A method for assimilating FY-4A geostationary satellite to identify aerosols under non-clear sky conditions
Through the integrated model of FY-4A stationary satellite with CALIPSO VFM and ERA-5 data, the problem of aerosol recognition under non-clear sky conditions is solved, and aerosol classification products with high spatiotemporal resolution is realized, which improves identification accuracy and coverage. , overcome the problems of low spatial coverage and discontinuity of time of traditional methods.
Patent Information
- Application Number
- CN202310691446.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-13
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-06-13
AI Technical Summary
The prior art is difficult to accurately identify aerosols under non-clear sky conditions, especially in cloud-covered areas. Traditional methods cannot meet the needs of aerosol observations, and there are problems of low spatial coverage and discontinuity of time.
Using the integrated model of FY-4A stationary satellite data, CALIPSO VFM products and ERA-5 data, the atmospheric characteristics and aerosol classification integration model are constructed to achieve high spatial and temporal resolution aerosols Classified products.
Effective identification and classification of aerosols under non-clear sky conditions were achieved, identification accuracy and coverage were improved, and a full coverage aerosol classification product with high temporal and spatial resolution was generated, reducing artificial labeling errors, and increasing sample size.
Smart Images

Figure CN117911884B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of atmospheric pollutant monitoring, and in particular relates to a method for assimilating my country's FY-4A geostationary satellite to identify aerosols under non-clear sky conditions. Background Art
[0002] Atmospheric aerosols can directly affect the energy budget of the atmosphere by absorbing and scattering solar shortwave radiation and the earth's longwave radiation, and can also indirectly affect it by acting as cloud condensation nuclei. Therefore, atmospheric aerosols play a vital role in global climate, cloud microphysical properties and hydrological cycle. The sources of atmospheric aerosols are mainly divided into natural and anthropogenic. East Asia has complex geographical conditions and diverse natural aerosol sources. Anthropogenic aerosol emissions have accounted for an increasing proportion in recent decades. East Asia has become a key area for aerosol research. On the other hand, different forms, particle size distributions and chemical compositions of aerosols often produce different optical properties. These optical properties determine the earth's radiation budget and the extent of the impact of aerosols on the global climate. It can be seen that it is particularly important to accurately identify and classify aerosols, which not only helps to understand the source, transformation and feedback mechanism of aerosols, but also helps to improve the accuracy of satellite retrieval and quantify the radiation impact of aerosols on climate.
[0003] In previous related work, a variety of methods have been proposed to achieve aerosol classification through ground-based remote sensing (such as AERONET) and polar-orbiting satellite remote sensing (such as MODIS) data, such as aerosol optical properties parameters (such as aerosol optical depth (AOD), single scattering albedo (SSA), The aerosol type is unsupervisedly classified and regressed based on the aerosol index (AE) and aerosol relative optical depth (AROD). These threshold-based aerosol classification methods have been applied to detect aerosol types worldwide.
[0004] Due to the limitation of spatial resolution and the influence of aerosol uncertainty, it is difficult for traditional regression models and physical retrieval methods to fully estimate aerosol information. In recent years, as various ML algorithms have been gradually applied to satellite remote sensing aerosol related research, such as Random Forest (RF), vector machine (SVR), convolutional neural network (CNN), etc., machine learning algorithms have shown great advantages in computational efficiency, accuracy, coordination, flexibility, ease of use, etc. At present, in terms of machine learning, relevant studies have been conducted to predict PM 2.5 、PM 10, AOD and aerosol height, as well as the identification and classification of aerosols, machine learning models have unique advantages in solving the challenging problems of aerosol remote sensing. Some scholars have achieved aerosol classification through the MODIS multi-band dataset based on the random forest algorithm, and used the AERONET aerosol type dataset as the target vector, with an overall accuracy of 73%; some scholars have used the VIIRS multi-band dataset as the FFNN model variable input, marked the target vector through the CALIPSOVertical Feature Mask product, and classified and identified dust aerosols. The accuracy of this model on land, ocean and global scales is 81%, 89% and 85% respectively.
[0005] Due to the low spatial coverage of ground-based remote sensing (such as AERONET), many areas lack ground observation data; MODIS and VIRRS sensors are both carried on polar-orbiting satellites, which pass through the region less frequently every day, and the data products obtained have low temporal resolution. Aerosol events have a short life cycle and a fast rate of change. Aerosol classification products generated based on ground-based remote sensing and polar-orbiting satellite data cannot meet the needs of aerosol observations. Therefore, a set of aerosol classification products with long time series and continuous observation data will be beneficial to aerosol research. Summary of the invention
[0006] In order to solve the shortcomings and deficiencies of the prior art in the above-mentioned background technology, the object of the present invention is to provide a method for assimilating the FY-4A geostationary satellite to identify aerosols under non-clear sky conditions.
[0007] The present invention is implemented as follows: a method for assimilating FY-4A geostationary satellite to identify aerosols under non-clear sky conditions, the method comprising the following steps:
[0008] (1) Use CALIPSO VFM products to label CALIPSO on-orbit pixels and generate labels for atmospheric characteristics and aerosol types;
[0009] (2) The grid sample data converted from FY-4A TOAR data and ERA-5 data were spatially and temporally matched with atmospheric characteristics and aerosol type labels to generate on-orbit sample data sets for atmospheric characteristics classification model and aerosol classification model respectively;
[0010] (3) The corresponding integrated model is constructed based on the on-orbit dataset and labels of atmospheric characteristics. The regional grid dataset is input into the atmospheric characteristics classification integrated model for identification and classification, and cloud-free aerosol areas and cloud-aerosol areas are detected.
[0011] (4) Based on the on-orbit dataset and labels of aerosol types, aerosol classification integrated models for cloud-free and cloud-covered areas are constructed respectively. The cloud-free aerosol areas and cloud-covered aerosol areas detected by the atmospheric feature classification integrated model are identified and classified respectively, and the aerosol types of the corresponding areas are detected.
[0012] Preferably, in step (1), the label includes labels of atmospheric characteristic classification labels and aerosol type labels.
[0013] Preferably, the step (2) comprises the following specific steps:
[0014] (2-1) The FY-4A TOAR dataset was projected into a plane coordinate system, and the spatial resolution of the ERA-5 data was adjusted to 4 km, the same as that of FY-4A, by bilinear interpolation, and then merged into grid sample data;
[0015] (2-2) In the spatiotemporal matching between the grid sample data and the CALIPSO on-orbit pixels, the FY-4A and CALIPSO data with an orbital time interval within 30 minutes are selected. Each CALIPSO orbit pixel is matched with a grid pixel with the nearest longitude and latitude to generate an on-orbit sample data set.
[0016] Preferably, in step (2), the grid sample data set is expressed by the formula:
[0017] Grid i,j =TOAR m,i,j +RH h,i,j +T h,i,j +Z h,i,j +U h,i,j +V h,i,j +H i,j +LON i,j +LAT i,j +Time i,j +LH i,j ····(1)
[0018] The on-orbit sample dataset is represented as:
[0019] Input i,j =TOAR m,i,j +RH h,i,j +T h,i,j +Z h,i,j +U h,i,j +V h,i,j +H i,j +LON i,j +LAT i,j +Time i,j +LH i,j ····(2)
[0020] Among them, Input represents the on-orbit dataset that matches the CALIPSO orbit, Input is the model input variable, Grid represents the grid dataset of the entire study area, i represents the position of the grid, j represents the time, m represents the channel number, and h represents the vertical height of 100hpa / 500hpa / 850hpa;
[0021] The model inputs include: relative humidity RH, temperature T, wind speed U / V, high vegetation index LH, altitude H, time Time: 0 to 365 days, longitude LON, and latitude LAT.
[0022] Preferably, the step (3) comprises the following specific steps:
[0023] (3-1) Divide the validation data set and the training data set in a ratio of 3:7, and select the optimal variables of the model through the sequential forward selection method SFS, such as formula (3) to build an integrated voting model;
[0024] Ensemble Model n =f voting (ET n +RF n +XGB n +LGT n +CAT n )···(3)
[0025] Among them, fvo ting represents the soft voting process, n represents the type of model, which are the atmospheric feature classification integrated model, the aerosol classification integrated model in cloud-free areas, and the aerosol classification integrated model in cloud-covered areas. The dependent variables are five submodels: Random-Forest (RF), Extra-Trees (ET), XGBoost (XGB), LightGBM (LGT) and CatBoost (CAT). Note: Both the atmospheric feature classification integrated model and the aerosol classification integrated model are composed of five submodels;
[0026] (3-2) The integrated model input variables are matched and trained with the target vector to construct an integrated model for atmospheric feature classification, as shown in formula (4):
[0027] Class n,i,j =f Ensemble Model (Grid i,j )·········(4)
[0028] Among them, f represents the integrated model, n represents the type of integrated model, i represents the position of the grid, j represents the time, Class represents the recognition result of the model, and the atmospheric feature classification results are: clean atmosphere-cloud area-aerosol in cloudless area-aerosol in cloud area, and the aerosol classification results are: dust-polluted dust-non-dust aerosol-mixed aerosol;
[0029] (3-3) The regional grid data set is input into the atmospheric characteristic integration model for identification and classification, and aerosol areas including cloud-free areas and cloud-covered areas are detected.
[0030] Preferably, the step (4) comprises the following specific steps:
[0031] (4-1) The steps of constructing the aerosol classification model are similar to those of the atmospheric feature classification model. The validation data set and the training data set are divided into a ratio of 3:7, and the features are screened by the sequential forward selection method SFS. As shown in formulas (3) to (4), the aerosol classification integration model for cloud-free areas is constructed by the on-orbit data set and corresponding labels of cloud-free areas, and the aerosol classification integration model for cloud-covered areas is constructed by the on-orbit data set and corresponding labels of cloud-covered areas.
[0032] (4-2) In step (3-3), the aerosol area information of the cloud-free area and the cloudy area has been retrieved by the atmospheric characteristic classification model. In this step, the corresponding grid data of these two areas are respectively input into the aerosol classification models of the cloud-free area and the cloudy area to detect the aerosol type of the corresponding area.
[0033] The present invention overcomes the shortcomings of the prior art and provides a method for assimilating the FY-4A geostationary satellite to identify aerosols under non-clear sky conditions. Fengyun-4A (FY-4A) is the first satellite of the second generation of geostationary meteorological satellites, which can cover the entire territory and is equipped with a multi-channel scanning imaging radiometer (AGRI). It can image the entire disk 205 times a day and image the East Asian region 40 times, which can provide sufficient data support for the present invention. In addition, clouds will interfere with satellite remote sensing observations, so there are many missing values in satellite data in cloud-covered areas. Therefore, in current related studies, most of the remote sensing satellite data are first subjected to cloud screening. However, using the data after cloud screening, not only can the aerosol type in the cloud area not be identified, but also the aerosol contribution in the cloud area cannot be estimated. Therefore, the present invention develops a method for realizing aerosol classification based on an integrated model for the FY-4A TOAR data set, using the CALIPSO VFM product to mark the aerosols and their types in cloudless and clouded areas, and introducing meteorological elements and surface information from ERA-5 as model input supplements. Finally, a full-coverage aerosol classification product with high temporal and spatial resolution is generated, which has the ability to effectively identify aerosols under non-clear sky conditions.
[0034] Compared with the shortcomings and deficiencies of the prior art, the present invention has the following beneficial effects:
[0035] (1) Compared with the traditional aerosol threshold algorithm, the machine learning algorithm has more advantages, and the integrated model of the present invention combines the advantages of multiple machine learning models, is more stable and has higher accuracy; in addition, compared with the method of manually labeling target vectors in some ML studies, the present invention uses the CALIPSO VFM product to assist in labeling, which can greatly improve work efficiency, reduce subjective errors in manual labeling, and help increase the sample size. The sample size in each season reaches more than 400,000;
[0036] (2) Previous aerosol classification studies mostly used ground-based remote sensing and polar-orbiting satellite data, which had problems such as low spatial coverage and temporal discontinuity. The FY-4A TOAR dataset used in this paper can cover the entire territory, with higher temporal and spatial resolutions. The aerosol classification products obtained by building an integrated model based on this data have higher temporal and spatial resolutions, which is more conducive to long-term continuous monitoring. The product has a temporal resolution of 1 hour and a spatial resolution of 4 km.
[0037] (3) Most of the current aerosol classification studies do not involve cloud-covered areas, and only identify cloudless areas. The present invention obtains a fully covered aerosol classification product based on an integrated model, and realizes the effective identification of aerosols and their types under non-clear sky conditions. Among them, the accuracy of the atmospheric characteristic classification model (four seasons) is 0.853, 0.825, 0.861, and 0.872, respectively; the accuracy of the aerosol classification model is 0.92, 0.91, 0.87, and 0.84 (cloudless area) and 0.89, 0.88, 0.85, and 0.84 (cloudy area). BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of the structure of the integrated voting model of the present invention;
[0039] Figure 2 It is a schematic diagram of the construction and prediction of the integrated voting model of the present invention;
[0040] Figure 3 It is an evaluation table of the on-orbit recognition performance of the atmospheric feature classification integrated model;
[0041] Figure 4 It is the on-orbit recognition performance evaluation table of the aerosol classification ensemble model in cloud-free areas;
[0042] Figure 5 It is a table for evaluating the on-orbit recognition performance of the integrated aerosol classification model in cloudy areas. DETAILED DESCRIPTION
[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0044] The embodiment of the present invention discloses a method for assimilating a FY-4A geostationary satellite to identify aerosols under non-clear sky conditions, the method comprising the following steps:
[0045] (1) Figure 2 As shown in step 1, the CALIPSO VFM product is used to label the CALIPSO on-orbit pixels and generate labels for atmospheric characteristics and aerosol types.
[0046] Fengyun-4A (FY-4A) is the first satellite of the second generation of geostationary meteorological satellites. One of its main optical payloads, the multi-channel scanning imaging radiometer AGRI (Advanced Geosynchronous Radiation Imager), has 2 visible light channels (0.47, 0.65μm), 1 near-infrared channel (0.83μm), 3 short-wave infrared channels (1.38, 1.61 and 2.23μm) and 8 mid-infrared channels. The present invention obtained the FY-4A Level1 TOAR full disk daytime data set (70°-140°E, 15-55°N) from March 1, 2018 to February 28, 2021 from the National Satellite Meteorological Center (NSMC).
[0047] In addition, the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite was successfully launched in April 2006, carrying the Orthogonal Polarization Cloud-Aerosol Lidar (CALIOP), which is the first satellite-borne Lidar system to provide global aerosol observations. Among them, the V4 Vertical Feature Mask product can provide a set of characteristic classification marks for atmospheric vertical profiles, including atmospheric feature types, cloud types, aerosol types and subtype data (Pan et al., 2019). This product is used to mark pixels of cloud-aerosols and multiple aerosol types. The present invention obtained the Vertical Feature Mask data product from NASA during the day from March 1, 2018 to February 28, 2021.
[0048] At the same time, passive remote sensing sensors are also susceptible to cloud interference during detection, and there may be uncertainty in using TOAR data alone to identify aerosols. Existing studies have shown that the formation and propagation of cloud-aerosols are closely related to factors such as wind speed, relative humidity, altitude, and surface type. The meteorological parameters and geographic information used in the present invention are from the ERA-5 (ECMWF Reanalysis v5) reanalysis data set, mainly including phase surface pressure (SP, Pa), high and low vegetation index (LL, LH), relative humidity (%), temperature (K), wind vector (U / V, m / s) and surface type data (LC).
[0049] In the embodiment of the present invention, the atmospheric characteristics and aerosol type data (-0.5km-20.2km) of the CALIPSO VFM product are used to mark the CALIPSO track pixels. It should be noted that thick clouds may cause the CALIPSO sensor receiving signal to attenuate and there may be missing values. In this case, since it is impossible to determine whether there are aerosols in the undetected area under the clouds, this part of the missing data is not used.
[0050] In step (1), the atmospheric feature classification labels are first labeled. The "cloud" and "Tropospheric aerosol" data in the Feature Type of the VFM product are selected to label each track pixel, and finally all samples are labeled as four types of pixels: "clean atmosphere-cloud layer-aerosol in cloudless areas-aerosol in cloudy areas". Secondly, the aerosol type labels are labeled. The five aerosol type data of Dust, Polluted continental, and Polluted dust in the Aerosol sub-type are used to label the aerosol types in "aerosol in cloudless areas-aerosol in cloudy areas", which are divided into "sand dust-polluted sand dust-non-sand dust aerosol-mixed aerosol" (Note: Since the present invention only targets land aerosols in the troposphere, the marine aerosol data in the VFM product is not used).
[0051] (2) Figure 2 As shown in step 2, the grid sample data converted from FY-4A TOAR data and ERA-5 data are temporally and spatially matched with atmospheric characteristics and aerosol type labels to generate on-orbit sample data sets for atmospheric characteristics classification model and aerosol classification model respectively.
[0052] In step (2), the grid sample data set is expressed by the formula:
[0053] Grid i,j =TOAR m,i,j +RH h,i,j +Th,i,j +Z h,i,j +U h,i,j +V h,i,j +H i,j +LON i,j +LAT i,j +Time i,j +LH i,j ····(1)
[0054] The on-orbit sample dataset is represented as:
[0055] Input i,j =TOAR m,i,j +RH h,i,j +T h,i,j +Z h,i,j +U h,i,j +V h,i,j +H i,j +LON i,j +LAT i,j +Time i,j +LH i,j ···(2)
[0056] Among them, Input represents the on-orbit dataset matching the CALIPSO orbit, Input is the model input variable, Grid represents the grid dataset of the entire study area, i represents the position of the grid, j represents the time, m represents the channel number, and h represents the vertical height (100 hpa / 500 hpa / 850 hpa).
[0057] The model inputs include: relative humidity (RH), air temperature (T), wind speed (U / V), high vegetation index (LH), altitude (H), time (Time: 0-365), longitude (LON), and latitude (LAT).
[0058] In the embodiment of the present invention, step (2) includes the following specific steps:
[0059] (2-1) The FY-4A TOAR dataset is projected into a plane coordinate system, and the spatial resolution of the ERA-5 data is adjusted to 4 km, the same as that of FY-4A, by bilinear interpolation, and then merged into a grid sample data set.
[0060] In step (2-1), the FY-4A TOAR dataset is projected into a plane coordinate system, and the spatial resolution of the ERA-5 data is adjusted to 4 km, the same as that of FY-4A, by bilinear interpolation, and then merged into grid sample data (as shown in Formula 2).
[0061] (2-2) In the spatiotemporal matching between the grid sample data and the CALIPSO on-orbit pixels, the FY-4A and CALIPSO data with an orbital time interval within 30 minutes are selected. Each CALIPSO orbit pixel is matched with a grid pixel with the nearest longitude and latitude to generate an on-orbit sample data set.
[0062] In step (2-2), since the trajectory of CALIPSO on-orbit data is a curve on a two-dimensional plane, in order to match it with the sample data distributed in a grid, the grid sample data needs to be matched to the curve data of the same trajectory. Therefore, in the spatiotemporal matching of the grid sample data (4km x 4km) and CALIPSO on-orbit data (VFM product: resolution 5km), data with an orbital time difference within 30 minutes are selected for matching, and each CALIPSO orbit pixel is matched with a grid pixel with the nearest longitude and latitude to generate an on-orbit sample data set (consistent with the CALIPSO orbital position).
[0063] It is worth mentioning that the present invention uses the CALIPSO VFM product dataset from March 2018 to February 2021, so there is sufficient data support. Overall, the satellite trajectories of CALIPSO data are spread across the study area, and the on-orbit dataset generated by matching with the CALIPSO VFM product can cover the entire study area and contain data from multiple time periods in the same area. The prediction results of the integrated model constructed in this way are scientific, practical, accurate, flexible, and adaptable.
[0064] (3) Basis Figure 2 In the subsequent process of step 2, the corresponding integrated model is constructed based on the on-orbit dataset and labels of atmospheric characteristics. The structure of the integrated model is as follows: Figure 1 As shown in the figure, the grid data set of the region is input into the atmospheric feature classification integrated model for identification and classification, and cloud-free aerosol areas and cloud-aerosol areas are detected.
[0065] In step (3), the present invention uses a variety of machine learning algorithms to build an integrated voting model, such as Figure 1 As shown in the figure, it includes Random-Forest, Extra-Trees, XGBoost, LightGBM and CatBoost. The integrated voting model uses the soft voting method. Five sub-models are constructed by the above five machine learning algorithms. Each sub-model outputs the classification probability of the sample. After averaging the classification probabilities of all models, the category with the highest probability is adopted.
[0066] In addition, the model is trained and identified as:
[0067] Ensemble Model n =f voiing (ETn +RF n +XGB n +LGT n +CAT n )···(3)
[0068] Among them, f voting represents the soft voting process, n represents the type of ensemble model (atmospheric feature classification ensemble model / aerosol classification ensemble model in cloudless area / aerosol classification ensemble model in cloudy area), and the dependent variables are five submodels: Random-Forest (RF), Extra-Trees (ET), XGBoost (XGB), LightGBM (LGT) and CatBoost (CAT).
[0069] Class n,i,j =f EnsembleModel (Grid i,j )·······(4)
[0070] Among them, f represents the integrated model, n represents the type of the integrated model, i represents the position of the grid, j represents the time, and Class represents the recognition result of the model.
[0071] The validation data set and the training data set are divided into a ratio of 3:7, as shown in formula (3), and the atmospheric feature classification integration model is constructed. The optimal variables for the atmospheric feature classification integration model are selected by the sequential forward selection method (SFS). As shown in formula (4), the input variables of the atmospheric features are matched and trained with the target vector, and the atmospheric feature classification integration model is constructed. The entire study area is searched and divided into four types of areas: clean atmosphere-cloud area-aerosol in cloudless area-aerosol in cloudy area, and aerosol areas (including cloudless area and cloudy area) are selected.
[0072] (4) Figure 2 As shown in step 3, aerosol classification integrated models for cloud-free areas and cloud-covered areas are constructed based on the on-orbit dataset and labels of aerosol types. The cloud-free aerosol areas and cloud-covered aerosol areas detected by the atmospheric feature classification integrated model are identified and classified, and the aerosol types of the corresponding areas are detected.
[0073] In step (4), the method of constructing the aerosol classification integrated model is similar to that of the atmospheric characteristic classification integrated model, as shown in formulas (1) to (4). The variable inputs of the aerosol classification model for the cloud-free area and the cloud-covered area are respectively selected by the sequential forward selection method (SFS), and the aerosol classification integrated model for the corresponding area is constructed respectively. The structure of the aerosol classification integrated model is consistent with that of the atmospheric characteristic classification integrated model (e.g. Figure 1Based on the aerosol classification integrated model, the aerosol areas detected by the atmospheric feature classification integrated model are identified and classified, and the aerosol types are divided into dust-polluted dust-non-dust aerosol-mixed aerosol, and finally a fully covered aerosol classification product is generated.
[0074] The above implementation was carried out based on on-orbit data, and the test results obtained are as follows Figures 3 to 5 shown.
[0075] Figure 3 The on-orbit recognition performance of the atmospheric feature classification integrated model was demonstrated. The average accuracy of the model in the four seasons was 85%, with the winter model having the highest accuracy of 87% and the summer model having the lowest accuracy of 82%. Overall, the atmospheric feature classification integrated model has the best recognition effect on clean atmosphere samples, with F1 values of 0.89, 0.85, 0.92, and 0.93 respectively; the recognition performance of aerosol samples in cloudy areas is slightly worse, with F1 values of 0.76, 0.77, 0.75, and 0.74 respectively.
[0076] Figure 4 The on-orbit recognition performance of the integrated aerosol classification model in cloud-free areas was demonstrated, with an overall accuracy of 0.89. The spring model had the highest accuracy, with an accuracy of 0.92, while the winter model had the lowest accuracy, with an accuracy of 0.84. Overall, dust had the best recognition results, with an average F1 of 0.958 for all four seasons. The recognition results for mixed aerosols were slightly worse, with an average F1 of 0.765 for all four seasons.
[0077] Figure 5 The on-orbit recognition performance of the integrated aerosol classification model in cloudy areas was demonstrated. Although affected by clouds, the model still had good recognition capabilities, with an overall accuracy of 87%. Among them, the spring model had the highest accuracy, with an accuracy of 0.89, and the winter model had the lowest accuracy, with an accuracy of 0.84. In the recognition of different aerosols, the model had the best effect on dust, with an average F1 of 0.94 in all four seasons. The model had a good recognition effect on polluted dust, with an average F1 of 0.86 in all four seasons. The model had the worst recognition effect on mixed aerosols, with an average F1 of 0.67 in all four seasons.
[0078] In order to understand the retrieval capability of the atmospheric characteristic classification integrated model and the aerosol classification integrated model outside the CALIOP orbit, this study was based on the sandstorm event on October 21, 2020, and was evaluated and verified in combination with CALIPSO VFM, Himawari-8AOD and other products.
[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for assimilating FY-4A geostationary satellite to identify aerosols under non-clear sky conditions, characterized in that: The method comprises the following steps: (1) Use CALIPSO VFM products to label CALIPSO on-orbit pixels and generate labels for atmospheric characteristics and aerosol types; (2) The grid sample data converted from FY-4A TOAR data and ERA-5 data were spatially and temporally matched with atmospheric characteristics and aerosol type labels to generate on-orbit sample data sets for atmospheric characteristics classification model and aerosol classification model respectively; (3) The corresponding integrated model is constructed based on the on-orbit dataset and labels of atmospheric characteristics. The regional grid dataset is input into the atmospheric characteristics classification integrated model for identification and classification, and cloud-free aerosol areas and cloud-aerosol areas are detected. (4) Based on the on-orbit dataset and labels of aerosol types, aerosol classification integrated models for cloud-free and cloud-covered areas are constructed. The cloud-free and cloud-covered aerosol areas detected by the atmospheric feature classification integrated model are identified and classified, and the aerosol types of the corresponding areas are detected. The step (3) comprises the following specific steps: (3-1) Divide the validation data set and the training data set in a ratio of 3:7, and select the optimal variables of the model through the sequential forward selection method SFS, such as formula (3) to build an integrated voting model; Ensemble Model n =f voting (ET n +RF n +XGB n +LGT n +CAT n )……………………………(3) Among them, f voting represents the soft voting process, n represents the type of model, which are the atmospheric feature classification integrated model, the aerosol classification integrated model in cloud-free areas, and the aerosol classification integrated model in cloud areas. The dependent variables are five submodels: RF, ET, XGB, LGT, and CAT. RF represents Random-Forest, ET represents Extra-Trees, XGB represents XGBoost, LGT represents LightGBM, and CAT represents CatBoost. Both the atmospheric feature classification integrated model and the aerosol classification integrated model are composed of five submodels. (3-2) Match and train the integrated model input variables with the target vector to build an integrated model for atmospheric feature classification, as shown in formula (4): Class n,i,j =f EnsembleModel (Grid i,j )……………………………(4) Among them, Grid represents the grid dataset of the entire study area, f represents the integrated model, n represents the type of the integrated model, i represents the location of the grid, j represents the time, Class represents the recognition result of the model, and the atmospheric feature classification results are: clean atmosphere-cloud area-aerosol in cloudless area-aerosol in cloud area, and the aerosol classification results are: dust-polluted dust-non-dust aerosol-mixed aerosol; (3-3) Inputting the regional grid data set into the atmospheric feature classification integrated model for identification and classification, and detecting the aerosol areas including cloud-free areas and cloud-covered areas; The step (4) comprises the following specific steps: (4-1) The validation dataset and the training dataset are divided into a ratio of 3:7, and the features are screened by the sequential forward selection method SFS, as shown in formulas (3) to (4). The aerosol classification ensemble model for the cloud-free area is constructed by the on-orbit dataset and the corresponding labels of the cloud-free area, and the aerosol classification ensemble model for the cloud-covered area is constructed by the on-orbit dataset and the corresponding labels of the cloud-covered area; (4-2) In step (3-3), the aerosol area information of the cloud-free area and the cloudy area has been retrieved by the atmospheric characteristic classification model. In this step, the corresponding grid data of these two areas are respectively input into the aerosol classification models of the cloud-free area and the cloudy area to detect the aerosol type of the corresponding area.
2. The method according to claim 1, characterized in that In step (1), the labels include labels of atmospheric characteristic classification labels and aerosol type labels.
3. The method according to claim 1, characterized in that The step (2) comprises the following specific steps: (2-1) The FY-4A TOAR dataset was projected into a plane coordinate system, and the spatial resolution of the ERA-5 data was adjusted to 4 km, the same as that of FY-4A, by bilinear interpolation, and then merged into grid sample data; (2-2) In the spatiotemporal matching between the grid sample data and the CALIPSO on-orbit pixels, the FY-4A and CALIPSO data with an orbital time interval within 30 minutes are selected. Each CALIPSO orbit pixel is matched with a grid pixel with the nearest longitude and latitude to generate an on-orbit sample data set.
4. The method according to claim 1, characterized in that In step (2), the grid sample data set is expressed by the formula: Grid i,j =TOAR m,i,j +RH h,i,j +T h,i,j +Z h,i,j +U h,i,j +V h,i,j +H i,j +LON i,j +YEARS i,j +Time i,j +LH i,j ····(1) The on-orbit sample dataset is represented as: Input i,j =TOAR m,i,j +RH h,i,j +T h,i,j +Z h,i,j +U h,i,j +V h,i,j +H i,j +LON i,j +LAT i,j +Time i,j +LH i,j ···(2) Among them, Input represents the on-orbit dataset that matches the CALIPSO orbit, Input is the model input variable, Grid represents the grid dataset of the entire study area, i represents the position of the grid, j represents the time, m represents the channel number, and h represents the vertical height of 100hpa / 500hpa / 850hpa; The model inputs include: relative humidity RH, temperature T, wind speed U / V, high vegetation index LH, altitude H, time Time: 0 to 365 days, longitude LON, and latitude LAT.
Citation Information
Patent Citations
Sea fog detection method based on MODIS satellite data
CN114022782A
Night aerosol optical thickness estimation method and system based on ground monitoring data
CN115081557A
Systems and methods for predicting outcomes and conditions of chemical reactions with high reliability based on a highly diverse and accurate dataset
US20230131234A1
