embedded machines with limited computing power. It greatly reduces the computational requirements for data classification, thereby achieving real-time data classification, reducing the time for multi-layer cloud detection result output, and accelerating the detection speed.

[0185] (S220) The threshold segmentation stage includes:

[0186] Using the confidence parameters obtained in the 3D lookup table reading stage and the optimal segmentation threshold obtained in the optimal segmentation threshold determination stage, the point P to be classified is determined. c The specific decision weights obtained from the k-th three-dimensional lookup table are as follows:

[0187] If η kt ≥T k (15)

[0188] Then τ k It equals 1;

[0189] If η kt <T k '(16)'

[0190] Then τ k Equals 0;

[0191] In equations (15) and (16), T k ' is the optimal segmentation threshold for the k-th 3D lookup table, τ k Let P be the point c Based on the decision weights obtained from the k-th 3D lookup table; repeat the 3D lookup table reading stage and the threshold segmentation stage, and obtain K decision weights for the K 3D lookup tables;

[0192] This invention uses threshold segmentation of the three-dimensional lookup table to determine the classification results, which can achieve a consistent and effective combination of classification results from individual classification results to multiple classification results, thereby increasing the effectiveness of the classification method.

[0193] The weight combination stage (S230) includes:

[0194] The K decision weights of the K three-dimensional lookup tables obtained in the threshold segmentation stage are set, and the decision coefficients of the decision weights are obtained by linear combination, as follows:

[0195]

[0196] In equation (17), μ k The coefficient of determination is set to 1 for all values; τ k The weight is used to determine the confidence level; D is the confidence level, which in this embodiment is D∈[0,20];

[0197] The confidence scores are normalized as follows:

[0198]

[0199] In equation (18), D' is the final confidence level, D'∈[0,1]; D is the confidence level. min D is the minimum confidence level. min =0; D max This represents the maximum confidence level.

[0200] The final confidence level is the point P that needs to be classified. c To determine whether a point P belongs to a multi-layered cloud, the closer the final confidence level is to 0, the more likely it is to be classified. c The more likely a point P is to contain a single layer of cloud, the closer its final confidence level is to 1, and the more likely it is to be classified. c It is more likely to contain multiple layers of clouds.

[0201] The invention uses the linear combination of the judgment values ​​after threshold segmentation of each three-dimensional lookup table as the confidence level (classification standard). This helps to establish a good linear correlation between the classification accuracy and the final confidence level. The accuracy of data classification can be directly judged by the magnitude of the final confidence level. At the same time, it enables a more refined judgment of the multi-layer cloud detection results. The higher the final confidence level, the more likely the judgment point is to be a multi-layer cloud, which helps to improve the quality of cloud level 2 products.

[0202] To visually demonstrate which categories point P belongs to within the test dataset c It could be a multi-layered cloud. By coloring it using the final confidence level, we can intuitively see that the darker the color, the more it represents the category of point P. c The more likely it is to be a multi-layered cloud structure. A crucial prerequisite for determining the final confidence level is providing the detection accuracy for the multi-layered cloud corresponding to that final confidence level. Detection accuracy increases linearly with the final confidence level, ensuring accuracy for any classification point P. c The final confidence level is higher than that of point P in the other classification. c The final confidence level is then the classification point P. c The probability of point P being a multi-layered cloud is higher than that of the other category. c Simultaneously, it ensures that during the coloring process, if the final confidence level changes by the same proportion, the accuracy of multi-layer cloud detection also changes by approximately the same proportion, enabling the direct differentiation of points P in various scenarios based on the degree of color change. c The degree of variation in the probability of whether it is a multi-layered cloud allows for a more intuitive and efficient comparison of the relative probability of a multi-layered cloud among the observed pixels to be classified in different test datasets.

[0203] Experiment Example 1: Analysis of Multi-layer Cloud Detection Results

[0204] The detection results obtained by applying the method in Example 1 are compared and analyzed with the results of several common machine learning classification methods. This invention was performed on an Intel Core i7 eight-core processor, 16GB of memory, a 1TB hard drive, and a Windows operating system, using PyCharm software. The machine learning methods used for comparison are programs from Scikit-learn. The datasets used in this invention from three instruments (AGRI, CALIOP, and MODIS) are as follows: For AGRI multi-layer cloud detection, the mid-wave infrared, water vapor channel, and long-wave infrared channel from AGRI Level 1 SDR data are used, specifically 3.75, 6.25, 7.1, 8.5, 10.7, and 12 μm; the operational MODIS Collection 6 cloud product (MYD06_L2) containing multi-layer cloud markers; and the Level 2 product Layer Information from the CALIPSO instrument, which includes the number of clouds detected by CALIOP, as well as the cloud phase and cloud top height of each cloud layer; the 532nm Layer Optical Depth data contains the optical depth information of the cloud layer at each observation point detected by CALIPSO, serving as a benchmark for evaluating the effectiveness of the detection method, as detailed in Table 2.

[0205] Table 2 Experimental Data

[0206]

[0207] (1) Creation of the training dataset of this invention

[0208] For matching the observation data from the three satellites mentioned above, this invention mainly uses a combination of temporal and spatial matching. For temporal matching, the observation time points of CALIPSO and Aqua satellites are used, limited to within 5 minutes after the observation time of FY-4A satellite. For spatial matching, there are two restrictions: First, the minimum difference in Euclidean distance between latitude and longitude is used for selection, and the maximum Euclidean distance between matching points is limited to 0.1. Second, since AGRI's observation data has a 4km resolution, while MODIS and CALIOP have a 5km resolution, to maximize the overlap of the observation areas of the three matched satellites, the Euclidean distance between MODIS and CALIOP, and between AGRI and CALIOP, is limited to no more than half the Euclidean distance between adjacent CALIOP observation points. Regarding the time span of the experimental data, this invention selects data within 5 minutes of the hour throughout 2018. The matching results of CALIPSO and FY-4A satellites form the entire dataset, totaling 335,053 data sets. Because the bias in FY-4A satellite observation data is related to the satellite's zenith angle, the satellite elevation angle is limited to within 60 degrees to ensure more effective cloud observation. The detection method designed in this invention distinguishes between single-layer and multi-layer clouds. Therefore, the observed pixels must be observation points with clouds, meaning only observation data containing cloud layer observations are retained. Furthermore, since the cloud observation characteristics are more consistent over the ocean surface, ocean surface data is used as the training dataset, totaling 97,456 sets, accounting for 29.087% of the total dataset of 335,053 sets (land + ocean). This invention labels multi-layer clouds as those detected by CALIOP with two or more cloud layers, and single-layer clouds as those detected by CALIOP with one cloud layer.

[0209] (2) Verification of the method of the present invention

[0210] The verification of this invention was compared with machine learning methods and with the multi-layer cloud detection results of MODIS. The machine learning (selecting the Scikit-learn program) package included 11 classification methods (HistGradient Boosting Classifier, XGBoost, Voting Classifier, Random Forest Classifier, Multi-layer Perception Classifier, Gradient Boosting Classifier, Calibrated Classifier CV, Bagging Classifier, Logistic Regression, and Linear Discriminant Analysis) for comparison. The test dataset consisted of all 335,053 data sets from the matching results of the CALIPSO and FY-4A satellites. 30% of this dataset, approximately 100,516 data sets, was randomly sampled as the test dataset. The aforementioned methods were used to classify multi-layer clouds in the FY-4AAGRI observation data. Clouds detected by CALIOP with two or more layers were labeled as multi-layer clouds, while clouds detected with only one layer were labeled as single-layer clouds.

[0211] For any multi-layer cloud detection method, the final confidence score D' ∈ [0,1] calculated for the test dataset. Since this invention has 20 lookup tables, there are a total of 21 confidence scores in this embodiment, and the final confidence score is also 21. The distribution range of the final confidence score of the machine learning multi-layer cloud detection method is also [0,1]. Although its distribution range is continuous, to calculate its evaluation index value and reduce the impact of the number of confidence scores on the evaluation index, the final confidence score of the machine learning method is divided into 21 parts, the same number as the final confidence score of the method in this invention. MODIS multi-layer cloud labels range from 2 to 9, with a total of 8 different confidence scores, which, after normalization, yield 8 final confidence scores.

[0212] Let z be the final confidence level for any multi-layer cloud detection method. As mentioned above, z = 21 in the multi-layer cloud detection method of this invention, z = 21 in the machine learning multi-layer cloud detection method, and z = 8 in the MODIS multi-layer cloud detection method.

[0213] For the i-th term, i∈[1,z], the final confidence level is D' i Detection accuracy The calculation method is as follows:

[0214]

[0215] In equation (19), The final confidence score for the test dataset is calculated as D'. i The number of multi-layered clouds, The final confidence score for the test dataset is calculated as D'. i The number of all clouds

[0216] According to formula (19), the corresponding i-th final confidence level D' can be calculated. i Detection accuracy

[0217] At the final confidence level D' i Detection accuracy The value, which is also the final confidence level D' of multi-layer clouds. i Percentage below. (in D') i (Confidence) is the x-axis. An accuracy curve can be plotted with (Accuracy) on the ordinate. To determine the number of all clouds at different final confidence levels, D' i (Confidence) is the x-axis, and (Quantity) can be used as the vertical axis to plot a quantity curve. The quantity curve represents the trend of the number of all clouds under different final confidence levels after the test dataset is calculated by the multi-layer cloud detection method. The more clouds there are, the more reliable the detection accuracy calculated using formula (19) is.

[0218] To verify that the final confidence level is linearly related to its detection accuracy, the final confidence level D' and the detection accuracy A are used. D' R 2 As the first evaluation indicator. (By D') i and Calculate R 2 The specific steps are as follows:

[0219] D' i The average values ​​are as follows:

[0220]

[0221] Total Sum of Squares (SS) tot as follows:

[0222]

[0223] Sum of Squared Residuals (SS) res as follows:

[0224]

[0225] From equations (20), (21) and (22), we can obtain

[0226]

[0227] From (20) and (23), the first evaluation index value R of different multi-layer cloud detection methods can be calculated. 2 Ideally, the accuracy of multi-cloud detection methods increases with increasing final confidence. That is, the accuracy curve should show a monotonically increasing trend with the final confidence level, and the detection accuracy should increase or decrease proportionally when the final confidence level increases or decreases by the same value. This requires high linearity in accuracy. If the accuracy curve is monotonically increasing and has high linearity, it can be colored according to the final confidence level, allowing for a visual comparison of the probability of multi-cloud formations at different observation points. Furthermore, the intensity of color changes can directly compare the magnitude of changes in the probability of multi-cloud formations. Therefore, the first evaluation metric, R0, is crucial. 2 The closer to 1, the better.

[0228] To evaluate which multi-layer cloud detection method is best suited to use the final confidence score as the predicted probability of identifying a multi-layer cloud, a second evaluation metric, Calibration (C), needs to be used. The specific calculation is as follows:

[0229]

[0230] Ideally, the predicted probability equals the percentage of multi-layered clouds at that confidence level. At this point, D' i The horizontal axis is... The curve plotted on the vertical axis is the perfectly calibrated curve. At this point, the final confidence level is used as the prediction probability and the effect is the best. According to formula (24), the calibration value is equal to 0 at this point. Therefore, the closer the calibration value is to 0, the better. The closer the calibration value is to 0, the more reliable it is to use the final confidence level as the prediction probability.

[0231] (3) Comparison of experimental results

[0232] For detailed experimental and machine learning results using the method of this invention, please refer to [link / reference]. Figures 3 to 15 The names of the various classification methods in machine learning are shown in the titles of the graphs.

[0233] like Figure 3 As shown, the accuracy curve and quantity curve of the method of the present invention (compared with machine learning methods) are shown; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfectly calibrated curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0234] like Figure 4 As shown, the Hist GradientBoosting Classifier has an accuracy curve and a quantity curve; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis represents the final confidence level, and the vertical axis represents the quantity and percentage.

[0235] like Figure 5 As shown, the accuracy curve and quantity curve of XGBoost are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0236] like Figure 6 As shown, the accuracy curve and quantity curve of Voting Classifier are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0237] like Figure 7As shown, the accuracy curve and quantity curve of Random Forest Classifier are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0238] like Figure 8 As shown, the accuracy curve and quantity curve of the Multi-layer Perception Classifier are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis represents the final confidence level, and the vertical axis represents the quantity and percentage.

[0239] like Figure 9 As shown, the accuracy curve and quantity curve of Gradient Boosting Classifier are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0240] like Figure 10 As shown, the accuracy curve and quantity curve of Calibrated Classifier CV are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0241] like Figure 11As shown, the accuracy curve and quantity curve of Bagging Classifier are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0242] like Figure 12 As shown, the accuracy curve and quantity curve of Logistic Regression are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfectly calibrated curve; the horizontal axis represents the final confidence level, and the vertical axis represents the quantity and percentage.

[0243] like Figure 13 As shown, the accuracy curve and quantity curve of Linear Discriminant Analysis are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis is the final confidence level, and the vertical axis is the quantity and percentage.

[0244] Through experiments Figures 3 to 13 It can be seen that the linearity between the final confidence level and the detection accuracy of this invention is the best. While the accuracy curves of Gradient Boosting Classifier and Multi-layer Perception Classifier exhibit linear characteristics, the linearity is poor at both high and low final confidence levels. Furthermore, the curve values ​​are low when the final confidence level is above 0.8, indicating that the accuracy curves are unreliable at these levels. Bagging Classifier, Voting Classifier, and XGBoost, although showing a single increase in accuracy with confidence, exhibit a curvilinear characteristic. In contrast, the accuracy curves of other detection methods show poor linearity, meaning that the final confidence level C of the other methods for classifying multi-layer clouds is low. o nfid e n ceThe size of the coloring index is not a good representation of the classification accuracy under this classification method. It is unsuitable for directly determining prediction accuracy using graphs, nor is it suitable for comparing the probability of different observation points having multi-layered clouds using the final confidence score. Therefore, to quantitatively compare which classification method is more suitable for coloring based on the final confidence score, and thus distinguish the probability of different observation points in the observation area having multi-layered clouds, the first evaluation index is... 2 The closer to 1, the better. Table 3 shows that this invention (compared to machine learning methods) [is effective]. 2 It is closest to 1, therefore the linearity of this invention is the best. If the final confidence level is used as the predicted probability of determining whether it is a multi-layered cloud, then the calibration value (C) a libr a ti o The closer n is to 0, the better. As shown in Table 3, this invention (compared to machine learning methods), C a libr a ti o The second evaluation metric is the smallest, therefore using the final confidence level of this invention as the prediction probability of multi-layered clouds is the most suitable.

[0245] Table 3 Objective Evaluation Indicators

[0246]

[0247] The classification performance of other machine learning methods, regardless of the linearity evaluation metric R... 2 Looking at the second evaluation metric, none of them perform as well as the method described in this invention. Therefore, the method of this invention is most suitable for multi-layer cloud detection by meteorological satellites, and its overall performance is superior. Meanwhile, from... Figure 3 It can be seen that the method of the present invention has a high detection accuracy when the final confidence level is 1, and the highest detection accuracy is about 90%.

[0248] In addition to comparing with existing machine learning classification methods, this invention also compared with the multi-layer cloud label Cloud_Multi_Layer_Flag in the MODIS cloud detection product MYD06_L2—MODIS / Aqua Clouds 5-Min L2 Swath 1km and 5km, which is currently in orbit. The training dataset for this invention's multi-layer cloud detection method still uses partial data from the ocean surface, totaling 97,456 data sets. The test dataset uses data from three satellites—FY-4A, Aqua, and CALIPSO—with a temporal resolution of less than 5 min; observation accuracies of 4km, 5km, and 5km respectively; and a constraint that the Euclidean distance between latitude and longitude is less than 0.1 degrees. After matching, a total of 40,406 data sets were obtained as the test dataset. This was used to compare the effectiveness of this invention's multi-layer cloud detection method for AGRI with that of MODIS. The objective evaluation indicators for the comparison are shown in Table 4. In this case, this invention detected 1,381 multi-layer cloud groups with an accuracy greater than 80%. For the multi-layer cloud confidence scores D∈{2,3,4,5,6,7,8,9} of MODIS, the MODIS D min D is the minimum confidence level. min =2,D max The maximum value D in the confidence level max =9. Normalization according to formula (18) yields the final confidence level D'∈{0, 0.14, 0.28, 0.42, 0.57, 0.71, 0.85, 1.00} for MODIS's determination of multi-layer clouds. Accuracy curves, quantity curves, and perfect calibration curves are plotted according to formulas (19) and (24) on page 19 of the instruction manual. Simultaneously, the first evaluation index R is calculated. 2 And the second evaluation metric, Calibration, see details. Figure 14 , 15 And Table 4.

[0249] Table 4 Objective Evaluation Indicators

[0250]

[0251] like Figure 14 As shown, the MODIS accuracy curve and quantity curve are displayed; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis represents the final confidence level, and the vertical axis represents the quantity and percentage. Figure 15As shown in Table 4, the accuracy curve and quantity curve of the method of the present invention (compared with MODIS) are presented; where the accuracy curve is the accuracy curve, the quantity curve is the quantity curve, and the perfectly calibrated curve is the perfect calibration curve; the horizontal axis represents the final confidence level, and the vertical axis represents the quantity and percentage. Table 4 shows that the multilayer cloud detection method of the present invention (compared with MODIS) achieves the best results in the first evaluation index R... 2 Compared to MODIS, the multi-layer cloud detection method of this invention exhibits better linearity. Furthermore, the second evaluation metric (Calibration) of this invention's method (compared to MODIS) is smaller, making the multi-layer cloud detection method of this invention more suitable for directly using confidence level as the predicted probability of a multi-layer cloud.

[0252] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A multi-layer cloud detection method based on a three-dimensional lookup table, characterized in that, This method constructs and creates a 3D lookup table during model training, and then uses the constructed 3D lookup table to find confidence parameters and construct confidence scores during model invocation. The final confidence score is then used to determine the likelihood of a multi-layered cloud. The method includes the following steps: (S100) The model training part includes the data feature creation stage, the feature combination stage, the 3D lookup table creation stage, and the optimal segmentation threshold determination stage, as detailed below: (S110) Data feature creation stage The mid-wave infrared and long-wave infrared channels in the training dataset are used as data features, or the mid-wave infrared and long-wave infrared channels, the channel difference of mid-wave infrared and the channel difference of long-wave infrared in the training dataset are used as data features. (S120) The feature combination stage includes: The data features obtained in the feature creation stage are used to generate data feature combinations through permutation and combination. From the total number of data features, three features are randomly selected as the three dimensions of each data feature combination, as follows: （1） In equation (1), The total number of data features. K This represents the total number of combinations of data features; (S130) The stage of creating a three-dimensional lookup table includes the following steps: (S131) ​​Construct a three-dimensional lookup table The features obtained in the feature combination stage K Each data feature combination is used to construct a three-dimensional lookup table, with the three dimensions of each data feature combination serving as the three dimensions of the corresponding three-dimensional lookup table, as detailed below: For the A three-dimensional lookup table, containing the third of its three dimensions. The dimension, according to the first The distribution range of the observations in the dimension, using node pairs. The distribution range of the observations in the dimension is divided to obtain the _th ... The resolution of each dimension and its relationship with the distribution range of the observed values ​​are as follows: （2） In equation (2), For the first j Resolution in each dimension For the first j The distribution range of observations in each dimension, For the first j The number of nodes in each dimension; After determining the first After resolving the resolution in each dimension, for any combination of data features, the corresponding three-dimensional lookup table can be used. The 3D lookup table is created by considering the resolution and number of nodes in each dimension. (S132) Calculate the parameters of the three-dimensional lookup table. Suppose there are a total of Observation points It is represented as: （3） In equation (3), For the first One sample point, All observation points A set; Let any number of... k Any node in a three-dimensional lookup table The coordinates are ( ),in For the first dimension, For the second dimension, For the third dimension, let's define any observation point in the training dataset. The coordinates are ( , , ),in For the observations in the first dimension, For the observations in the second dimension, To determine the observation point for the observation value in the third dimension. Does it belong to a node? Three conditions must be met simultaneously, as follows: （4） （5） （6） In equations (4), (5) and (6), For the first k The resolution of the first dimension in a 3D lookup table; For the first k The resolution of the second dimension in a three-dimensional lookup table; For the first k The resolution of the third dimension in a three-dimensional lookup table; If the observation points in the training dataset If all three conditions (4), (5), and (6) are met, then the observation point... It is a node in the three-dimensional lookup table. ; Suppose there are a total of Observation points It belongs to the node ,exist Observation points The CCP Observation points It belongs to the multi-layered cloud category; if If the value is 0, then the node... There is no corresponding observation point in the training dataset. ,set up It is an empty value; if If it is not equal to 0, then according to and calculate : ， （7） In equation (7), For the first k Nodes in the lookup table ; By calculating step by step using the above formula (7), we can obtain any number of... k All nodes in a 3D lookup table of parameter; (S140) The stage of determining the optimal segmentation threshold includes the following steps: (S141) Set a threshold, and determine the first [missing information] based on the set threshold. k Each node in a three-dimensional lookup table Does it belong to a multi-layered cloud? In the k Take a threshold from a 3D lookup table This threshold and any first value obtained during the stage of creating the three-dimensional lookup table k Each node in a three-dimensional lookup table of The relationship is as follows: （8） （9） In equations (8) and (9), For the threshold, , ; like If the formula (8) is satisfied, then the node Classified as belonging to the multi-layered cloud category; if If the formula (9) is satisfied, then the node The portion classified as not belonging to multi-layered clouds; (S142) Determine the optimal segmentation threshold using the threshold evaluation index. The evaluation metric for the threshold is the absolute value of the difference between the true positive rate and the true negative rate, as shown below: （10） In equation (10), For the threshold, , for The number of clouds belonging to the multi-layered cloud system; for It belongs to a multi-layered cloud, and its satisfy Quantity; for The number of clouds that do not belong to the multi-layered cloud system; for It does not belong to multi-layered clouds, and its satisfy Quantity; For true yang rate, The true negative rate; When the calculation result of formula (10) is the maximum value, the corresponding threshold is... The optimal segmentation threshold is denoted as . ,Right now For the first k The optimal segmentation threshold for a three-dimensional lookup table; (S200) The model invocation section includes the three-dimensional lookup table reading stage, the threshold segmentation stage, and the weight combination stage, as detailed below: (S210) The stage of reading the three-dimensional lookup table includes: Points in the observation dataset that need to be classified In the k The observation value on the first dimension of the three-dimensional lookup table is The observed value in the second dimension is The observed value in the third dimension is Then the points need to be classified. The coordinates are ( , , ), determine the points that need to be classified Does it belong to a node? The determination criteria require that three conditions be met simultaneously. These three conditions are as follows: （11） （12） （13） In equations (11) to (13), For the first k The resolution of the first feature dimension in a 3D lookup table; For the first k The resolution of the second feature dimension in a three-dimensional lookup table; For the first k The resolution of the third feature dimension in a three-dimensional lookup table; If points need to be categorized If the above formulas (11) to (13) are satisfied simultaneously, then the point... It belongs to the node in the lookup table. , will node of As a point In the k A three-dimensional lookup table : （14） In equation (14), For point According to the k The confidence parameters obtained from a three-dimensional lookup table query; (S220) The threshold segmentation stage includes: Using the confidence parameters obtained in the 3D lookup table reading stage and the optimal segmentation threshold obtained in the optimal segmentation threshold determination stage, the points that need to be classified are determined. In the k The decision weights obtained from the three-dimensional lookup table are as follows: like (15) but It equals 1; like (16) but Equal to 0; In equations (15) and (16), For the first k The optimal segmentation threshold for a three-dimensional lookup table. For point According to the k The decision weights are obtained from a three-dimensional lookup table; Repeatedly reading the 3D lookup table stage and the threshold segmentation stage, targeting K A three-dimensional lookup table is obtained. K Each judgment weight; (S230) The weight combination stage includes: Obtained during the threshold segmentation stage K A three-dimensional lookup table K Each decision weight has a decision coefficient, and the confidence level is obtained using a linear combination method, as follows: （17） In equation (17), The coefficient of determination can be any real number; To determine the weight; For confidence level, ; The confidence scores are normalized as follows: （18） In equation (18), For the final confidence level, ; D For confidence level, This is the minimum value in the confidence level. ; This represents the maximum confidence level. ; The final confidence level is the number of points that need to be classified. To determine whether a cloud belongs to a multi-layered cloud, the closer the final confidence level is to 0, the better the determination. The more likely it is to contain a single layer of cloud, the closer the final confidence level is to 1 for a judgment. It is more likely to contain multiple layers of clouds.

2. The method according to claim 1, characterized in that, During the stage of creating the 3D lookup table, different numbers of nodes result in different resolutions and thus different precisions of the 3D lookup table.

3. The method according to claim 2, characterized in that, In the stage of creating the three-dimensional lookup table, if ≥50, resulting in a high-precision 3D lookup table; if If the value is less than 50, the resulting 3D lookup table will be a low-precision 3D lookup table.

4. The method according to claim 3, characterized in that, In the stage of reading the 3D lookup table, all 3D lookup tables are read to obtain the points that need to be classified. The decision weights in all 3D lookup tables, where the point is in a high-precision 3D lookup table. If the value is null, a low-precision three-dimensional lookup table will be used to find the result. Replace or use the nearest neighbor method to find the nearest belonging point in the high-precision 3D lookup table.

5. The method according to claim 1, characterized in that, During the stage of creating the 3D lookup table, the creation of the 3D lookup table is updated according to the creation time of different training datasets.

6. The method according to claim 5, characterized in that, The three-dimensional lookup table can be uploaded to the satellite to directly determine whether it is a multi-layered cloud; or the satellite observation data can be transmitted to the ground observation center and combined with the three-dimensional lookup table to determine whether it is a multi-layered cloud.

7. The method according to claim 1, characterized in that, The final confidence level is linearly related to the detection accuracy of the final confidence level.

Images