Supercooled water clouds (SWCs) are prevalent in the atmosphere and crucial for global and local radiation balance, aviation safety, and weather modification techniques like artificial precipitation. Therefore, there is an imperative need for the continuous and precise monitoring of SWCs at high temporal and spatial resolution, encompassing observations under all sky conditions. This study aims to enhance the identification of SWCs by leveraging thermal infrared (TIR) channels of the Himawari-8 geostationary satellite, regardless of solar illumination or sun glint effects, which can be problematic for reflectivity bands. Principal component analysis (PCA) is utilized to perform a sensitivity analysis on a dataset comprising TIR bands from the Himawari-8 satellite and labels derived from the Cloud-Aerosol LiDAR and Infrared Pathfinder Satellite Observation (CALIPSO) cloud profile products, to assess the efficacy of the data in differentiating between supercooled water and ice clouds (ICs). Subsequently, machine learning techniques are employed to develop an all-day SWC identification model. The model is assessed using a time-independent dataset, yielding an overall accuracy rate for cloud phase (CPH) identification of over 90%, as well as high performance for detecting SWCs. The model demonstrates consistent performance across various surfaces, times of day, and seasons. Notably, it outperforms traditional algorithms that rely on reflectivity bands by accurately identifying SWCs even in sun glint regions, thus improving the reliability of CPH detection for applications in meteorology, climate research, and aviation safety.

To enhance the accuracy of cloud classification, this study proposes cloud classification models based on machine learning algorithms. The models take as input the observed reflectance or brightness temperature of 12 channels of the Advanced Geostationary Radiation Imager (AGRI) on Fengyun-4A satellite, and multi-channel clear sky brightness temperature. The classification results of the CPR-CALIOP merged product are used as the truth for training and validating the models. These models are developed to reliably detect and classify the clouds during daytime as well as for all-time (including both day and night). The results obtained from the developed models show better accuracies relative to those of the Fengyun 4A Level-2 cloud products in terms of cloud detection and classification. The models provide a feasible method for the detection of multi-layer clouds and classification of clouds at night. The applicability of cloud classification results based on CPR-CALIOP from the perspective of spectral sensitivity is analyzed on AGRI observations, providing valuable prior knowledge for cloud classification methods based on geostationary satellite imagers. The accuracies of single-layer cloud type classification during the day and all-time are 83.4% and 79.4%, respectively. Compared with the ISCCP classification method, the model’s identification of Nimbostratus and the Deep convection clouds (Ni/DC) has better consistency with precipitation observed by GPM satellite, which helps to track and monitor precipitation processes. This study also evaluates the model results using CALIPSO products and ground-based cloud radar, demonstrating that they can obtain accurate and robust results in different time periods and regions.

The soil complex dielectric permittivity (CDP) is a fundamental physical quantity utilized in microwave remote sensing to estimate soil moisture. Although several four-parameter (soil temperature, moisture, clay content, and frequency) semiempirical complex dielectric mixing models (SEMs) have been developed, they do not consider the changes in soil CDP caused by the conversion of bound water to free water due to temperature variations. This study proposes an improved SEM (also known as the Dobson-Jin model) based on the Dobson SEM. The improvements include: 1) rectifying the distortion phenomenon in the imaginary part of the Dobson model and minimizing the error in calculating sand using the Dobson model; 2) reducing the number of input parameters from seven to four; and 3) dividing soil water into strongly bound water and soil-effective water (the sum of weakly bound water and free water) and successfully simulating the “competitive mechanism” of soil CDP variation with temperature by establishing the weakly bound water to free water conversion functions and the weakly bound water proportion functions. Experimental validation using measured data from 31 types of soils has demonstrated that the Dobson-Jin model is capable of accurately calculating the CDP of sand, loam, and clay within the range of 1.4–18 GHz and water content volume of 0–0.5 (cm$^{3}\cdot \text { cm}^{-3}$). In summary, the Dobson-Jin model is expected to improve the accuracy of microwave remote sensing inversion of soil moisture.

A physics-driven machine learning (ML) algorithm that integrates radiative transfer model (RTM) simulation and ML techniques has been developed to facilitate multilayer cloud property retrievals from advanced Himawari imager (AHI) measurements based on the geostationary satellite Himawari-8. Theoretical sensitive study revealed that integrating RTM-simulated clear-sky radiances into the retrieval model holds substantial potential to enhance multilayer cloud property retrieval. Thus, a convolutional neural network (CNN) model called CNN_TL was designed for simultaneous retrieval of ice and water properties in multilayer clouds, based on the combined use of visible/near-infrared (VNIR) and thermal infrared (TIR) measurements and RTM-simulated clear-sky radiances as predictors. The CNN_TL model was initially trained using the simulated datasets generated by RTM to gain better model initiation and generalization and further tuned using the standard references from collocated active sensor products through transfer learning (TL) method. Validation on independent dataset shows CNN_TL-retrieved cloud top heights agreeing well with active sensor products for both overlaying ice and underlying water (RMSEs: 1.219 and 0.863 km), substantially outperforming AHI official products. Additionally, due to separate retrieval of ice and water microphysical properties in multilayer clouds, CNN_TL retrieved overlying ice microphysical properties exhibited a substantial correlation with active sensor products, showing Pearson coefficients above 0.78. Notably, CNN_TL model outperformed the purely ML-based model, demonstrating the advantage of integrating physical RTM simulation and ML techniques through TL method. Finally, a novel investigation into the spatial distribution of multilayer cloud properties across full disk is conducted, revealing distinct seasonal variations and notable latitudinal dependencies.

A hybrid algorithm based on radiative transfer simulations and machine learning for dust aerosol detection, is developed for the Advanced Himawari Imager (AHI) carried by the geostationary satellite Himawari-8. The sensitivities of the AHI thermal infrared (TIR) channels for dust aerosols are analyzed through radiative transfer simulations. The sensitivity study demonstrates that the simulated clear-sky brightness temperatures (BTs) show an obvious improvement in identifying dust aerosols compared to brightness temperature difference techniques, especially optically thin dust. Therefore, the simulated clear-sky BTs and AHI TIR observed BTs, in addition to ground information, are used as inputs to add physical knowledge in the machine learning model. The performance of an artificial neural network constructed for dust aerosol detection is evaluated by comparing its results with those of active Cloud-Aerosol Lidar with Orthogonal Polarization measurements. The proposed algorithm effectively achieves dust aerosol detection during both daytime and nighttime, with a precision of over 86% and a recall of over 85% on an independent testing dataset. The proposed algorithm is applied to three typical dust events to further illustrate its applicability. Although some thin dust aerosols near the ground are misclassified due to weak signals, most dust aerosols are successfully detected, and the identification is generally not affected by other types of aerosols. The results of the regional classification demonstrate that our algorithm is superior in detecting tenuous dust aerosols compared to Dust RGB images using the AHI TIR channels and physical-based algorithm.

Clouds play an important role in the Earth’s climate system; however, various observational methods describe clouds differently, leading to cloud products being described with different characteristics, and affecting our understanding of cloud effects. To address this problem, this study integrates different cloud products into the transfer-learning procedure of a deep learning model and determined the Cloud Effective Radius (CER), Cloud Optical Thickness (COT), and Cloud Top Height (CTH) from Himawari-8 thermal infrared measurements. The retrieval results were independently evaluated against the Moderate-resolution Imaging Spectroradiometer cloud products and further compared with Himawari-8 cloud products during the day. The Root Mean Squared Errors (RMSE) of the model for the CER, COT, and CTH were 4.490 μm, 11.198, and 1.904 km, respectively, which are lower than those of Himawari-8 cloud products (RmSe:11.172 μm, 14.755, and 2.860 km). Moreover, validation results against active sensors show that the model performs slightly better during the day than at night, and both are generally better than the Himawari-8 cloud product. Overall, the model maintains stable performance during both day and night, and its accuracy is higher than that of Himawari-8 cloud products.

The Advanced Geostationary Radiation Imager (AGRI) onboard the Fengyun-4A (FY4A) satellite has good cloud observation ability, but it still absents all-weather and high-precision official cloud products. This study develops a deep-learning ResUnet model for all-weather retrieval of cloud phase (CLP) and cloud properties using the brightness temperature from water vapor and longwave infrared channels of AGRI. The ResUnet model is trained with the Himawari-8 satellite Level-2 (H8-L2) cloud products as true targets, and adopts image-by-image way to learn the spatial structure information of clouds, which compensates for the difficulty of retrieving thick clouds by thermal infrared radiation at night to some extent. On an independent testing dataset, the model has an overall accuracy of 90.64% for CLP identification and performs well at retrieving cloud top height (CTH). Even without using visible and near-infrared radiation, the root mean square error of cloud effective radius (CER) and cloud optical thickness (COT) estimations still reaches 7.14 μm and 9.01 in the range of 0–60. To further illustrate the reliability and applicability, CLP and cloud properties provided by the CALIPSO and MODIS are used as benchmarks to assess the quality of cloud products from FY4A satellite Level-2 (FY4A-L2), H8-L2 and ResUnet model retrieval. The ResUnet model provides a significant improvement over FY4A-L2 for the accuracy of cloud identification and in the quality of CTH products. In the range of 0–40 μm (0–60), the CER (COT) product of ResUnet model retrieval has a reliable and higher precision that is comparable with H8-L2.

Cloud segmentation is fundamental in obtaining many parameters of clouds. However, traditional cloud segmentation performs far from satisfactory, due to the fuzzy boundaries and complex textures of clouds. Although deep learning methods have shown superior performance in cloud segmentation, they are constrained by limited labels in ground-based cloud image data sets. This letter established a new Ground-Based Cloud Segmentation (GBCS) data set with 1742 accurately labeled images. Then to evaluate how well deep learning models perform in cloud segmentation, 12 state-of-the-art semantic segmentation networks are selected, among which DeepLabV3+ outperformed all others. Since 1742 images are not enormous, a novel Transfer learning (TL)-DeepLabV3+ model was developed by TL: DeepLabV3+ network was trained with the PASCAL VOC 2012 data set, then retrained in GBCS. TL-DeepLabV3+ showed a high ability of cloud segmentation, scoring the Mean Intersection-over-Union (MIoU) of 91.05% in GBCS and further verified in the UTILITY data set and the Cirrus Cumulus Stratus Nimbus (CCSN) data set.

A deep-learning-based cloud detection and classification algorithm for advanced Himawari imager (AHI) measurements from the geostationary satellite Himawari-8 has been developed. It is found that a combination of observed radiances and simulated clear-sky radiances can substantially improve cloud phase discrimination, especially for optically thin clouds. Therefore, cloud detection, cloud phase classification, and multilayer cloud detection are obtained simultaneously from multispectral observed radiances and simulated clear-sky radiances using deep neural networks (DNNs). Two DNN models are established for all-day and daytime-only applications, respectively, using active Cloud Profiling Radar (CPR) and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) merged cloud products from 2016 as reference labels. The independent dataset from 2017 is used to validate the DNN models. It is shown that both the DNN models outperform the official Moderate Resolution Imaging Spectroradiometer (MODIS) and AHI products in cloud detection and phase discrimination, and the enhancement is more significant over land than over water surface. For multilayer cloud detection, the probability of detecting multilayer clouds reaches ~60% for the all-day model and is increased to ~70% for the daytime model, which is substantially better than MODIS and AHI products. In practical cases, multilayer cloud detection by DNN models is more consistent with CPR/CALIOP than two official products. In addition, the DNN models have superior capability in detecting the optically thin cirrus, which is omitted by MODIS and AHI products. Specifically, the cases also demonstrate that the DNN models can provide effective mixed-phase cloud identification. This deep-learning-based algorithm has the potential for measurements from other similar instruments.

Advances in Internet of Things (IoT) and cloud/edge computing systems could precisely monitor the meteorological elements and environmental conditions. Remote automated observation system (RAOS) makes the full use of IoT to communicate with other sensors, enabling the active responses from passive devices for smart weather. Cloud observation and classification have been regarded as a successful application that could automatically perform emergency tasks in RAOS. However, with the increasing growth of resource exploitation, the performance of communications among the automatic observation platforms, and the efficiency of task allocation among them has become a critical challenge. In this article, an ensemble learning method and resource allocation scheme are proposed to realize the cloud observation and classification with the help of reliable and controllable infrastructures. On the one hand, several ensemble methods, like Bagging, AdaBoost, and Snapshot are selected as a base classifier to capture the cross-semantic and structure features of cloud, while applying them to the ensemble using convolutional neural networks with different base learners and residual neural networks with different depths. on the other hand, a particular cloud-edge distributed framework is proposed for cloud classification approach based on the intelligent network, to overcome the difficulty in the massive data transmission. The experimental results verify that the proposed ensemble approach achieves high accuracy of cloud classification, and effectively improves the number of allocated tasks. Ensemble methods can generate a more accurate prediction than any single classifier or the majority algorithms. It consistently yields lower error rates than single state-of-the-art models at no additional training cost.