JournalJune 2019 of the Meteorological Society of Japan, 97(3),H. 689−710,HIROSE et2019. al. doi:10.2151/jmsj.2019-040 689

High Temporal Rainfall Estimations from Himawari-8 Multiband Observations Using the Random-Forest Machine-Learning Method

Hitoshi HIROSE, Shoichi SHIGE, Munehisa K. YAMAMOTO

Graduate School of Science, Kyoto University, Kyoto, Japan

and

Atsushi HIGUCHI

Center for Environmental Remote Sensing, Chiba University, Chiba, Japan

(Manuscript received 27 March 2018, in final form 26 Febuary 2019)

Abstract

We introduce a novel rainfall-estimating algorithm with a random-forest machine-learning method only from Infrared (IR) observations. As training data, we use nine-band brightness temperature (BT) observations, obtained from IR radiometers, on the third-generation geostationary meteorological satellite (GEO) Himawari-8 and precipitation radar observations from the Global Precipitation Measurement core observatory. The Himawari-8 Rainfall-estimating Algorithm (HRA) enables us to estimate the rain rate with high spatial and temporal resolu- tion (i.e., 0.04° every 10 min), covering the entire Himawari-8 observation area (i.e., 85°E – 155°W, 60°S – 60°N) based solely on satellite observations. We conducted a case analysis of the Kanto–Tohoku heavy rainfall event to compare HRA rainfall estimates with the near-real-time version of the Global Satellite Mapping of Precipitation (GSMaP_NRT), which combines global rainfall estimation products with microwave and IR BT observations obtained from satellites. In this case, HRA could estimate heavy rainfall from warm-type precipitating clouds. The GSMaP_NRT could not estimate heavy rainfall when microwave satellites were unavailable. Further, a sta- tistical analysis showed that the warm-type heavy rain seen in the Asian monsoon region occurred frequently when there were small BT differences between the 6.9-μm and 7.3-μm of water vapor (WV) bands (ΔT6.9 – 7.3). Himawari-8 is the first GEO to include the 6.9-μm band, which is sensitive to middle-to-upper tropospheric WV. An analysis of the WV multibands’ weighting functions revealed that ΔT6.9 – 7.3 became small when the WV amount in the middle-to-upper troposphere was small and there were optically thick clouds with the cloud top near the middle troposphere. Statistical analyses during boreal summer (August and September 2015 and July 2016) and boreal winter (December 2015 and January and February 2016) indicate that HRA has higher estima- tion accuracy for heavy rain from warm-type precipitating clouds than a conventional rain estimation method based on only one IR band.

Keywords warm-type heavy rain; Himawri-8; GSMaP; GPM; machine-learning

Citation Hirose, H., S. Shige, M. K. Yamamoto, and A. Higuchi, 2019: High temporal rainfall estimations from Himawari-8 multiband observations using the random-forest machine-learning method. J. Meteor. Soc. Japan, 97, 689–710, doi:10.2151/jmsj.2019-040.

Corresponding author: Hitoshi Hirose, Center for Environ- mental Remote Sensing, Chiba University, 1-33 Yayoi, Inage, Chiba 263-8522, Japan E-mail: [email protected] J-stage Advance Published Date: 15 March 2019 ©The Author(s) 2019. This is an open access article published by the Meteorological Society of Japan under a Creative Commons Attribution 4.0 International (CC BY 4.0) license (https://creativecommons.org/licenses/by/4.0). 690 Journal of the Meteorological Society of Japan Vol. 97, No. 3

it is difficult to use liquid water emission over land, 1. Introduction the MWR algorithm estimates the rain rate mainly by Global rainfall observation datasets with high accu- scattering signatures from of ice crystals. Therefore, racy are important role in climatology and hydrology, the MWR algorithm also assumes that deeper and especially in terms of disaster countermeasures. colder clouds tend to cause heavier rain, like the IR al- Satellite observations are the most suitable means of gorithm (Spencer 1984; Ferraro et al. 2005). Hamada obtaining global observation data. Rainfall estimates et al. (2015) reported that the heaviest rain was caused from satellite signatures were first made using infrared by the clouds with lower echo top height (ETH), (IR) BT and visible (VIS) reflectance from cloud tops rather than those with the highest ETH, by using the observed via geostationary meteorological satellites Tropical Rainfall Measurement Mission (TRMM; (GEOs) (Barrett 1970). Observational studies in the Kummerow et al. 1998) precipitation radar (PR). United States created a classical model of heavy rain Sohn et al. (2013) reported that heavy rainfall over the from deep cumulonimbus clouds (Byers and Braham Korean peninsula was mainly caused by the clouds 1949). with ETH lower than 8 km. The rain rate of this type Based on this conceptual model, IR-radiometer- heavy rainfall was comparable to that of heavy rain based rainfall-estimating algorithms assume that from the deep convective cloud seen in Oklahoma in deeper and colder clouds tend to cause heavier rain the United States of America, although the ETH was (Richards and Arkin 1981). However, the estimation obviously low. They referred to low-level clouds, accuracy of IR-radiometer-based techniques is low which are associated with heavy rainfall, as warm- for optically thin clouds, such as cloud anvils, due to type clouds, and we use “warm-type” to indicate the the weak linkage between the cloud-top temperature same meaning in this study. Significant underestima- (CTT) and precipitation (Adler et al. 1993). For more tion of rainfall occurred in the coastal mountains in accurate and precise precipitation estimates, research- the Asian monsoon region when using the GSMaP ers have attempted to use microwave radiometer MWR algorithm (Aonashi et al. 2009; Shige et al. (MWR) signatures, which observe microwave emis- 2009). Shige et al. (2013) reported that such heavy sions from liquid hydrometeors and scattering from rainfall, with low precipitation-top heights, occurs due ice particles. Compared to IR information obtained to orographically forced upward motion over coastal only from the cloud top, MWR algorithm can estimate mountains where large amounts of WV constantly the vertical structure inside the precipitating cloud converge. An orographic–nonorographic rainfall clas- (Arkin and Ardanuy 1989; Smith et al. 1998). sification scheme to identify orographic rainfall with However, there are gaps in the overpasses of MWR low precipitation-top heights has been incorporated satellites observation networks by low-Earth-orbiting into the GSMaP MWR algorithm (Shige et al. 2013; satellites when more frequent rainfall outputs are re- Taniguchi et al. 2013; Shige et al. 2015; Yamamoto quired. The Global Satellite Mapping of Precipitation and Shige 2015; Yamamoto et al. 2017). Using this (GSMaP) uses two consecutive IR GEO observations scheme, the rainfall estimate accuracy is improved at 1 h intervals to calculate the cloud-moving vector over almost all Asian regions; however, the rainfall (Ushio et al. 2009). Using the moving vector, the rate complemented by the IR algorithms in GSMaP rainfall observed by an MWR satellite is propagated still depends only on CTT information. The number along with the moving vector to interpolate gaps in of observation bands on the past GEOs has been very the MWR observation coverage. In addition to the limited, as shown in Table 1, and GSMaP uses only cloud-moving vector, GSMaP uses a Kalman filter one IR band, 10.8 μm. to update the rainfall intensity correspond to the IR As the performance of GEOs has improved in TB after propagation along with the moving vector recent years, studies of rainfall estimates by GEOs (GSMaP_MVK; Ushio et al. 2009). By combining have progressed. Upadhyaya and Ramsankaran (2016) MWR and IR, the global rainfall can be estimated proposed the multispectral rainfall estimation algo­ from satellite observations at high frequency. rithm using the Indian National Satellite System The matching algorithms of MWR and IR have (INSAT). They improved the estimate accuracy by succeeded, to a certain extent, where deep convection incorporating topographical information into the em- or large organized convection dominates, such as over pirical formula between the GEO-observed BT and the the tropical ocean or over continental North America. rain rate in all the climatic regions in India. However, The MWRs provide the liquid water emission of the INSAT radiometer has only three spectral bands, 8 lower frequency bands over the ocean. However, since km spatial resolution of IR and WV bands and 30 min June 2019 H. HIROSE et al. 691 temporal resolution. to-upper troposphere using three WV bands. Using The problem of spatial resolution and the number of the Himawari-8 observational data, we can expect to GEO observation bands was solved after the Meteosat obtain more detailed precipitation related information Second Generation (MSG; Aminou 2002) satellite was to analyze “warm-type” rain in the Asian monsoon launched in December 2005. The Spinning Enhanced region, as mention in Sohn et al. (2013). Visible and Infrared Imager (SEVIRI) sensor on MSG To investigate how the estimate accuracy of warm- operates in 12 spectral bands: 3 VIS bands with a spa- type heavy rain improves when using the multiband IR tial resolution of 1 km and 8 IR and WV bands with of Himawari-8, we created high-frequency precipita- a spatial resolution of 3 km. Roebeling and Holleman tion data by applying the RF machine-learning method (2009) developed a rainfall-estimating algorithm to the multiband IR observations of Himawari-8. This based on an empirical relation between the cloud makes it possible to estimate rainfall over the entire physical properties estimated from MSG SEVIRI and observation range of Himawari-8 (85°E – 155°W) the rainfall observed via weather radar in Europe. using the Global Precipitation Measurement (GPM; Bergès et al. (2010) built a neural network based on Hou et al. 2014) satellite equipped with the Dual- training data created from simultaneous observations frequency Precipitation Radar (DPR; Kojima et al. of MSG SEVIRI and TRMM PR and estimated rain 2012) instead of the ground-based radar for machine probability and intensity in African regions. Kühnlein learning. GPM was launched in February 2014 with a et al. (2014) provided a new rainfall retrieval tech- large orbital inclination angle of 65° and can observe nique using CTT, cloud phase, and cloud water path a wide range from tropical to mid-latitude regions. data retrieved from the MSG SEVIRI multiband We performed a case analysis of the Kanto–Tohoku observations in Germany with the random-forest (RF) heavy rainfall event to validate the accuracy of the machine-learning algorithm. The RF method, original- Himawari-8 Rainfall estimation Algorithm (HRA). In ly proposed by Breiman (2001), is a highly accurate, this case, it was difficult to estimate the rainfall using machine-learning algorithm that can deal with the the GEO IR interpolation method of GSMaP because multiband data from MSG. Rainfall estimates using there was heavy rainfall from warm-type precipitating other machine-learning methods have been attempted clouds. This is why we chose this case. Further, we (Capacci and Conway 2005); however, the RF method analyzed how the multiband IR, including the newly can process a large number of inputs with substantial- available Himawari-8 6.9-μm WV band, contributed ly lower calculation costs compared to conventional to estimating warm-type heavy rain. This study aims methods. Tebbi and Haddad (2016) investigated the to create homogeneous daytime and nighttime rainfall cloud classification potential using high-resolution, estimation products; therefore, we use only IR multi- MSG SEVIRI images from northern Algeria. They band without VIS bands. estimated rainfall using neural networks with MSG 2. Data and rain gauge observations as training data. The above studies successfully estimated rainy The Himawari-8 AHI has 6 VIS, 9 IR and 1 short­ areas, rain types, and rainfall rates from the multi- wave IR bands; however, we only used the IR bands, band observations of MSG SEVIRI and significantly as shown in Table 1. Three bands in the WV absorp- improved the estimation accuracy compared to the tion band (6.2, 6.9, and 7.3 μm) are available, while conventional estimation method using GEOs. How- the single 6.6 μm band was available in the Japan ever, the analysis ranges of these studies are limited Advanced Meteorological Imager (JAMI) of the by the observational range of MSG SEVIRI (60°W – Multi-functional Transport Satellite (MTSAT; Japan 60°E) and that of ground-based radars (in Europe or Meteorological Agency 2003) series, which is a Africa). Therefore, the objects of their analyses were previous generation Japanese GEO. These three WV primarily deep convective rainfall and frontal rainfall. bands are sensitive to the middle-to-upper troposphere The latest GEO, called Himawari-8, was launched and enable us to estimate the vertical distribution of in October 2014 and can observe most of the Asian WV from the BT difference between 6.2, 6.9, and 7.3 monsoon region (85°E – 155°W, 60°S – 60°N) with μm bands. Six bands in the atmospheric window band the nine IR bands of the Advanced Himawari Imager (8.6, 9.6, 10.4, 11.2, 12.4, and 13.3 μm) are used to (AHI), as shown in Table 1 (Bessho et al. 2016). monitor IR radiation from cloud tops without WV in- Himawari-8 is the first GEO equipped with a 6.9-μm terference. BT10.4 provides information about cloud WV band; therefore, it can obtain more detailed top height (CTH). When a deep convective cloud information about the WV distribution in the middle- reaches the tropopause, the BT difference between the 692 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Table 1. IR bands of MTSAT/JAMI and Himawari-8/AHI the first space-borne DPR (Seto and Iguchi 2011; specifications. Kojima et al. 2012), including Ku and Ka bands MTSAT HMWR-8 (13.6 and 35.5 GHz, respectively). KuPR has greater (µm) (µm) Prime measurement objective sensitivity than TRMM PR because of its higher trans- 8.6 Cloud phase mitted peak power that achieves a minimum detection threshold of 0.3 mm h−1 (a corresponding reflectivity 9.6 Total ozone of ~ 14.53 dBZ; Toyoshima et al. 2015; Hamada and 10.4 Cloud top temperature 10.8 Takayabu 2016). The KaPR observations are especial- 11.2 ly effective at weak rain and snow detection; however, 12.0 12.4 the cross-track swath width of KaPR is 120 km, which is approximately half of the swath width of KuPR (245 13.3 km). For our analysis period of one year, KaPR had 6.2 Water vapor (WV) in difficulty collecting sufficient samples for simultane- upper-troposphere 6.6 ous observations with the GEOs. Therefore, we used 6.9 WV in upper and only observations from KuPR (product version 3) as middle troposphere training data. 7.3 WV in middle troposphere For comparison, we used GSMaP which estimates global precipitation with a high temporal resolution by combining MWR and IR observations (Ushio et al. 2009). To fill the gap in the MWR satellites observa- WV bands and the atmospheric window bands is near tion network, the GSMaP estimates the precipitating zero because there is nearly no WV above the cloud. cloud’s destination, observed by an MWR satellite, Therefore, ΔT7.3 – 13.3 also provides information by using cloud-moving vector calculated from two about CTH (Schmetz et al. 1997; Reudenbach et al. successive GEO IR images. Additionally, the GSMaP 2001). Note that ΔT represents the BT difference applies a Kalman filter to the rainfall intensity after between the two IR bands, and the subscripts indicate propagation along with the moving vector to make the center wavelengths of the observation bands. The the intensity correspond to IR BT (GSMaP_MVK). extinction coefficients for water in the atmospheric GSMaP_NRT is another rainfall product for near-real- window band with longer wavelengths are larger than time requirements and has reduced data latency than those with shorter wavelengths in all the solid, liquid, GSMaP_MVK. GSMaP_NRT is expected for rain and gaseous states. Therefore, the cloud water path disaster mitigation but is inferior to GSMaP_MVK (CWP) can be described as the BT differences of two in terms of accuracy because of the increased weight atmospheric window bands, called IR split-window of supplementary products by GEOs. Therefore, (SW) technique (Inoue et al. 1985; Lensky and GSMaP_NRT’s accuracy is expected to improve by Rosenfeld 2003; Thies et al. 2008a). The original using HRA with high temporal resolution; we used Himawari-8 standard data were provided by the Japan GSMaP_NRT in this study. The GSMaP_NRT data- Meteorological Agency (JMA). The data were con- sets have 0.1° spatial resolution in the range between verted to latitude-longitude coordinates by the Center 60°N – 60°S. The method to fill the gap in the GSMaP for Environmental Remote Sensing (CEReS), Chiba MWR network is based on CTH information obtained University, Japan. CEReS applied precise geometric from only one IR band of GEO. We separated the data corrections to the original Himawari-8 standard data of GSMaP_NRT into MWR and IR estimations at the to eliminate the very small deviations in the AHI ob- time/place where the microwave satellite is available servational positions between each band. Himawari-8 to investigate the accuracy difference of the one IR observes the IR radiance every 10 min and has 2 km × band algorithm of GSMaP_NRT and that of HRA with 2 km spatial resolution at sub-satellite points (Bessho multiband IR. A product that is directly estimated from et al. 2016). A wide range of Himawari-8 (85°E – MWR (Aonashi et al. 2009; Shige et al. 2009; Kida 155°W, 60°S – 60°N) can cover most of the Asian et al. 2017) is called GSMaP_MWR, and a product monsoon area. that is estimated from GEO IR is called GSMaP_IR. The reliable rainfall observation data used for our To separate GSMaP_MWR and GSMaP_IR, we used machine-learning algorithm consisted of near-surface a satellite flag provided by GSMaP. This satellite flag rainfall observations from the GPM core observatory provides information on whether the GSMaP product (Hou et al. 2014). The GPM core observatory carries is observed by MWR satellites or complemented by June 2019 H. HIROSE et al. 693

Fig. 1. Overview of the stepwise scheme for rainfall rate assignment.

GEO IR. have peaks of observation frequency at different For the validation, we used the radar–Automated BT, we define threshold as a value that can best Meteorological Data Acquisition System (AMeDAS) separate cloud-free pixels from cloud-contam- composite, which is created from a composite of JMA inated pixels. Here, we defined the cloud area operational radars calibrated by AMeDAS rain gage rather broadly so that actual clouds were not data (Makihara et al. 1996; Makihara 2000). The excluded. This procedure is performed to reduce radar–AMeDAS composite covers the entire Japanese the large amount of unnecessary cloud-free areas. region with a temporal resolution of 10 min. All (ii) Rain-type separation: Precipitating cloud pixels analyzes with this radar–AMeDAS as truth were per- are classified into weak and strong rain pixels. formed in Japanese region (120 – 150°E, 20 – 50°N). (iii) Rain rate estimation: The rain rates of weak and The European Center for Medium-Range Weather strong rain pixels are estimated. Forecast (ECMWF) Interim reanalysis (ERA-Interim; In these three steps, classification models (rain or Dee et al. 2011) was used to describe the atmospheric no rain, weak or strong) and regression models are environment. In this study, we used the temperature, created as independent RF models and is executed relative humidity and geopotential height data with a stepwisely. resolution of 0.125° × 0.125° every 6 h. There is no standard threshold for distinguishing between weak and strong rain in step (ii); therefore, 3. Methods we referred to Thies et al. (2008b). If rain rates ob- 3.1 Learning and estimating procedure served by GPM KuPR are higher than 1.8 mm h−1, In Kühnlein et al. (2014), the RF method requires a then the precipitation pixels are classified as strong, stepwise scheme to estimate the rain rate because the otherwise they are classified as weak. relationship between the rain rate and the multiband BT is strongly nonlinear. Therefore, we used the step- 3.2 Preparation of training data and validation data wise scheme shown in Fig. 1. To estimate rainfall using the RF method for the (i) Rain area detection: Only those pixels that are multiband IR observations of Himawari-8, we first classified as cloudy by using Himawari-8 BT collected simultaneous observations from Himawari-8 observations are considered. For the cloud mask, and GPM to construct our training dataset (Fig. 2). the algorithm developed and implemented by Because of the different spatial resolutions between Cermak (2006) and Cermak and Bendix (2008) is Himawari-8 (about 2 km) and GPM (about 5.2 km), applied. In this method, cloud-free areas are de- we used average BT of Himawari-8 in each observa- fined where three IR bands (BT8.6, BT10.4 and tion pixel of GPM KuPR. Because Himawari-8 scans BT12.4). Since threshold depends on region and every 10 min, we collected simultaneous Himawari-8 season, no fixed value can be used. Therefore, and GPM observations where the difference in the a procedure for dynamic retrieval of a proper observation times between Himawari-8 and GPM was threshold was used with reference to above within 5 min. RF classification may perform poorly previous research. In this method, histograms of when learning from an extremely unbalanced dataset BT distribution are created as the simultaneous (Liu et al. 2006) in the case when the rain samples observation of Himawari-8 and GPM exceeds accounts for approximately 20 % of all samples, while 250,000 pixels for each season and area. Since only 30 % of the rain samples correspond to the heavy cloud-free pixels and cloud-contaminated pixels rain class. Kühnlein et al. (2014) suggested a tech- 694 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Fig. 2. Schematic diagram for making an RF model for the rain/no rain classification.

nique to address an imbalance in the ratio between the simultaneous observational datasets are separated into two classes. In this method, sample numbers are re- two datasets by random selection: training (1/10 of the duced by random selection so that the sample number entire dataset) and validation (9/10 of the entire data- in the large class does not exceed a certain ratio based set). The ratio of 1:9 is based on the result of Kühnlein on the sample number of the small class. Since this et al. (2014) who found that reducing the training data method is for mitigating the imbalance of the number by one tenth of the original sample greatly improves of samples between two classes with RF classification, calculation speed but hardly decreases estimation ac- it is not necessary for RF regression without class curacy. Because the number of original samples used separation. Following this approach, we corrected the in this study is much larger than that used in the previ- sample ratio between two classes by choosing samples ous paper, we decided that a sample ratio of 1:9 would randomly from each class at the fixed ratio shown in be sufficient to maintain the estimation accuracy. Each Table 2. The optimal class ratio parameters in Table RF model is created using the training datasets, and 2 are the values when the classification accuracy is the validation data were used to determine the above maximized. For validation of the technique, these optimal class ratio parameters. The Kanto–Tohoku June 2019 H. HIROSE et al. 695

Table 2. Class ratio parameters used within the RF models for the rain/no rain and rain-type classifications. Each class is sampled separately by RF models. Here, R denotes the number of observations of raining pixels and C indicates the number of observations of strong rain pixels. Step Parameter name Land-AS Ocean-AS Sample size rain 0.1 × R 0.1 × R (i) Rain area Sample size no rain 2.0 × 0.1 × R 1.6 × 0.1 × R Sample size strong rain 0.1 × C 0.1 × C (ii) Rain type Sample size weak rain 0.9 × 0.1 × C 0.9 × 0.1 × C

heavy rainfall event, which is used for the case study variables (P) randomly selected from the Q predictor in Section 4, is removed from the training data. variables. In the training process of the random forest method, only the training dataset is used. An out-of- 3.3 Random forest box (OOB) sample is used to calculate the estimate The basic mechanism of the RF method is first to error when determining the optimum combination of create many classification and regression trees based the explanatory variables or the optimum value of P. on training data, and then to obtain the final prediction The determination of P is made after the combina- result by majority vote (classification, rain or no rain, tion of the optimum explanatory variables has been weak or strong rain) or averaging (regression; rain decided. Because we determine the optimal hyper- rate) of each tree (Breiman and Cutler 2013). The RF parameter using a parameter sweep to minimize the algorithm begins the random selection by replacing estimate error, we believe there is a small influence of bootstrap samples with learning samples, as shown uncertainty due to this artificial choice. The method in an example of the RF classification (Fig. 2). Next, for determining a set of explanatory variables this when creating a tree from each bootstrap sample, a study used is described in Section 4.1. A sensitivity split is created at each node to separate two classes analysis was performed to determine the optimum most appropriately. If the number of bootstrap samples value of P because the best value of P depends on the created is 500, the final number of classification trees problem. The value of P was increased by 1 from P (ntree) is also 500, forming a forest. For example, if = 1 to P = Q with ntree = 500, and then the optimum ntree = 500, and 400 trees predict rain and the remain- value was determined based on the RF model with the ing 100 trees predict no rain, the final prediction result highest estimate accuracy. The optimum value of P is will be rain. Once the classification trees are created, determined using the method shown in Fig. 3. First, a the classification result can be obtained at any time by different subset of roughly 2/3 of the Himawari-8 and inputting the observations of Himawari-8 into the RF GPM match-up dataset is used to make each tree, as model. In the case of estimating rain rate, a regression illustrated in Fig. 2. The remaining 1/3 of the dataset tree is created instead of the classification tree and the is not used for learning but is stored as the OOB final prediction result is obtained by averaging predic- sample for testing (Breiman 2001). Because the OOB tion results of all regression trees (Breiman 2001). data contain the rainfall observation result of GPM KuPR, it is possible to calculate the estimation error 3.4 Tuning of RF models by inputting the OOB data into the RF models created Before executing the RF model, adjustment of two using each P. Therefore, the value of P leading to the parameters is necessary. The first parameter is ntree smallest OOB error is selected as the appropriate RF and an increase in the ntree leads to a decrease in esti- model. Ultimately, we chose the value of P for the mation error, but the computational load also increases rain/no rain classification as 4, the value for the rain- (Breiman 2001). Kühnlein et al. (2014) have shown type classification as 5, and the value for the rain rate that 500 trees are sufficient to produce a stable rain/no estimation as 6. rain classification and rain rate estimation. Following As summarized in Table 1, the number of observa- Kühnlein et al. (2014), we set ntree to 500 in all the tion bands of Himawari-8 in the IR region was RF models. substantially increased from three in the past GEO to The second parameter is the total number of nine. To account for the two different IR band combi- predictor variables (Q) and the number of predictor 696 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Fig. 3. Schematic diagram for calculating OOB error.

9! nations, there are C = = 36 combination 4. Results 9 2 29!( -2)! bands and nine individual bands, as shown in Table 3. 4.1 Selecting explanatory variables used in RF Therefore, to learn with a high level of accuracy with models the smallest possible calculation cost, it is necessary Figure 4 shows the importance of the explanatory to determine the optimal combination of explanatory variables calculated during boreal summer (August/ variables. The importance of each explanatory vari- September 2015/July 2016) from each stepwise able is determined with the same method used to scheme for rainfall-rate assignment and shows the top determine parameter P. However, one difference is five in the 9 single IR band and the 36 IR band com- that a data string of a specific explanatory variable is binations (Table 3) in descending order of importance. sorted randomly in the OOB data to make synthetic Horizontal axis of Fig. 4 indicates classification error OOB data. When the RF model is applied to these and route mean square error (RMSE) averaged over synthetic OOB data, the estimation accuracy is small- all trees when rainfall is estimated without a specific er than that of the original OOB data. If the decline index. The rain area classification part (Fig. 4a) and rate of the estimation accuracy for these two OOB rain type classification part (Fig. 4b) show the clas- datasets is larger than that of any other variables, the sification error without the BT10.4 index is largest. explanatory variable is considered to be the most This means BT10.4 is most important explanatory important. variable for rain area and rain type classification. On the other hand, for rain rate estimates in strong rain June 2019 H. HIROSE et al. 697

Table 3. Combination of Himawari-8 observation band that be considered as an explanatory variable. Numbers indicate ob- servation center wavelength. µm 6.2 6.9 7.3 8.6 9.6 10.4 11.2 12.4 13.3 6.2 6.2 6.2 – 6.9 6.2 – 7.3 6.2 – 8.6 6.2 – 9.6 6.2 – 10.4 6.2 – 11.2 6.2 – 12.4 6.2 – 13.3 6.9 6.9 6.9 – 7.3 6.9 – 8.6 6.9 – 9.6 6.9 – 10.4 6.9 – 11.2 6.9 – 12.4 6.9 – 13.3 7.3 7.3 7.3 – 8.6 7.3 – 9.6 7.3 – 10.4 7.3 – 11.2 7.3 – 12.4 7.3 – 13.3 8.6 8.6 8.6 – 9.6 8.6 – 10.4 8.6 – 11.2 8.6 – 12.4 8.6 – 13.3 9.6 9.6 9.6 – 10.4 9.6 – 11.2 9.6 – 12.4 9.6 – 13.3 10.4 10.4 10.4 – 11.2 10.4 – 12.4 10.4 – 12.4 11.2 11.2 11.2 – 12.4 11.2 – 12.4 12.4 12.4 12.4 – 13.3 13.3 13.3

Fig. 4. (a) Classification error averaged over all trees when classifying rainfall or no rainfall without a specific index. Here, a large error shown on the horizontal axis indicates the index has a large influence on the classification accuracy. Similar to panel (a) but (b) for the classification of rain type, (c) route mean square error averaged over all trees for estimating the weak rain rate without a specific index. Similar to panel (c) but (d) for estimating the strong rain rate. ΔT represents the BT difference between the two IR channels, and the subscripts indicate the center wavelengths of the observation bands. cases, ΔT7.3 – 13.3 is the most important explanatory indices are smaller than those in Fig. 4a. Thies et al. variable (Fig. 4d). Both BT10.4 and ΔT7.3 – 13.3 are (2008a) reported that the indices of ΔT8.6 – 10.8 and indices on CTH, and the importance of CTH indices ΔT10.4 – 12.4 were effective for obtaining information shown in Fig. 4 is consistent with previous studies concerning CWP. Therefore, we used ΔT8.6 – 10.4 and (Schmetz et al. 1997; Reudenbach et al. 2001). ΔT10.4 – 12.4 as indices of CWP. From the results in Therefore, this study could regard our importance of Figs. 4c and 4d, it is seen that ΔT6.9 – 7.3 is the most explanatory variables as reliable and used BT10.4 and important variable for estimating the weak rain rate ΔT7.3 – 13.3 as indices of CTH. and is an important factor with ΔT7.3 – 13.3 of the Conversely, in Figs. 4b and 4d, the differences in CTH index for estimating strong rain intensity. The ΔT importance between the four indices (ΔT8.6 – 10.4, of the WV bands in the upper and middle troposphere ΔT10.4 – 12.4, ΔT6.9 – 7.3, and ΔT6.2 – 6.9) and CTH using BT6.9 became possible, for the first time, by 698 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Table 4. Composite band list of Himawari-8 used for RF Table 5. Results of accuracy indicators for four types of machine learning. ΔT represents the difference in BT be- satellite rainfall-estimating data. These indicators are cal- tween the two observation bands, and the suffixes identi- culated by comparing each satellite estimation result with fy the center wavelength of each observation band. the radar–AMeDAS observations for the case of Kanto– Tohoku heavy rainfall event. The results of GSMaP_ Predictor variables MWR and HRA are calculated on 17:00 GMT and that BT10.4 Cloud top height of GSMaP_IR is calculated on 16:00 GMT September 9. ΔT7.3 – 13.3 For the HRA, accuracy indicators were also calculated ΔT8.6 – 10.4 on 16:00 GMT, but the value change was very small. Re- Cloud water path ΔT10.4 – 12.4 sults were calculated separately for the case of extreme rain within the square dashed line in Fig. 5 (139 – 140.5°E, ΔT6.2 – 6.9 Water vapor 34.5 – 39°N) and the outside region (137 – 145°E, 33 – ΔT6.9 – 7.3 46°N). (mm hr−1) ME MAE RMSE Extreme −0.06 3.10 6.03 Himawari-8. We used ΔT6.2 – 6.9 and ΔT6.9 – 7.3 as GSMaP_MWR Outside −0.16 0.42 1.20 indices for the WV vertical profile information. Given Extreme −2.70 2.75 6.67 the above, we determined the optimal combination of GSMaP_IR the Himawari-8 observation bands and identified the Outside 0.10 0.52 1.16 explanatory variables for our machine-learning algo- Extreme −0.52 1.69 3.47 HRA rithm based on the results shown in Table 4. Outside −0.33 0.61 1.07 HRA Extreme −2.66 2.71 6.03 4.2 Verifying accuracy in the Kanto–Tohoku heavy (BT10.4 only) Outside 0.38 0.59 1.24 rainfall event We compared the estimation results of HRA with the radar–AMeDAS rainfall for the case of the Kanto– Tohoku heavy rainfall event. The Kanto–Tohoku 3.10 mm h−1, and 6.03 mm h−1, respectively. However, heavy rainfall event refers to five days from Septem- GSMaP_IR could not estimate the heavy rainfall due ber 7 to 11, 2015, in the Kanto–Tohoku region (139.5 – to the lack of MWR observations in the left panel of 140°E, 36 – 38°N) of Japan. The accumulated rainfall Fig. 5b. The ME, MAE, and RMSE of GSMaP_IR exceeded 500 mm. The Meteorological Research for extreme rainfall were −2.70 mm h−1, 2.75 mm h−1, Institute (MRI) of JMA reported that this heavy rain- and 6.67 mm h−1, respectively. It is concluded that fall event occurred because of a moist band of WV GSMaP_IR has a systematic underestimation error brought from Tropical Cyclone Etau west of Japan and for extreme rainfall because the result of GSMaP_IR Typhoon Vamco southeast of Japan (Meteorological shows that the value of |ME| is as large as the value of Research Institute 2015). Radar–AMeDAS obser- MAE. Conversely, the estimated rainfall intensity of vations showed extremely heavy rain within a very HRA, shown in Fig. 5c, shows good agreement with narrow band in the Kanto–Tohoku region (139.5 – the radar–AMeDAS result. The ME, MAE, and RMSE 140°E, 36 – 38°N) (Fig. 5a). As indicators show the of HRA for extreme rainfall were −0.52 mm h−1, 1.69 estimate accuracy of the rainfall, mean error (ME = mm h−1 and 3.47 mm h−1, respectively. To confirm bias), mean absolute error (MAE) and RMSE were the importance of the Himawari-8 multiband IR, we calculated and are shown in Table 5. The results were re-estimated the HRA results using only BT10.4 of calculated separately for the extreme rainfall within the CTH information. Figure 5d shows that the HRA the square dashed line in Fig. 5 (139 – 140.5°E, 34.5 – with only BT10.4 cannot estimate heavy rainfall, and 39°N) and outside that region (137 – 145°E, 33 – 46°N). it is very similar to the result of GSMaP_IR. The ME, The Advanced Microwave Scanning Radiometer MAE, and RMSE for extreme rainfall were −2.66 2 (AMSR2) and Advanced Microwave Sounding mm h−1, 2.71 mm h−1 and 6.03 mm h−1, respectively. Unit (AMSU-A)/Microwave Humidity Sounder The RMSE was relatively small in the areas other than of GSMaP_MWR were able to estimate the heavy those with extreme rainfall, and the difference in the rainfall band well, indicating the good performance accuracy between GSMaP and HRA was small. of MWR in the right panel of Fig. 5b. The ME, MAE We examined the ETH of the Kanto–Tohoku heavy and RMSE of GSMaP_MWR for extreme rainfall rainfall event using GPM KuPR observations. Figure calculated using radar–AMeDAS were −0.06 mm h−1, 6a indicates the radar–AMeDAS rain rate and the June 2019 H. HIROSE et al. 699

Fig. 5. Estimated rainfall results for the Kanto–Tohoku heavy rainfall event from (a) radar–AMeDAS and (b) GSMaP_NRT. The result at 16:00 GMT was complemented by GEO IR and that at 17:00 GMT was observed by MWR. The dashed line indicates the boundary of the two different MWR satellites. (c) HRA. (d) HRA estimated without the SW and WV information shown in Table 3. The results of two consecutive observations are shown from left (16:00 GMT) to right (17:00 GMT). The dashed circular frame lines show the effective range of the ground-based radar. The square dashed line shows the region where extreme rain exists.

KuPR observing range (the gray-shaded region). It lower than approximately 4 km. can be seen in Fig. 6b that the ETH defined by the In this case, the indices of ETH and CTT do not threshold of 17 dBZ was abruptly lowered in the area correspond to the radar−AMeDAS rain rate; however, south of 37.5°N and that the CTT indicated by BT10.4 Fig. 6c shows that the values of ΔT10.4 – 12.4 and increased toward the south side. However, the radar– ΔT6.9 – 7.3 (SW and WV indices respectively) have AMeDAS rain rate abruptly increased in the area peak values near 37.0°N at the same location as the south of 37.5°N. Figure 6b indicates there was weak maximum rain rate. Since variations in the other WV rain from clouds with ETH higher than approximately indices (ΔT6.2 – 6.9, ΔT6.2 – 7.3, BT6.2, BT6.9, and 4 km and there was heavy rain from clouds with ETH BT7.3 shown in Fig. 6d) do not correspond to the 700 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Fig. 6. (a) The top panel shows the radar–AMeDAS rain rate, and the gray shade indicates KuPR observation swath onboard GPM core observatory. (b) The blue line shows latitudinal variations in the ETH calculated by GPM KuPR, the red line shows BT10.4 observed by Himawari-8, and the black line shows the rainfall intensity observed by the radar–AMeDAS averaged longitudinally in the square area in panel (a). (c) Same as panel (b), but the blue line shows ΔT6.9 – 7.3 of the WV bands and the pink line shows ΔT10.4 – 12.4 of the SW channels. (d) Same as panel (b), but the blue and yellow dashed lines indicate ΔT6.2 – 6.9 and ΔT6.2 – 7.3, respectively. The black, red, and green lines indicate BT6.2, BT6.9, and BT7.3, respectively. (e) vertical profiles of RH and horizon-

tal winds from ERA-Interim. The solid line indicates d θ e /dz from 850 hPa to 500 hPa. All results are snapshots at the nearest time to 14:00 GMT when the GPM satellite passed over in its orbit. June 2019 H. HIROSE et al. 701 radar−AMeDAS rain rate and behave similarly to In Fig. 7b, variations of the rainfall intensity in the CTT in Fig. 6b, ΔT6.9 – 7.3 is considered a special Kanto−Tohoku area, estimated from HRA, show good index in WV indices. Figure 6e shows the vertical agreement with the radar–AMeDAS observations, profile of the relative humidity (RH) and horizontal and the correlation coefficient is 0.72 (Fig. 7b). The winds by ERA-Interim data, and there is a region with time series of the GSMaP_NRT rain intensity shows RH lower than 50 % in the southerly wind region of an extremely sharp increase and decrease. As the the middle troposphere (higher than 600 hPa). The microwave observation passes, heavy rainfall can be increase in ΔT6.9 – 7.3 corresponds to the decrease in estimated by GSMaP_MWR; however, in other time RH of the middle troposphere. In addition, as the RH zones, the non-MWR observation (i.e. GSMaP_IR) of the middle troposphere declines, the convective rainfall intensity tended to be underestimated, and the instability indicated by the vertical gradient of the correlation coefficient from GSMaP_IR only is 0.07. equivalent potential temperature (d θ e /dz) of the lower Considering the above results, HRA is more suitable troposphere (from 850 hPa to 500 hPa) increases. We than GSMaP_NRT to capture rapid changes in local speculate that the heavy rainfall from the warm-type heavy rainfall. precipitating clouds in the Kanto–Tohoku rainfall event was also caused by a high convective instability 4.3 Statistical analysis due to the vertical WV gradient. Statistical analysis was performed to confirm The temporal variations in the Kanto–Tohoku whether the relationship among ΔT6.9 – 7.3, RH, rainfall event are indicated in Fig. 7. Figure 7a shows and heavy rain, as seen in the case of the Kanto– the radar–AMeDAS rain rate of the Kanto–Tohoku Tohoku rainfall event, can be observed in other cases. heavy rainfall event that accumulated from 12:00 on Figure 8a shows mean rainfall intensity as a function September 8 to 03:00 on September 10, 2015 (GMT). of ΔT10.4 – 12.4 versus ΔT6.9 – 7.3 for warm-type The rain rate in the Kanto–Tohoku area, indicated by precipitating clouds with a CTT of 252 K higher. a square area in Fig. 7a (139.5 – 140°E, 36 – 38°N), Mean rainfall intensity was calculated from the radar– rapidly increased after 00:00 GMT on September 9 AMeDAS composite and Himawari-8 simultaneous and sustained a heavy rain rate for a long period from observations during boreal summer (August/Septem- 12:00 GMT to 18:00 GMT on September 9 (Fig. 7b). ber 2015/July 2016) over land near Japan (120 – 150°E,

Fig. 7. (a) Radar–AMeDAS rain intensity of the Kanto–Tohoku heavy rainfall event that accumulated from 12:00 on September 8 to 03:00 on September 10, 2015 (GMT). (b) Time series of the rainfall intensity averaged in the square area (139.5 – 140°E, 36 – 38°N) in panel (a). The black line represents observations from the ground- based radar, the green line represents estimated values from HRA, the blue line represents observed values from GSMaP_NRT, and the red dashed lines indicate times when the MWR satellites passed over the analysis region. 702 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Fig. 8. Radar–AMeDAS rainfall intensity as a function of ΔT10.4 – 12.4 versus (a) ΔT6.9 – 7.3, (b) ΔT6.2 – 7.3, (c) ΔT6.2 – 6.9, (d) BT6.2, (e) BT6.9, and (f) BT7.3 for clouds with CTT greater than 252 K during boreal summer (August/September 2015/July 2016).

20 – 50°N). As ΔT10.4 – 12.4 is small, the precipitating to using only CTT information. Similar results are cloud is optically thick, and the average rain rate is obtained in the case of ΔT6.2 – 7.3 (Fig. 8b), which is greater than approximately 6 – 9 mm h−1. Further, for slightly less sensitive to the middle-tropospheric WV such an optically thick raincloud, as ΔT6.9 – 7.3 is compared to ΔT6.9 – 7.3. In Fig. 8b, maximum rainfall small, the rain rate increases to approximately 25 – 30 intensity as a function of ΔT6.2 – 7.3 does not reach 30 mm h−1 or more. This result shows that using the mm h−1. This result indicates that the index ΔT6.2 – 7.3 optical thickness (ΔT10.4 – 12.4) and WV (ΔT6.9 – 7.3) is less sensitive to rainfall intensity in warm-type information increases the estimate accuracy compared precipitating cloud than that of ΔT6.9 – 7.3. In other June 2019 H. HIROSE et al. 703

in warm-type precipitating clouds. Therefore, we conclude that SW and WV information are indis- pensable for detecting heavy rain from warm-type precipitating clouds. Considering the WV band of conventional GEOs was focused only on WV in the upper troposphere, the newly added WV bands (BT6.9 and BT7.3) of Himawari-8 captured the change in the vertical WV content of the middle troposphere. We compared the estimate accuracies of HRA and GSMaP_IR during the same period. Using radar– AMeDAS as a ground truth, Hit (H) indicates the number of pixels where the RF model can correctly estimate the observed rainfall. Here, the rain pixel indicates that the observed or estimated rain rate is greater than x mm h−1. False alarm (F) indicates the number of pixels where the RF model erroneously estimated rainfall even though rain was not observed, and a miss (M) indicates the number of pixels where the observed rainfall could not be estimated by the RF model. Based on these elements, Threat score (TS) and Bias score (BS) are calculated as follows:

Fig. 9. Vertical profile of RH as a function of Hx TSx = , (1) ΔT6.9 – 7.3, calculated by using simultaneous Hx ++Fx Mx observations of Himawari-8 and ERA-Interim Hx + Fx during boreal summer (August and September BSx = . (2) 2015 and July 2016) near Japan (130 – 150°E, Hx + Mx 30 – 50°N). This result is only from precipitat- ing clouds with BT10.4 greater than 252 K and Figure 10 shows the TS ratio between GSMaP_NRT ΔT10.4 – 12.4 less than 3 K. The solid line indi- and HRA calculated with radar–AMeDAS as truth

cates d θ e /dz from 850 hPa to 500 hPa. during (a – c) boreal summer (August and September 2015 and July 2016) and (d – f) boreal winter (Decem- ber 2015 and January and February 2016). Figure 11 is same as Fig. 10, but the BS is shown. HRA showed WV indices ΔT6.2 – 6.9, BT6.2, BT6.9, and BT7.3, the higher TS than the GSMaP_IR in the case, including rainfall intensity reaches, at most, approximately 20 all rain intensity and CTT. Additionally, we divided mm h−1 and the performance for classifying heavy rain the result into three cases with rain rates greater than further decreases, as shown in Figs. 8c – f respectively. 1.0 mm h−1, 5.0 mm h−1, and 10.0 mm h−1. The TS To investigate why ΔT6.9 – 7.3 has high sensitivity difference for the warm rain case is larger than that to rain intensity in warm-type precipitating clouds, for the cold rain case, and this tendency is clearer in we analyzed the atmospheric environment using summer than in winter. The BS results in Fig. 11 also ERA-Interim data. In Fig. 9, the convective instability shows that HRA lowers the risk of underestimating d θ e /dz of the lower troposphere (from 850 hPa to 500 warm-type rain. The TS difference for the strong rain hPa) increases as the RH of the middle troposphere case (10 mm h−1) is relatively larger than that for the (from 650 hPa to 400 hPa) decreases, similar to the weak rain cases (5 mm h−1 or 1 mm h−1). Conversely, result in our case study of the Kanto–Tohoku heavy the TS for clouds with low CTT is clearly lower in rainfall event. The increases in ΔT6.9 – 7.3 from −3 K winter than in summer regardless of rain intensity. to 3 K correspond to the decrease in the RH in the This appears to be due to the difficulty in estimating middle troposphere. From these results, it was statis- the snow rate. These results suggest that HRA’s es- tically suggested that ΔT6.9 – 7.3 detected dry air in timation accuracy is higher for heavy rain produced the middle troposphere and the potentially unstable from warm-type precipitating clouds in the Asian atmospheric condition. This seems to be the reason monsoon region compared to conventional rainfall why ΔT6.9 – 7.3 has high sensitivity to rain intensity estimation methods based only on CTT. 704 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Fig. 10. TS calculated with radar–AMeDAS as truth during (a – c) boreal summer (August/September 2015/July 2016) and (d – f) boreal winter (December 2015 and January and February 2016). The red lines indicate the TS of HRA, and the green lines indicate the TS of GSMaP_IR. The horizontal axis indicates that TS is calculated only from clouds with CTT higher than the threshold. The results were calculated with rain rates greater than 1.0, 5.0, and 10.0 mm h−1, respectively.

4.4 Weighting functions of Himawari-8 WV bands the presence of clouds and by the WV amounts above retrieved by RTTOV the clouds. We used Radiative Transfer for TIROS The results, so far, have shown that the index Operational Vertical sounder (RTTOV: version 12.1; ΔT6.9 – 7.3 is important for estimating warm-type James et al. 2019) to calculate the weighting functions heavy rain. Next, we performed a simple radiation of the newly available Himawari-8 WV multiband. analysis to investigate how ΔT6.9 – 7.3 is affected by RTTOV is the fast radiative transfer model for TOVS June 2019 H. HIROSE et al. 705

Fig. 11. Same as Fig. 10 but BS is calculated.

and was developed at ECMWF in the early 1990s shown below. to simulate the satellite spectrum radiance of TOVS (i) The composite WV and temperature profile aver- IR and MWR. We used the ERA-Interim reanalysis aged over all cases for precipitating clouds with a data for information on all the environmental field CTT lower than 252 K. input to RTTOV during boreal summer (August and (ii) The composite WV and temperature profile September 2015 and July 2016) over land near Japan averaged over all cases for warm-type precipi- (120 – 150°E, 20 – 50°E) as same as in the case of Fig. tating clouds with a CTT of 252 K or higher, a 8 in Section 4.3. To separately analyze the effects of ΔT10.4 – 12.4 of 0 K or higher, and a ΔT6.9 – 7.3 clouds and WV, the radiation transfer calculation by of 0 K or higher. RTTOV set up two types of environmental fields as Figure 12a shows the vertical profile of the com- 706 Journal of the Meteorological Society of Japan Vol. 97, No. 3 posite RH for cases (i) and (ii). It is clear in the shows that each WV band rapidly attenuates due to middle-to-upper troposphere (600 – 300 hPa) that the the large amount of WV and that the peaks of each RH in the warm-type heavy rain case (ii) is smaller weighting function are at the top of the troposphere. than that in case (i). To investigate how this differ- Conversely, Fig. 12c for case (ii) shows that each WV ence in the vertical distribution of the RH affects band reaches a relatively lower height and that each ΔT6.9 – 7.3, the weighting functions of BT6.2, BT6.9, weighting function peak is lower. Table 6 shows the and BT7.3 were calculated in the clear sky case (cloud BT of each WV band calculated by RTTOV. Paying radiative effect removed). Figure 12b for case (i) attention to the BT differences in the case of (ii) – (i)

Fig. 12. (a) The composite RH vertical profile calculated by ERA-Interim reanalysis data during boreal summer (August and September 2015 and July 2016) over land near Japan (120 – 150°E, 20 – 50°E). (i) In all cases for precipitating clouds with CTT lower than 252 K, (ii) in all cases for warm-type precipitating clouds with a CTT of 252 K higher and ΔT10.4 – 12.4 of 0 K higher and ΔT6.9 – 7.3 of 0K higher. (b) – (d) The weighting function of BT6.2, BT6.9 and BT7.3 calculated by RTTOV with WV and temperature profile obtained by the ERA-Interim reanalysis data. (b) in the case of (i) without cloud effects, (c) in the case of (ii) without cloud effects and (d) in the case of (ii) with cloud effects.

Table 6. BT6.2, BT6.9 and BT7.3 calculated by RTTOV by using ERA-Interim reanalysis data during boreal summer (August and September 2015 and July 2016) over land near Japan (120 – 150°E, 20 – 50°E). (i) In all cases for precipitating clouds with CTT lower than 252 K, (ii) in all cases for warm-type precipitating clouds with a CTT of 252 K higher and ΔT10.4 – 12.4 of 0 K higher and ΔT6.9 – 7.3 of 0K higher. Clear means without cloud effects and cloudy mean with cloud effects. (K) (i) clear (ii) clear (ii) cloudy (ii) – (i) (clear) (ii) cloudy–clear BT6.2 230.42 236.77 234.01 6.35 −2.76 BT6.9 239.49 245.03 241.08 5.54 −3.95 BT7.3 250.22 254.34 245.09 4.12 −9.25 ΔT6.9 – 7.3 −10.73 −9.31 −4.01 1.42 5.30 June 2019 H. HIROSE et al. 707

(clear sky), the attenuation of BT due to WV in the pause. Additionally, they reported that the above middle-to-upper troposphere is the largest in case of phenomenon was primarily caused by the influence BT6.2 and is the smallest in the case of BT7.3. There- of a calibration error. It was found from Fig. 3 in fore, the large amount of WV in the middle-to-upper Nishi et al. (2017) that, even though the calibration troposphere reduces ΔT6.9 – 7.3. error was smaller than in the case of MTSAT, ΔT10.4 Next, we investigated the effect of warm-type (HT) – 12.4 (LT) still had a negative value even for clouds on ΔT6.9 – 7.3. When considering the influence Himawari-8. Some of the causes of the negative value of clouds, we need to pay attention to cloud emissiv- of ΔT10.4 (HT) – 12.4 (LT) and the positive value of ity spectral dependence. However, in this study, we ΔT6.9 (LT) – 7.3 (HT) in this study are also thought to focused on the ΔT6.9 – 7.3, which is the difference be calibration errors. Conversely, the opposite sign of between the WV bands. Since the WV bands have ΔT in Fig. 8a occurs with strong rain. Therefore, we significantly lower transmittance to cloud than the speculate that warm-type heavy rain may lead to the atmospheric window band (10.4 – 12.4 μm), we opposite sign of ΔT due to a small temperature inver- thought that the influence of cloud emissivity spectral sion at the convective cloud top that is small enough dependence on ΔT6.9 – 7.3 was small. Therefore, we to be invisible at the resolution of the reanalysis data considered that the cloud height and the WV amount in the RTTOV analysis. Note that the opposite sign of above clouds were the main cause of ΔT6.9 – 7.3, ΔT has a negligible impact on our observation-based and RTTOV analysis was performed with the cloud rain estimates of HRA. emissivity of 1.0. An optically thick (emissivity = 5. Summary 1.0) warm-type cloud was added to the case of (ii) below 500 hPa and the radiative transfer calculation The HRA method constructed a rainfall-estimating was performed again. As shown in Fig. 11d, the product featuring high time and spatial resolutions of weighting function of BT6.9 was greatly influenced 10 min and 2 km, respectively, using the RF machine- by the added cloud and a highly intense peak devel- learning method and multiband observations of a oped at the cloud top. Looking at the BT differences third-generation GEO, Himawari-8, and rain observa- in case (ii) (cloudy – clear) in Table 6, it can be seen tion data obtained via KuPR on board the GPM core that BT7.3 is greatly reduced when interrupted by an observatory as training data. To investigate how IR added cloud. These results indicate another cause for multiband observations contribute to the accuracy of decreases in ΔT 6.9 – 7.3. To summarize, a small value rainfall estimation using Himawari-8, we conducted of ΔT6.9 – 7.3 indicates that the amount of WV in the a case study of the September 2015 Kanto–Tohoku middle-to-upper troposphere is small and that clouds heavy rainfall event in Japan. This case was a warm- have developed in the middle troposphere. type heavy rain in a potentially unstable condition. The value of ΔT6.9 – 7.3 in case (ii) (cloudy – clear) HRA was able to estimate detailed time variations in shown in Table 6 is smaller than that of other cases; the rainfall at a very high temporal resolution (every however, it does not have a positive value, unlike 10 min) when the GSMaP_NRT was not able to the observed value of ΔT6.9 – 7.3 shown in Fig. 8. estimate the heavy rainfall in the Kanto area during The difference ΔT occurs due to the difference in the the gap between the MWR overpasses under the same transmittance of the two IR bands for clouds or WV. conditions, as shown in Fig. 7b. As a result of this Specifically, because the high transmittance (HT) band case analysis, it was confirmed that HRA clearly has can receive IR radiation from lower altitudes, the high precision for the Kanto Tohoku heavy rain. In observed BT becomes higher and generates a ΔT the summer statistical analysis, it was found that HRA relative to the low transmittance (LT) band. Therefore, could classify warm-type precipitating clouds into op- in principle, ΔT10.4 (HT) – 12.4 (LT) should have tically thick clouds with strong rain and optically thin a value of zero or more (positive) and ΔT6.9 (LT) – clouds with weak rain, primarily using information 7.3 (HT) should have a value of 0 or less (negative). concerning the optical thickness from the ΔT10.4 – Schmetz et al. (1997) reported that ΔT WV (LT) – 12.4 SW bands. Using the index of ΔT6.9 – 7.3 in addi- IR (HT) usually only had a negative value under the tion to the SW band of ΔT10.4 – 12.4, HRA can judge standard tropospheric lapse rate; however, it had a whether the thick warm-type clouds resulted in heavy positive value with a strong tropopause inversion rain. Himawari-8 was the first GEO to include BT6.9 layer. Hamada and Nishi (2010) reported that ΔT10.8 and is sensitive to WV at lower levels than MTSAT. (HT) – 12.0 (LT) observed by MTSAT had a negative Further, as a result of the RTTOV calculations, it was value even for clouds that did not reach the tropo- found that the small value of ΔT6.9 – 7.3 suggested 708 Journal of the Meteorological Society of Japan Vol. 97, No. 3 there was dry air in the upper-middle troposphere and estimation; Its concepts and implementation for the developed low clouds. AMMA experiment. Ann. Geophys., 28, 1–20. Research on rain estimates using GEO multiband Bessho, K., K. Date, M. Hayashi, A. Ikeda, T. Imai, H. observations has been conducted primarily in Europe Inoue, Y. Kumagai, T. Miyakawa, H. Murata, T. based on MSG (e.g., Bergès et al. 2010; Kühnlein Ohno, A. Okuyama, R. Oyama, Y. Sasaki, Y. Shimazu, K. Shimoji, Y. Sumida, M. Suzuki, H. Taniguchi, H. et al. 2014). However, Himawari-8 is the first GEO Tsuchiyama, D. Uesawa, H. Yokota, and R. Yoshida, with more than 10 bands in the Asian monsoon region 2016: An introduction to Himawari-8/9 – Japan’s and is the first GEO with 3 WV bands in the world. new-generation geostationary meteorological satel- By developing HRA from Himawari-8, we found that lites. J. Meteor. Soc. Japan, 94, 151–183. an index using the WV multiband of ΔT6.9 – 7.3 could Breiman, L., 2001: Random forest. Mach. Learn., 45, 5–32. effectively estimate warm-type heavy rain in the Asian Breiman, L., and A. Cutler, 2013: Classification/clustering. monsoon region. Finally, we conclude that HRA has a Random forests. [Available at http://www.stat.berkeley. higher estimation accuracy for heavy rain from warm- edu/~breiman/RandomForests/cc_home.htm.] type precipitating clouds than a conventional rain Byers, H. R., and R. R. Braham, Jr., 1949: The Thunder- estimation methods based on only one IR band. From storm: Final Report of the Thunderstorm Project. U.S. these results, using HRA should improve the accuracy Government Printing Office, 282 pp. of global rainfall estimates by compensating for gaps Capacci, D., and B. J. Conway, 2005: Delineation of pre- in the microwave observation network. cipitation areas from MODIS visible and infrared imagery with artificial neural networks. Meteor. Appl., Acknowledgments 12, 291–305. Cermak, J., 2006: SOFOS–A new satellite-based operational This work was partly supported by the 8th Pre- fog observation scheme. Ph.D. Thesis, Philipps-Uni- cipitation Measurement Mission/RA from the Japan versity Marburg, 147 pp. Aerospace Exploration Agency (JAXA) and Virtual Cermak, J., and J. Bendix, 2008: A novel approach to fog/ laboratory/Ministry of Education, Culture, Sports, low stratus detection using Meteosat 8 data. Atmos. Science, and Technology. Himawari-8 data were Res., 87, 279–292. provided by CEReS, Chiba University, Japan. GSMaP Dee, D. P., S. M. Uppala, A. J. Simmons, P. Berrisford, P. products were produced by the Earth Observation Poli, S. Kobayashi, U. Andrae, M. A. Balmaseda, G. Research Center (EORC) in JAXA. We used the RF Balsamo, P. Bauer, P. Bechtold, A. C. M. Beljaars, library (package “randomForest”) created by Liaw L. van de Berg, J. Bidlot, N. Bormann, C. Delsol, R. and Wiener (2002) included in R statistical computing Dragani, M. Fuentes, A. J. Geer, L. Haimberger, S. B. Healy, H. Hersbach, E. V. Holm, L. Isaksen, P. Kåll- software (http://www.r-project.org/). berg, M. Köhler, M. Matricardi, A. P. McNally, B. M. References Monge-Sanz, J.-J. Morcrette, B.-K. Park, C. Peubey, P. de Rosnay, C. Tavolato, J.-N. Thepaut, and F. Vitart, Adler, R. F., A. J. Negri, P. R. Keehn, and I. M. Hakkarinen, 2011: The ERA-Interim reanalysis: Configuration and 1993: Estimation of monthly rainfall over Japan and performance of the data assimilation system. Quart. J. surrounding waters from a combination of low-orbit Roy. Meteor. Soc., 137, 553–597. microwave and geosynchronous IR data. J. Appl. Ferraro, R. R., F. Weng, N. C. Grody, L. Zhao, H. Meng, C. Meteor., 32, 335–356. Kongoli, P. Pellegrino, S. Qiu, and C. Dean, 2005: Aminou, D. M. A., 2002: MSG’s SEVIRI instrument. ESA NOAA operational hydrological products derived from Bull., 111, 15–17. [Available at http://www.esa.int/ the Advanced Microwave Sounding Unit (AMSU). esapub/bulletin/bullet111/chapter4_bul111.pdf.] IEEE Trans. Geosci. Remote Sens., 43, 1036–1049. Aonashi, K., J. Awaka, M. Hirose, T. Kozu, T. Kubota, G. Hamada, A., and N. Nishi, 2010: Development of a cloud- Liu, S. Shige, S. Kida, S. Seto, N. Takahashi, and Y. N. top height estimation method by geostationary satellite Takayabu, 2009: GSMaP passive microwave precipi- split-window measurements trained with CloudSat tation retrieval algorithm: Algorithm description and data. J. Appl. Meteor. Climatol., 49, 2035–2049. validation. J. Meteor. Soc. Japan, 87A, 119–136. Hamada, A., and Y. N. Takayabu, 2016: Improvement in Arkin, P. A., and P. E. Ardanuy, 1989: Estimating climatic- detection of light precipitation with the global pre- scale precipitation from space: A review. J. Climate, 2, cipitation measurement dual-frequency precipitation 1229–1238. radar (GPM DPR). J. Atmos. Oceanic Technol., 33, Barrett, E. C., 1970: The estimation of monthly rainfall from 653–667. satellite data. Mon. Wea. Rev., 98, 322–327. Hamada, A., Y. N. Takayabu, C. Liu, and E. J. Zipser, 2015: Bergès, J. C., I. Jobard, F. Chopin, and R. Roca, 2010: Weak linkage between the heaviest rainfall and tallest EPSAT-SG: A satellite method for precipitation storms. Nat. Commun., 6, 6213, doi:10.1038/ncomms June 2019 H. HIROSE et al. 709

7213. ing occurrence of Kanto-Tohoku heavy rain September Hou, A. Y., R. K. Kakar, S. Neeck, A. A. Azarbarzin, C. D. 2015. 5 pp. [Available at http://www.mri-jma.go.jp/ Kummerow, M. Kojima, R. Oki, K. Nakamura, and T. Topics/H27/270918/press20150918.html.] Iguchi, 2014: The global precipitation measurement Nishi, N., A. Hamada, and H. Hirose, 2017: Improvement of mission. Bull. Amer. Meteor. Soc., 95, 701–722. cirrus cloud-top height estimation using geostationary Inoue, T., 1985: On the temperature and effective emissivity satellite split-window measurement trained with determination of semi-transparent cirrus clouds by CALIPSO data. SOLA, 13, 240–245. bi-spectral measurements in the 10μm window region. Reudenbach, C., G. Heinemann, E. Heuel, J. Bendix, and M. J. Meteor. Soc. Japan, 63, 88–99. Winiger, 2001: Investigation of summertime convec- James, H., P. Rayer, D. Rundle, R. Saunders, M. Matricardi, tive rainfall in Western Europe based on a synergy of A. Geer, P. Brunel, and J. Vidot, 2019: RTTOV v12 remote sensing data and numerical models. Meteorol. users guide. 150 pp. [Available at https://www. Atmos. Phys., 76, 23–41. nwpsaf.eu/site/download/documentation/rtm/docs_ Richards, F., and P. Arkin, 1981: On the relationship rttov12/users_guide_rttov12_v1.3.pdf.] between satellite-observed cloud cover and precipita- Japan Meteorological Agency, 2003: JMA HRIT mission tion. Mon. Wea. Rev., 109, 1081–1093. specific implementation. Issue 1.2, 63 pp. [Available at Roebeling, R. A., and I. Holleman, 2009: SEVIRI rainfall https://www.data.jma.go.jp/mscweb/en/operation/type/ retrieval and validation using weather radar observa- HRIT/JMA_HRIT_Issue1.2.pdf.] tions. J. Geophys. Res., 114, D21202, doi:10.1029/ Kida, S., T. Kubota, S. Shige, and T. Mega, 2017: Devel- 2009JD012102. opment of a rain/no-rain classification methods over Schmetz, J., S. A. Tjemkes, M. Gube, and L. van de Berg, land for the microwave sounder algorithm. Remote 1997: Monitoring deep convection and convective Sensing of Aerosols, Clouds, and Precipitation. Islam, overshooting with METEOSAT. Adv. Space Res., 19, T., Y. Hu, A. Kokhanovsky, and J. Wang (eds.), Else- 433–441. vier, 362 pp. Seto, S., and T. Iguchi, 2011: Applicability of the iterative Kojima, M., T. Miura, K. Furukawa, Y. Hyakusoku, T. Ishi- backward retrieval method for the GPM dual- kiri, H. Kai, T. Iguchi, H. Hanado, and K. Nakagawa, frequency precipitation radar. IEEE Trans. Geosci. 2012: Dual-frequency precipitation radar develop- Remote Sens., 49, 1827–1838. ment on the global precipitation measurement (GPM) Shige, S., T. Yamamoto, T. Tsukiyama, S. Kida, H. core observatory. Proc. SPIE, 85281, 85281A, doi: Ashiwake, T. Kubota, S. Seto, K. Aonashi, and K. 10.1117/12.976823. Okamoto, 2009: The GSMaP precipitation retrieval Kühnlein, M., T. Appelhans, B. Thies, and T. Nauß, 2014: algorithm for microwave sounders–Part I: Over-ocean Precipitation estimates from MSG SEVIRI daytime, algorithm. IEEE Trans. Geosci. Remote Sens., 47, nighttime, and twilight data with random forests. J. 3084–3097. Appl. Meteor. Climatol., 53, 2457–2480. Shige, S., S. Kida, H. Ashiwake, T. Kubota, and K. Aonashi, Kummerow, C., W. Barnes, T. Kozu, J. Shiue, and J. Simp- 2013: Improvement of TMI rain retrievals in moun- son, 1998: The Tropical Rainfall Measuring Mission tainous areas. J. Appl. Meteor. Climatol., 52, 242–254. (TRMM) sensor package. J. Atmos. Oceanic Technol., Shige, S., M. K. Yamamoto, and A. Taniguchi, 2015: 15, 809–817. Improvement of TMI rain retrieval over the Indian Lensky, I. M., and D. Rosenfeld, 2003: A night-rain delin- subcontinent. Remote Sensing of the Terrestrial Water eation algorithm for infrared satellite data based on Cycle. Geophys. Monogr., No.206, American Geo- microphysical considerations. J. Appl. Meteor., 42, physical Union, 27–42. 1218–1226. Smith, E. A., J. E. Lamm, R. Adler, J. Alishouse, K. Liaw, A., and M. Wiener, 2002: Classification and regression Aonashi, E. Barrett, P. Bauer, W. Berg, A. Chang, R. by randomForest. R News, 2, 18–22. Ferraro, J. Ferriday, S. Goodman, N. Grody, C. Kidd, Liu, Y., N. V. Chawla, M. P. Harper, E. Shriberg, and A. D. Kniveton, C. Kummerow, G. Liu, F. Marzano, A. Stolcke, 2006: A study in machine learning from Mugnai, W. Olson, G. Petty, A. Shibata, R. Spencer, imbalanced data for sentence boundary detection in F. Wentz, T. Wilheit, and E. Zipser, 1998: Results of speech. Comput. Speech Lang., 20, 468–494. WetNet PIP-2 Project. J. Atmos. Sci., 55, 1483–1536. Makihara, Y., 2000: Algorithms for precipitation nowcasting Sohn, B. J., G.-H. Ryu, H.-J. Song, and M.-L. Ou, 2013: focused on detailed analysis using radar and rain­ Characteristic features of warm-type rain producing gauge data. Tech. Rep. of the Meteorological Research heavy rainfall over the Korean peninsula inferred Institute, 39, 63–111. from TRMM measurements. Mon. Wea. Rev., 141, Makihara, Y., N. Uekiyo, A. Tabata, and Y. Abe, 1996: 3873–3888. Accuracy of Radar-AMeDAS precipitation. IEICE Spencer, R. W., 1984: Satellite passive microwave rain rate Trans. Commun., E79-B, 71–762. measurement over croplands during spring, summer Meteorological Research Institute, 2015: Factors determin­ and fall. J. Climate Appl. Meteor., 23, 1553–1562. 710 Journal of the Meteorological Society of Japan Vol. 97, No. 3

Taniguchi, A., S. Shige, M. K. Yamamoto, T. Mega, S. Kida, 17. T. Kubota, M. Kachi, T. Ushio, and K. Aonashi, 2013: Upadhyaya, S., and R. Ramsankaran, 2016: Modified- Improvement of high-resolution satellite rainfall INSAT multi-spectral rainfall algorithm (M-IMSRA) product for Typhoon Morakot (2009) over Taiwan. J. at climate region scale: Development and validation. Hydrometeor., 14, 1859–1871. Remote Sens. Environ., 187, 186–201. Tebbi, M. A., and B. Haddad, 2016: Artificial intelligence Ushio, T., K. Sasashige, T. Kubota, S. Shige, K. Okamoto, systems for rainy areas detection and convective cells’ K. Aonashi, T. Inoue, N. Takahashi, T. Iguchi, M. delineation for the south shore of Mediterranean Sea Kachi, R. Oki, T. Morimoto, and Z. Kawasaki, 2009: during day and nighttime using MSG satellite images. A Kalman filter approach to the Global Satellite Atmos. Res., 178–179, 380–392. Mapping of Precipitation (GSMaP) from combined Thies, B., T. Nauß, and J. Bendix, 2008a: First results on passive microwave and infrared radiometric data. J. a process-oriented rain area classification technique Meteor. Soc. Japan, 87, 137–151. using Meteosat Second Generation SEVIRI nighttime Yamamoto, M. K., and S. Shige, 2015: Implementation of data. Adv. Geosci., 16, 63–72. an orographic/nonorographic rainfall classification Thies, B., T. Nauß, and J. Bendix, 2008b: Precipitation scheme in the GSMaP algorithm for microwave radi- process and rainfall intensity differentiation using Me- ometers. Atmos. Res., 163, 36–47. teosat Second Generation SEVIRI data. J. Geophys. Yamamoto, M. K., S. Shige, C.-K. Yu, and L.-W. Cheng, Res., 113, D23206, doi:10.1029/2008JD010464. 2017; Further improvement of the heavy orographic Toyoshima, K., H. Masunaga, and F. A. Furuzawa, 2015: rainfall retrievals in the GSMaP algorithm for mi- Early evaluation of Ku- and Ka-band sensitivities for crowave radiometers. J. Appl. Meteor. Climatol., 56, the Global Precipitation Measurement (GPM) Dual- 2607–2619. frequency Precipitation Radar (DPR). SOLA, 11, 14–