This study develops a machine learning (ML)-based multilayer cloud detection algorithm for the passive Advanced Himawari Imager (AHI) aboard the geostationary Himawari-8 satellite. AHI measurements in the 0.64-, 1.6-, 2.3-, 3.9-, 7.3-, 8.6-, 11.2-, and 12.4- $\mu \text{m}$ channels, their combinations, geolocations, and observational geometries are used as predictors, and collocated active CloudSat and CALIPSO data are used to accurately label multilayer cloud pixels as the reference/truth of the predictand. We develop an ML-based daytime model (ML-Day) that utilizes all the aforementioned predictors and an all-time one (ML-All) that excludes the solar channel-dependent variables. Among four ML algorithms, the random forest (RF) performs slightly better than the artificial neural network, K-nearest neighbor, and support vector machines. By comparing with the merged CloudSat and CALIPSO product, the ML-Day model correctly identifies ~89% single-layer clouds and ~70% multilayer clouds, outperforming the Moderate Resolution Imaging Spectroradiometer (MODIS) operational multilayer cloud product (~80% and ~40% given by Marchant et al. ). The success rates of ML-All for single-layer and multilayer clouds also reach ~85% and ~64%, respectively. The misclassification of our algorithm is mostly caused by missing optically thin clouds, a drawback of most radiometers without the 1.38- $\mu \text{m}$ channel. Furthermore, with multilayer cloud pixels well detected by our algorithm, the AHI operational cloud top height retrievals are found to be larger biased due to multilayer cloud occurrence and might be improved by considering cloud vertical structures.

Estimating cloud-base height (CBH) from satellite-based passive imaging radiometers is very challenging because the conventional visible and infrared radiance measured by satellite cannot be employed to profile clouds from top to base. The Advanced Himawari Imager (AHI) aboard Himawari-8 is the first of Japanese new-generation visible-infrared imaging spectroradiometer with an improved temporal and spatial resolution. A CBH estimation algorithm for Himawari-8 is imminently needed for aviation safety and climate research. This study presents a machine learning methodology for CBH estimation from Himawari-8 data using a random forest (RF) algorithm, which was conducted for the daytime focusing on single-layer clouds. About 2,876,230 matchups in the period from August 2015 to July 2016 are used in both RF training and validation processes for the purpose of learning the relationship between inputs from Himawari-8 Level-2 cloud products together with geolocation information and target CBH from CloudSat/CALIPSO combined measurements. With the trained RF model, the CBH prediction can be independently generated using Himawari-8 upstream products. Data of ground-based K_a radar in Beijing during 2017 and 2019 are used for independent validation. The results indicate that the Himawari-8 CBH estimation agrees well with those of the K_a radar. The CBH estimations for single-layer nonprecipitation clouds show the best performance with a mean difference of −0.2 ± 1.7 km. The presence of multi-layer clouds mostly leads to overestimated CBHs relative to the radar values, particularly for situations that cirrus clouds overlapping clouds at low-level. Since the ability of the K_a radar is limited to retrieve CBH in precipitation cases, significant overestimation also occurs for precipitation clouds with mean difference of 1.7 ± 2.4 km.

To evaluate the validity of cloud top height (CTH) retrievals from FY-4A, the first of China’s next-generation geostationary meteorological satellite series, the retrievals are compared to those from Himawari-8, CloudSat, Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO), and Moderate Resolution Imaging Spectroradiometer (MODIS) operational products from August to October 2017. Regarding CTHs from CloudSat, CALIPSO, and MODIS as truth, the results show that the performance of FY-4A CTH retrievals is similar to that of Himawari-8. Both FY-4A and Himawari-8 retrieve reasonable CTH values for single-layer clouds, but perform poorly for multi-layer clouds. The mean bias error (MBE) shows that the mean value of FY-4A CTH retrievals is smaller than that of Himawari-8 for single-layer clouds but larger for multi-layer clouds. For ice crystal clouds, both FY-4A and Himawari-8 obtain the underestimated CTHs. However, there is a tendency for FY-4A and Himawari-8 to overestimate the CTH values of CloudSat and CALIPSO mainly for low level liquid water clouds. The temperature inversion near the tops of water clouds may result in an overestimation of CTHs. According to the MBE change with altitude, FY-4A and Himawari-8 overestimate the CTHs mainly for clouds below 3 km, and the overestimation is slightly more apparent in Himawari-8 data than that in FY-4A values. As the cloud optical thickness (COT) increases, the CTH bias of FY-4A CTH retrievals gradually decreases. Two typical cases are analyzed to illustrate the differences between different satellites’ CTH retrievals in detail.

Under the assumption that clouds are homogeneous and single-layered (SL), most current operational cloud top height (CTH) products derived from passive radiometers may largely underestimate the CTH of overlapping clouds. This article proposes a statistics-based extrapolation algorithm for retrieving the CTHs of overlapping clouds using only existing cloud property products available for most operational radiometers, and the method is successfully employed for the advanced himawari imager (AHI) observations. Because regional clouds within the same “system” have relatively continuous geometric properties, especially CTH, due to similar atmospheric conditions, upper-layer ice cloud CTHs (ITHs) and lower-layer water cloud CTHs (WTHs) are inferred using the CTH retrievals of well-chosen neighboring SL ice and water clouds, respectively. The proposed algorithm uses the latest machine-learning-based model to reasonably distinguish overlapping clouds from SL clouds, and optimizes the extrapolation by considering three physical constraints on neighboring, cloud phase, and cloud optical thickness (COT). Validated using active observations from CloudSat and cloud-aerosol Lidar and infrared pathfinder satellite observation (CALIPSO), our algorithm improves the AHI CTH mean bias for overlapping clouds from −5.1 to −2.6 km. More importantly, the algorithm provides CTH information of underlying water clouds that are unavailable from existing radiometer-based products. With the simultaneous retrieval of ITH and WTH, this algorithm increases our capability to detect the vertical structures of overlapping clouds and better evaluate the cloud radiative effects (CREs).

Cloud base height (CBH) is crucial for parameterizing the cloud vertical structure (CVS), but knowledge concerning the temporal and spatial distribution of CBH is still poor owing to the lack of large-scale and continuous CBH observations. Taking advantage of high temporal and spatial resolution observations from the Advanced Himawari Imager (AHI) on board the geostationary Himawari-8 satellite, this study investigated the climatology of CBH by applying a novel CBH retrieval algorithm to AHI observations. We first evaluated the accuracy of the AHI-derived CBH retrievals using the active measurements of CVS from the CloudSat and CALIPSO satellites, and the results indicated that our CBH retrievals for single-layer clouds perform well, with a mean bias of 0.3 ± 1.9 km. Therefore, the CBH climatology was compiled based on AHI-derived CBH retrievals for single-layer clouds for the time period between September 2015 and August 2018. Overall, the distribution of CBH is tightly associated with cloud phase, cloud type, and cloud top height and also exhibits significant geographical distribution and temporal variation. Clouds at low latitudes are generally higher than those at middle and high latitudes, with CBHs peaking in summer and lowest in winter. In addition, the surface type affects the distribution of CBH. The proportion of low clouds over the ocean is larger than that over the land, while high cloud occurs most frequently over the coastal area. Due to periodic changes in environmental conditions, cloud types also undergo significant diurnal changes, resulting in periodic changes in the vertical structure of clouds.

Layered cloud cover (LCC), that is, cloud cover at different levels, is crucial for estimating cloud radiative effects and modeling climate change. However, accurate LCC characterization using passive satellite measurements is challenging because of the difficulties in resolving cloud vertical structures. In this study, we developed a novel method to estimate LCC from geostationary satellite radiometric measurements. The proposed method resolves cloud vertical structures by retrieving cloud-top and cloud-base heights for both single- and multi-layer clouds; thus, better estimating LCC. Our results agreed well with active satellite measurements, showing identification accuracies of 86%, 90%, and 91% for high, medium, and low clouds, respectively. Additionally, our LCC estimates derived from satellite measurements were used to evaluate those from atmospheric reanalysis. The annual averaged total, high, medium, and low cloud covers given by our methods were 0.681, 0.393, 0.356, and 0.455, respectively, while those from ERA-5 were 0.623, 0.415, 0.274, and 0.392, respectively. These results indicate that the total cloud cover determined by ERA-5 was lower than that derived from satellite measurements, potentially as a result of medium and low-level clouds.