Multi-channel cloud physical property inversion algorithm based on radiation patterns and machine learning

Abstract

This invention belongs to the field of cloud physical property inversion technology, specifically a cloud physical property inversion algorithm based on radiation patterns and machine learning. This invention inverts the physical properties (CTH, COT, and CER) of single-layer and multi-layer clouds from the visible light and infrared multi-channels of the Himawari 8 satellite AHI imager. The specific steps include: using the radiative transfer mode (ERTM) to simulate the cloud reflectivity / brightness temperature of each channel of the AHI imager under different cloud conditions and the clear-sky brightness temperature of the thermal infrared channel, constructing a cloud physical property—multi-channel simulated radiative value dataset; using a neural network, pre-training a cloud remote sensing inversion model with the dataset simulated by the radiation pattern; constructing a real satellite observation dataset using active radar observations as standard values, and using transfer learning methods to retrain and optimize the pre-trained cloud inversion model with the real satellite observation dataset; ultimately achieving higher accuracy inversion of cloud physical properties. This invention is more accurate and precise than traditional cloud inversion methods.

Description

Technical Field

[0001] This invention belongs to the field of cloud physical property inversion technology, specifically involving a visible light and infrared multi-channel cloud physical property inversion algorithm. Background Technology

[0002] With the widespread application of satellite data, cloud remote sensing inversion algorithms have become increasingly mature. The most representative cloud inversion method is that of Nakajima and King (1990). [3] A dual-channel visible light (VIS) / shortwave infrared (SWIR) scheme is proposed. This scheme is based on the fact that clouds have a lower absorption coefficient in the VIS channel, and their reflectivity is more sensitive to cloud optical thickness (COT). In the SWIR channel, the absorption coefficient is higher, and the reflectivity is primarily sensitive to the cloud particle effective radius (CER). Therefore, this scheme utilizes both VIS and SWIR channels to simultaneously invert COT and CER. Currently, this method has been widely applied to the inversion of operational cloud optical properties in spaceborne spectroscopic imagers. The joint inversion of cloud microphysical properties using VIS and SWIR channels is particularly effective for clouds with a COT greater than 1. [4] However, this does not apply to thin ice clouds with a COT less than 1. (Inoue (1985)) [5] The proposed infrared split-window method has a significant advantage in inverting ice clouds with COT values ​​between 0.1 and 5. However, the infrared split-window algorithm cannot accurately distinguish between thin high clouds and thick low clouds, and it requires prior determination of cloud and surface temperatures and background atmospheric information. This has spurred the development of cloud inversion algorithms using multiple infrared channels. (Wang et al., 2011) [6] The COT and CER of cirrus clouds were retrieved using the 8.5μm, 11μm, and 12μm thermal infrared (TIR) ​​channels of the MODIS sensor. Atmospheric temperature and water vapor profiles in the algorithm were derived from MERRA atmospheric reanalysis data, while CTH was obtained from CALIOP (Cloud-Aerosol Lidar with Orthogonal Polarization) and CPR (Cloud Profiling Radar) active radar terrestrial observations. This results in significant spatial limitations in the implementation of the algorithm.

[0003] Most current passive cloud remote sensing inversion algorithms are based on the single-layer cloud assumption. However, when multiple layers of clouds (referring to upper ice clouds covering lower water clouds) exist, this assumption can cause significant errors in the inversion of cloud physical properties. [7] Studies show that the probability of multi-layered clouds occurring globally is 25-28% on average; therefore, the inversion of the physical characteristics of multi-layered clouds should be considered in cloud remote sensing. (Iwabuchi et al., 2017) [8]For AHI imagers, a cloud inversion algorithm combining thermal infrared multi-channels has been further developed, applicable to multi-layered clouds. Due to the limited penetration of the thermal infrared channel, this algorithm assumes that the COT and CER of lower-level water clouds are constant. Based on this, it can invert the cloud physics characteristics of thin ice clouds overlapping water clouds and the cloud top characteristics of lower-level water clouds. (Teng et al., 2020) [9] Using the optimal estimation method, a two-layer cloud COT and CER inversion algorithm based on the shortwave channels (0.87, 1.61, 2.13, and 2.25 μm) of MODIS and VIIRS imagers was proposed. The COT of the upper ice cloud inverted by this algorithm is in good agreement with CALIOP observations. However, since the algorithm only uses the shortwave channels, there is some uncertainty in the inversion results for thin clouds. Although the optimal estimation method can basically realize the inversion of the physical properties of multi-layer clouds, it is very time-consuming and requires repeated iterations. Currently, it can only be used in the scientific research stage and cannot be widely used in remote sensing operations. In addition, physical methods based on radiative transfer models often use some approximate assumptions, such as the assumption of vertical homogeneity of clouds.

[10] The approximate assumption of planar parallel atmospheric radiative transfer (ignoring three-dimensional radiation effects and Earth's curvature effects).

[11] Simplified assumptions of the ice crystal scattering model

[12] These factors inevitably lead to a decrease in the accuracy of cloud inversion.

[0004] In recent years, the development of machine learning methods has provided new ideas for solving some nonlinear and complex problems in quantitative cloud remote sensing inversion. (2018)

[13] Using the CTH observed by CALIOP as the standard value, a machine learning-based CTH retrieval algorithm was proposed for the infrared channel of the MODIS sensor. Evaluation results show that the algorithm's results are significantly improved compared to the official MODIS cloud product. (Lin et al., 2022)

[14] The gradient boosting decision tree algorithm was used to retrieve cloud base height based on observations from the passive sensor ABI (Advanced Baseline Imager). The algorithm's retrieval results for the cloud base height of a single layer of non-strong convective clouds were largely consistent with CALIOP observations, with a root mean square error of 1.14 km after quality control. However, this algorithm is still not effectively applicable to the retrieval of the physical properties of multi-layered clouds. Importantly, pure machine learning methods often suffer from poor physical interpretability and are highly dependent on the quantity and quality of the training sample set. Summary of the Invention

[0005] The purpose of this invention is to provide a visible light and infrared multi-channel cloud physical property inversion algorithm based on radiation patterns and machine learning, so as to achieve higher accuracy inversion of cloud physical properties.

[0006] Currently, the widely used VIS / SWIR dual-channel scheme is only effective for clouds with a COT greater than 1, while the infrared window method is only effective for thinner clouds with a COT between 0.1 and 5. Therefore, using the visible infrared channel simultaneously in the inversion algorithm is expected to achieve the inversion of cloud physical characteristics over a larger COT range. On the other hand, most current passive cloud remote sensing inversion algorithms are based on the single-layer cloud assumption, but when multiple layers of clouds exist, this assumption will cause a large error in the inversion of cloud physical characteristics. However, most current multi-channel multi-layer cloud inversion algorithms are based on optimal estimation algorithms, which are very time-consuming and cannot be widely used in remote sensing operations. More importantly, optimal estimation methods based on radiative transfer patterns (physical methods) have some physical approximation assumptions, which greatly limit the accuracy of cloud remote sensing inversion. To address the aforementioned issues, this invention combines radiation mode simulation and machine learning methods to retrieve macroscopic and microscopic physical properties (i.e., cloud top height (CTH), cloud optical thickness (COT), and cloud particle effective radius (CER)) from the visible and infrared multi-channel imagesr of the Himawari 8 satellite AHI imager.

[0007] The visible and infrared multi-channel cloud physical property inversion algorithm based on radiation patterns and machine learning provided by this invention retrieves the cloud top height (CTH), cloud optical thickness (COT), and cloud particle effective radius (CER) of single-layer and multi-layer clouds from visible and infrared multi-channel data of the Himawari 8 satellite AHI using a cloud physical property remote sensing inversion model. The specific steps are as follows:

[0008] (1) Construct a training dataset for the remote sensing inversion model of cloud physical properties;

[0009] First, based on the SeeBor 5.0 atmospheric profile dataset, which can characterize the possible conditions of temperature and absorbing gases (water vapor and ozone) in the actual atmosphere. [1] Within a pre-defined range, cloud physics properties of single-layer and multi-layer clouds, as well as solar / satellite observation geometric parameters, are randomly generated; and combined with atmospheric profile sets, the forward radiative transfer model ERTM is used. [2] The cloud reflectance and brightness temperature of each channel of the AHI imager, as well as the clear-sky brightness temperature of the thermal infrared channel, were simulated under different cloud state assumptions, forming a dataset called "Cloud Physical Characteristics - Multi-channel Simulated Radiation Values". Here, the clear-sky brightness temperature is the brightness temperature simulated by the radiative transfer mode based on the background atmosphere and surface conditions under the assumption of clear, cloudless sky. Using multi-channel clear-sky brightness temperature as a predictor in the cloud remote sensing inversion model can, to some extent, eliminate the influence of background atmosphere and ground conditions, and is more conducive to the inversion of cloud microphysical characteristics. Finally, the simulation dataset was divided into a training set, a validation set, and a test set for use in step (2).

[0010] (2) Construct a remote sensing inversion model of cloud physical properties;

[0011] In this invention, the establishment of the final cloud physical property remote sensing inversion model requires two rounds of training. For the neural network (DNN) model, it is first pre-trained based on a dataset simulated by radiation patterns, and then further post-trained and optimized using a real satellite observation dataset. Pre-training the model on the simulated dataset yields better model initialization and generalization capabilities because the simulated data can cover different atmospheric scenarios and satellite / sun geometry and has clear and interpretable physical meaning.

[0012] A cloud remote sensing inversion model was pre-trained using a neural network (DNN) based on a radiation pattern simulation dataset. The model structure mainly consists of an input layer, a hidden layer, and an output layer. The intermediate hidden layer comprises a series of interconnected one-dimensional convolutional layers, activation layers, dropout layers, and fully connected layers. The one-dimensional convolutional layer acts as a feature extractor, effectively extracting key spectral information of different cloud physical properties from the input values ​​(multi-channel radiance values). Each convolutional layer uses a 5x5 convolutional kernel to generate a set of feature maps. Batch normalization is applied after each convolutional layer to improve the model's robustness and accuracy. The ReLU activation algorithm used in the model is relatively simple and has high learning efficiency. During training, the Adam optimizer is used to adjust the model parameters to minimize the root mean square error of cloud physical property inversion. A dropout rate of 0.5 is used in the hidden layer to avoid overfitting and improve the performance of the neural network model. The neural network model is pre-trained using the simulated dataset generated by the radiation model in step (1). The multi-channel radiation values ​​(brightness temperature / reflectivity and clear sky brightness temperature) are used as inputs, and the corresponding cloud physical characteristics are output. The learning rate is set to decrease exponentially with the increase of training times. The model performance is evaluated on the validation set, thereby adjusting the model hyperparameters to select the optimal model configuration and initially constructing a cloud remote sensing inversion model based on radiation model simulation.

[0013] (3) Further optimization of the remote sensing inversion model of cloud physical characteristics;

[0014] Because the radiation model uses some physical assumptions such as the planar parallel approximation, neglecting three-dimensional radiation effects and Earth curvature, and the assumption of vertical cloud uniformity, there are certain discrepancies between the simulated data and the data from real scenes. To eliminate these discrepancies, this invention will further optimize the cloud remote sensing inversion model. Spaceborne active sensors (CPR cloud radar and CALIOP lidar) can provide reliable cloud vertical profile information. Cloud top characteristics and cloud microphysical properties acquired by active remote sensing are often used as standard values ​​to evaluate passive sensor cloud products. Therefore, active radar observation data is used as the standard value for cloud physical properties. AHI observation data, ERA5 atmospheric reanalysis data, MODIS surface emissivity data, and CALIOP / CPR active radar data are spatiotemporally matched. Then, based on atmospheric reanalysis data and surface parameter data, the ERTM model is used to simulate the clear-sky brightness temperature of each channel of AHI thermal infrared, and combined with the observed reflectivity and brightness temperature of multiple channels to generate a dataset of "cloud physical properties - multi-channel radiation values" based on real satellite observations. Then, based on the real dataset, the cloud inversion model initially established in step (2) is retrained using transfer learning to optimize the model. In transfer learning, the parameters of the first two convolutional layers are fixed, and the remaining network structure parameters are adjusted and optimized. The resulting cloud physical property remote sensing inversion model can achieve higher accuracy inversion of cloud microphysical properties.

[0015] Features and advantages of the present invention:

[0016] This invention combines visible light and infrared channels to effectively invert the cloud physical properties of thin clouds while simultaneously retrieving the cloud physical properties of thick clouds. In other words, this invention integrates the advantages of the VIS / SWIR dual-channel method and the infrared split-window method.

[0017] When multi-layered clouds exist, this invention can simultaneously invert the CTH, COT, and CER of upper-layer ice clouds and lower-layer water clouds in multi-layered clouds, which is more accurate than traditional cloud inversion methods based on the assumption of a single-layered uniform cloud layer.

[0018] This invention first trains and constructs a cloud remote sensing inversion model using a radiation model simulation dataset. Then, it further optimizes and adjusts the initially constructed cloud inversion model using a real satellite observation dataset with active radar observation data as the standard. Pre-training the model on the radiation model simulation dataset yields better model initialization and generalization capabilities because the simulated data can cover different atmospheric scenarios and satellite / sun geometries and has clear and interpretable physical meaning. However, due to the use of some physical assumptions in the radiation model, such as the planar parallel approximation, neglecting three-dimensional radiation effects and Earth curvature, and the assumption of vertical cloud uniformity, there is a certain gap between the simulated data and the data from real-world scenarios.

[0019] This invention utilizes transfer learning to further optimize and adjust the pre-trained model based on real satellite data, ultimately achieving inversion accuracy superior to both pure physical methods and pure machine learning methods based on optimal estimation. Attached Figure Description

[0020] Figure 1 The inversion algorithm framework of this invention includes: (a) the construction of a simulated dataset based on radiation patterns; (b) the construction of a real dataset based on satellite observations; and (c) the configuration of a deep neural network model (DNN).

[0021] Figure 2 This is a case study from 07 UTC on January 8, 2017, located in the eastern Indian Ocean region. (a) AHI false-color images (0.65, 2.3, and 11 μm). (b) Cloud detection and cloud phase classification images (from...).

[15] (cf) Spatial distribution maps of CTH(c,e), COT(d,f) and CER(e,h) retrieved from the DNN model, where (e), (f) and (h) are the distribution maps of CTH, COT and CER of the lower water cloud in the presence of multi-layer clouds, respectively.

[0022] Figure 3 This is a vertical profile of cloud phases along the trajectory line for the CPR / CALIOP joint cloud product. Inversion results from the DNN model, MODIS, and AHI official products are marked with different symbols in the figure. Specifically, the points in the DNN model indicate the location of the lower-level water cloud CTH retrieved by the DNN model in the presence of multi-layered clouds. Detailed Implementation

[0023] (1) To best represent the possible conditions of temperature and absorbing gases in the actual atmosphere, we selected the SeeBor 5.0 atmospheric profile dataset provided by the Meteorological Satellite Research Institute at the University of Wisconsin-Madison. The SeeBor dataset contains 15,704 temperature, water vapor, and ozone profiles from around the world. Then, within a predefined range, cloud physics properties of single-layer and multi-layer clouds, as well as solar / satellite observation geometric parameters, were randomly generated and combined with standard profiles before being input into the forward radiative transfer model ERTM. [2] In the simulation, the cloud reflectivity / brightness temperature of each channel of AHI and the clear sky brightness temperature of the thermal infrared channel are obtained, forming a dataset of "cloud physical characteristics - multi-channel simulated radiance values" (e.g. Figure 1(a) shows the dataset, which is divided into training, validation, and test sets. Specifically, clear-sky brightness temperature is the brightness temperature simulated by the forward radiative transfer model based on the background atmospheric conditions under the assumption of clear, cloudless skies. Using multi-channel clear-sky brightness temperature as a predictor in cloud remote sensing models can, to some extent, eliminate the influence of background atmospheric and ground conditions, and is more conducive to the inversion of cloud microphysical properties.

[0024] (2) First, a pre-trained cloud remote sensing inversion model based on a radiation pattern simulation dataset is used to utilize a neural network (DNN). The DNN structure mainly consists of an input layer, a one-dimensional convolutional layer, an activation layer, a dropout layer, a fully connected layer, and an output layer, as shown in the figure. Figure 1 As shown in (c), the convolutional layer acts as a feature extractor, effectively extracting key spectral information of different cloud physical properties from the input samples. Each convolutional layer uses a kernel of size 5 to generate a set of feature maps. Batch normalization is applied after each convolutional layer to improve the robustness and accuracy of the model. The activation algorithm used in the model is the ReLU function, which is relatively simple and has high learning efficiency. The generated training dataset is then used to train the neural network, using multi-channel radiance values ​​(brightness temperature / reflectivity and clear-sky brightness temperature) as input and the corresponding cloud physical properties as standard values. The learning rate is set to decrease exponentially with increasing training iterations. The network performance is initially evaluated on the validation set, and the network hyperparameters are adjusted to select the optimal model configuration, thus initially constructing a cloud remote sensing inversion model.

[0025] (3) Spaceborne active sensors such as CPR (Cloud Profiling Radar) and CALIOP (Cloud-Aerosol Lidar with Orthogonal Polarization) can provide reliable cloud vertical profile information. Cloud top characteristics and cloud microphysical properties acquired by active remote sensing are often used as standard values ​​to evaluate cloud products from passive sensors. Therefore, in order to further improve and refine the initially established cloud remote sensing inversion model, we use CALIOP / CPR active radar data as the standard values ​​for cloud physical properties. We perform spatiotemporal matching of AHI observation data, ERA5 atmospheric reanalysis data, MODIS surface emissivity data, and CALIOP / CPR active radar data. Then, based on atmospheric reanalysis data, surface parameter data, and AHI observation geometric parameters, we use the radiative transfer model ERTM to simulate the clear-sky brightness temperature of each channel of AHI thermal infrared and combine it with AHI multi-channel observation reflectivity and brightness temperature. In this way, we construct a dataset of "cloud physical properties - multi-channel radiative values" (e.g., cloud top characteristics - multi-channel radiative values) using real satellite observation data. Figure 1(b) shows the cloud inversion model. Based on a real dataset, we use transfer learning to adjust and optimize the cloud inversion model initially established in step (2). In transfer learning, we fix the parameters of the first two convolutional layers and optimize the remaining network structure parameters to achieve higher accuracy inversion of cloud microphysical properties.

[0026] (4) Figure 2 The image shows a case from 07 UTC on January 8, 2017, located in the eastern Indian Ocean region. Figure 2 (a) and (b) present the false-color images synthesized with AHI channels of 0.65, 2.3, and 11 μm, and the cloud detection and classification results, respectively (source:

[15] The image shows a false-color DNN daytime model developed in China. Large areas of high-altitude ice clouds (appearing purple) cover patches of low-altitude undulating stratocumulus clouds (appearing brown) over the western coast of Sumatra and Java. Figure 2 (b) also shows that the area south of 10°S is mainly water clouds, while the area north of 10°S has a large number of ice clouds. These ice clouds overlap with the low-level water clouds to form large areas of multi-layered clouds. Figure 2 (cf) shows the spatial distribution of CTH(c,e), COT(d,f), and CER(e,h) retrieved by the DNN model. Figure 3 The CTH results of the DNN model, the internationally mainstream MODIS and AHI official products, and the CPR / CALIOP joint cloud product were compared. As shown in the figure, the MODIS and AHI official products generally yielded lower CTH inversion results for cirrus clouds with small COTs near 0°-10°S, and both MODIS and AHI products missed some thinner cirrus clouds near 2°S or 4°S. However, the DNN model's results were generally superior to the MODIS and AHI official products, effectively inverting the CTH of thin cirrus clouds and the upper ice clouds and lower water clouds in multi-layered clouds.

[0027] References

[0028] [1]BORBAS E, SEEMANN SW, HUANG HL, et al. Global profile training database for satellite regression retrievals with estimates of skintemperature and emissivity, F, 2005[C].

[0029] [2]LI W,ZHANG F,SHI Y-N,et al.Efficient radiative transfer model forthermal infrared brightness temperature simulation in cloudy atmospheres[J].2020,28(18):25730-25749.

[0030] [3]NAKAJIMA T,KING M D.Determination of the optical thickness andeffective particle radius of clouds from reflected solar radiationmeasurements.Part I:Theory[J].Journal of Atmospheric Sciences,1990,47(15):1878-1893.

[0031] [4]PLATNICK S,KING M D,ACKERMAN S A,et al.The MODIS cloud products:Algorithms and examples from Terra[J].IEEE Transactions on geoscience andRemote Sensing,2003,41(2):459-473.

[0032] [5]INOUE T.On the temperature and effective emissivity determinationof semi-transparent cirrus clouds by bi-spectral measurements in the 10μmwindow region[J].Journal of the Meteorological Society of Japan Ser II,1985,63(1):88-99.

[0033] [6]WANG C,YANG P,BAUM B A,et al.Retrieval of Ice Cloud OpticalThickness and Effective Particle Size Using a Fast Infrared RadiativeTransfer Model%J Journal of Applied Meteorology and Climatology[J].2011,50(11):2283-2297.

[0034] [7]DAVIS S M,AVALLONE L M,KAHN B H,et al.Comparison of airborne insitu measurements and Moderate Resolution Imaging Spectroradiometer(MODIS)retrievals of cirrus cloud optical and microphysical properties during theMidlatitude Cirrus Experiment(MidCiX)[J].2009,114(D2).

[0035] [8]IWABUCHI H,PUTRI N S,SAITO M,et al.Cloud property retrievalfrommultiband infrared measurements by Himawari-8[J].Journal of theMeteorological Society of Japan Ser II,2018,96B:27-42.

[0036] [9]TENG S,LIU C,ZHANG Z,et al.Retrieval of Ice-Over-Water CloudMicrophysical and Optical Properties Using Passive Radiometers[J].2020,47(16):e2020GL088941.

[0037]

[10] YEOM J M,YUM S S,SHAW R A,et al.Vertical Variations of CloudMicrophysical Relationships in Marine Stratocumulus Clouds Observed Duringthe ACE-ENA Campaign[J].2021,126(24):e2021JD034700.

[0038]

[11] DING K,GORDON H R.Atmospheric correction of ocean-color sensors:effects of the Earth’s curvature[J].Appl Opt,1994,33(30):7096-7106.

[0039]

[12] LETU H,ISHIMOTO H,RIEDI J,et al.Investigation of ice particlehabits to be used for ice cloud remote sensing for the GCOM-C satellitemission[J].Atmos Chem Phys,2016,16(18):12287-12303.

[0040]

[13] N,ADOK C,THOSS A,et al.Neural network cloud toppressure and height for MODIS[J].Atmos Meas Tech,2018,11(5):3177-3196.

[0041]

[14] LIN H,LI Z,LI J,et al.Estimate of daytime single-layer cloud baseheight from advanced baseline imager measurements[J].Remote Sensing ofEnvironment,2022,274:1-15.

[0042]

[15] LI W,ZHANG F,LIN H,et al.Cloud detection and classificationalgorithms for Himawari-8 imager measurements based on deep learning[J].2022。

Claims

1. A consistent cloud physical property inversion method based on radiation patterns and machine learning is used to invert the cloud top height (CTH), cloud optical thickness (COT), and cloud particle effective radius (CER) of single-layer and multi-layer clouds from visible and infrared multi-channel data of the Himawari 8 satellite AHI using a cloud physical property remote sensing inversion model. The specific steps are as follows: (1) Construct a training dataset for the remote sensing inversion model of cloud physical characteristics; First, based on the SeeBor 5.0 atmospheric profile dataset, which can characterize the temperature and possible absorption gas conditions in the actual atmosphere, cloud physics characteristics of single-layer and multi-layer clouds and solar / satellite observation geometric parameters are randomly generated within a pre-defined range. Then, these are combined with the atmospheric profile dataset, and the forward radiative transfer model (ERTM) is used to simulate the cloud reflectivity, brightness temperature, and clear-sky brightness temperature of each channel of the AHI imager under different cloud state assumptions, thus forming the "cloud physics characteristics - multi-channel simulated radiative values" dataset. Here, the clear-sky brightness temperature is the brightness temperature simulated by the radiative transfer model based on the background atmosphere and surface conditions under the assumption of clear sky and no clouds. Finally, the simulated dataset is divided into a training set, a validation set, and a test set for use in step (2). (2) Construct a remote sensing inversion model for cloud physical characteristics; The establishment of the cloud physical characteristics remote sensing inversion model involves two rounds of training. For the neural network model, it is first pre-trained using a dataset simulated by radiation models, and then further trained and optimized using a real satellite observation dataset. The pre-trained neural network model structure consists of an input layer, hidden layers, and an output layer; the intermediate hidden layers include a one-dimensional convolutional layer, an activation layer, a dropout layer, and a fully connected layer connected in sequence; among them, One-dimensional convolutional layers, acting as feature extractors, can extract key spectral information of different cloud physical properties from the input values, i.e., multi-channel radiance values. Each convolutional layer uses a kernel of size 5 to generate a set of feature maps. Batch normalization is applied after each convolutional layer to improve the robustness and accuracy of the model. The activation algorithm used in the model is the ReLU function. During training, the Adam optimizer is used to adjust the model parameters to minimize the root mean square error of cloud physical property inversion. At the same time, a dropout rate of 0.5 is used in the hidden layers to avoid overfitting and improve the performance of the neural network model. The neural network model is pre-trained using the simulated dataset generated by the radiation model in step (1). The multi-channel radiation values, namely brightness temperature / reflectivity and clear sky brightness temperature, are used as inputs, and the corresponding cloud physical characteristics are output. The learning rate is set to decrease exponentially with the increase of training times. The model performance is evaluated on the validation set, thereby adjusting the model hyperparameters to select the optimal model configuration and initially constructing a cloud remote sensing inversion model based on radiation model simulation. (3) Optimize the remote sensing inversion model of cloud physical characteristics; Since the spaceborne active sensors, namely CPR cloud radar and CALIOP lidar, can provide reliable cloud vertical profile information, the active radar observation data is used as the standard value of cloud physical characteristics. The AHI observation data, atmospheric and surface state data and CALIOP / CPR active radar data are spatiotemporally matched. Then, based on the atmospheric and surface data, the ERTM model is used to simulate the clear sky brightness temperature of each channel of AHI thermal infrared, and combined with the multi-channel observed reflectivity and brightness temperature to generate a cloud physical characteristics-multi-channel radiation value dataset based on real satellite observations. Based on the real dataset, the cloud physical characteristics remote sensing inversion model constructed in step (2) is retrained using the transfer learning method to optimize the model. In the transfer learning, the parameters of the first two convolutional layers are fixed, and the remaining network structure parameters are adjusted and optimized. The final cloud physical characteristics remote sensing inversion model can achieve higher accuracy inversion of cloud microphysical characteristics.

Images