Yong Chen , Jing Dun , Junling An , Huizhi Liu , Ulrih G?rsdorf , Frnz H. Berger

a State Key Laboratory of Atmospheric Boundary Layer Physics and Atmospheric Chemistry, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China

b State Key Laboratory of Severe Weather & Key Laboratory for Cloud Physics of China Meteorological Administration, Chinese Academy of Meteorological Sciences,

Beijing, China

c Meteorologisches Observatorium Lindenberg/Richard A βmann Observatory, Deutscher Wetterdienst, Lindenberg, Germany

Keywords:Satellite rainfall estimation Rainfall variability Micro Rain Radar TRMM

A B S T R A C T Uncertainties in satellite rainfall estimation may derive from both the local rainfall characteristics and its subpixel variability. To study this issue, Micro Rain Radars and a rain gauge network were deployed within a 9-km satellite pixel in the semi-arid Xilingol grassland of China in summer 2009. The authors characterized the subpixel variability with the coefficient of variation (CV) and evaluated the satellite rainfall estimation for this semi-arid area. The results showed that rainfall events with a high CV were mostly convective with a small amount of rainfall. Spatially inhomogeneous rainfall was most likely to occur at the edges of small clouds producing rain. The performance of the TRMM (Tropical Rainfall Measuring Mission) 3B42V7 product for daily rainfall was better than that of the CMORPH (Climate Prediction Center morphing technique) and PERSIANN (Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks) products, although the TRMM product tended to overestimate rainfall in a lake area of the semi-arid grassland.

1. Introduction

Satellite estimates of rainfall are important in semi-arid regions where rain gauge (RG) and weather radar networks are sparse, but accurate knowledge of the spatial distribution of rainfall is needed. Different climate zones are characterized by different types of rain, so algorithms for the estimation of rainfall from satellite measurements need to consider the impact of the rainfall characteristics in a particular climate zone ( Liu et al., 2005 ; Wang et al., 2016 ). In addition, the subpixel variability of rainfall on the ground may cause some uncertainty in the estimation of rainfall by satellites as a result of mismatch problems between point measurements and areal estimates ( Krajewski et al., 2003 ;Peters et al., 2005 ; Kirstetter et al., 2015 ). Therefore, the algorithms need to consider both the effect of different climate zones and the variability of rainfall at the subpixel scale.

The most widely used methods to estimate rainfall from satellite data are based on infrared (IR) sensors, microwave sensors, and precipitation radar ( Kidd and Levizzani, 2011 ). The IR method first discriminates clouds producing rain via a multi-IR channel and then estimates the rain rate (

R

) using the relationship between

R

and the IR brightness temperature (

T

) ( Kurino, 1997 ; Vicente et al., 1998 ; Ba and Gruber, 2001 ). However, this method contains some bias because only cloud-top information is obtained by the IR method. To obtain a more accurate spatial distribution of rainfall, global satellite rainfall estimation products combining IR, microwave, and precipitation radar data have been developed, such as the CMORPH (Climate Prediction Center morphing technique) ( Joyce et al., 2004 ), PERSIANN (Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks) ( Hsu et al., 1997 ; Hong et al., 2004 ), and TRMM (Tropical Rainfall Measuring Mission) ( Huffman et al., 2007 ) products. Evaluations of one or more of these rainfall estimation methods and products have been carried out worldwide ( Liu et al., 2005 ; Ebert et al., 2007 ;Zhang et al., 2016 ; Kimani et al., 2017 ; Wang et al., 2019 ; Peng et al.,2020 ), and some studies have focused on semi-arid zones ( Wei et al.,2018 ; Duan et al., 2019 ; Mahbod et al., 2019 ). However, there have been few studies that combined the evaluation of satellite rainfall data,error analysis for multiple products, and the local

R

—

T

IR relationships in semi-arid grasslands.The spatial coefficient of variation (CV: the ratio of the standard deviation to the arithmetic mean) can be used to characterize the subpixel variability in rainfall ( Pedersen et al., 2010 ). The CV changes with the timescale and the median CV for a 9-km box has been reported to decrease from 15 min to 3 h and then remain in the range 0.17—0.42 for timescales

>

3 h ( Chen et al., 2015 ). The CV also depends on the total amount and type of rainfall ( Pedersen et al. 2010 ; Jiao and Wang 2001 ),with a larger CV for small amounts of total rainfall and local heavy rainfall. However, the effect of different types of rain (convective, stratiform,and mixed) and the local high-impact synoptic situation on the CV is unclear.

Xilingol is a typical temperate grassland in a semi-arid region of northern China. Rainfall during the grass-growing period (May—September) is one of the key factors affecting the grassland ecosystem,and most of this rainfall is affected by the Mongolian cold vortex. The Sino—German collaborative project MAGIM (Matter Fluxes in Grasslands of Inner Mongolia as Influenced by Stocking Rate) was carried out in the Xilingol grassland area to measure the subpixel rainfall ( Chen et al.,2015 , 2016 ). Based on the characteristics of rainfall in this semi-arid grassland, the CV for different types of rain and high-impact synoptic situations (the Mongolian cold vortex) were studied. The impact of the macrophysical characteristics of clouds producing rain on the variability of rainfall was considered. The satellite rainfall products and the errors in multi-IR channels were evaluated and analyzed.

2. Measurements, data, and methods

2.1. Measurements

The field campaign for the MAGIM-P11 project was carried out from 18 May to 14 September 2009. Several Micro Rain Radars (MRRs) and RGs were placed within a 9-km satellite pixel (43.57°N, 116.67°E) in the Xilin River catchment, Inner Mongolia Autonomous Region, China( Fig. 1 ). Two intense observation periods (IOP), including IOP1 (18 May—10 July) and IOP2 (11 July—14 September), were conducted, with four MRRs (MRR1—MRR4) and three MRRs (MRR2—MRR4) used to represent 9-km and 4-km satellite pixels in this study ( Fig. 1 (b)). A more detailed overview of this field campaign has been reported previously( Chen et al., 2015 ).

Fig. 1. (a) Location of the Xilin River catchment and (b) topography of the field campaign area and locations of weather equipment. The numbers indicate the IDs of the MRRs. Location 1 on the map indicates the location of IMGERS (Inner Mongolia Grassland Ecosystem Research Station).

Based on the MRR measurements and meteorological observations,four rainfall types, including convective rainfall impacted by the normal synoptic situation (C), convective rainfall affected by a cold vortex (Cl),stratiform rainfall (S), and mixed rainfall (M), were classified. A rain event was defined as having started if rainfall

>

0.2 mm, whereas it was defined as having ended if without rainfall

>

0.2 mm during 6 h in the pixel. A rainfall event, with a bright band measured by the MRR, was defined as S-type if no cumulonimbus cloud was reported; otherwise, it was defined as M-type. Convective rainfall was defined as cumulonimbus cloud or thunderstorms having been recorded, but no bright band measured. A cold vortex was defined as a well-established closed low existing at the 500-hPa surface (Fig. S1); the continuing cold air can cause more showers from the western part of the cold vortex. The Cl type was selected independently in this study because (1) the Mongolian cold vortex is one of the main raining-weather systems, and (2) most showers caused by cold vortex in this semi-arid area.

Fig. 2. Box-and-whisker plots of the coefficient of variation (CV) of rainfall as a function of (a) rainfall type and (b) total rainfall based on the 24 rainfall events measured by MRRs at 300 m above ground level. The four rainfall types are convective rainfall impacted by the normal synoptic situation (C), convective rainfall affected by a cold vortex (Cl), stratiform rainfall (S), and mixed rainfall (M). The 9-km box is based on MRR1—4, whereas the 4-km box is based on MRR2—4.

Ten rainfall events (mean rainfall of four MRRs

>

1 mm) were collected in IOP1 (Table S1) and 14 rainfall events in IOP2 (Table 4 in Chen et al. (2015) ). The whole IOP included 12 C-, five Cl-, four M-,and three S-type rainfall events. The mean rainfall measured by the MRRs was 2.1—18.2 mm and 1.4—13.5 mm in IOP1 and IOP2, respectively —similar to that by the RGs (1.7—19.7 mm and 0.5—19.7 mm). All five Cl events were observed in July (IOP1) when the Mongolian cold vortex was active. The subpixel rainfall variability was characterized based on MRR observations.

2.2. Satellite rainfall data and methods

Firstly, satellite-based rainfall data of three summers were used to evaluate their availability in this semi-arid area. Then, based on the detailed ground measurements in summer 2009 in section 2.1 , the sources of possible errors and improvements of the IR method were discussed.

Two satellite datasets in the summers of 2007—2009 were used in this study: (1) The data from the Chinese Fengyun (FY-2C) geostationary meteorological satellite (Table S2; Wang and Guo, 2014 ; Zhang et al., 2019 )were mapped at a resolution of 0.05°×0.05°with an interval of 1 h; the image navigation of FY-2C and six IR satellite rainfall estimation methods based on FY-2C are listed in Table S3. (2) The global rainfall products of CMORPH, PERSIANN, and TRMM_3B42V7 ( Hong et al., 2004 ;Huffman et al., 2007 ; Joyce et al., 2004 ) at a resolution of 0.25°×0.25°with an interval of 3 h.

The daily rainfall estimated by satellite in a 0.25°×0.25°box was mainly evaluated by using two RG observations, at Xilinhot station and IMGERS (Inner Mongolia Grassland Ecosystem Research Station), with the statistics including the correlation coefficient (COR), the probability of detection (POD), and the false alarm ratio (FAR) ( Ebert et al., 2007 ).Finally, based on the above evaluation and error analysis for multiple products, combining the 1-h

R

—

T

IR relationships and criteria of raining cloud detection measured by MRR and FY-2C in summer 2009 at IMGERS, some reasons for the IR method’s availability and limitations in the Xilingol grassland were analyzed.

3. Results and discussion

3.1. Subpixel rainfall variability

Fig. 2 shows the CVs of the 4- and 9-km boxes depending on rainfall type and total rainfall measured by MRRs at 300 m above ground level.The median CV of the rainfall type in the 4-km box generally decreased in the order C

>

Cl

>

M

>

S; and a similar trend in the median CV was found in the 9-km box, except for M and S ( Fig. 2 (a)). The mean CVs of the 9-km box for convective (C and Cl), mixed (M), and stratiform (S)rain events were 0.36, 0.27, and 0.20, respectively; whereas those of the 4-km box were 0.26, 0.11, and 0.05, respectively. This shows that the spatial variability of the rainfall depends on both the spatial scale and the type of rainfall —for example, a high CV was observed at a larger spatial scale and for convective rainfall events.The rain events were divided into two classes, with a total rainfall of ≤ 5 and

>

5 mm ( Fig. 2 (b)). For both the 4- and 9-km boxes, the CV of the total rainfall ≤ 5 mm had a wider range than that of rainfall

>

5 mm. The 4-km median CV for small amounts of rainfall ( ≤ 5 mm)was larger than that for heavy rain (

>

5 mm), whereas opposite results were obtained for the 9-km boxes ( Fig. 2 (b)). The mean CVs of the 9-km(4-km) boxes for small amounts of rainfall ( ≤ 5 mm) and heavy rain rainfall (

>

5 mm) were 0.36 (0.22) and 0.28 (0.18), respectively. This shows that high CVs were observed in most of the events with low total rainfall.

The maximum CV in the 9-km box occurred during the event on 29 July 2009. This event was convective, and the total rainfall was ≤ 5 mm, with one station recording continuous rain and the other station without rain during a 1-h period, although the two stations were only 6 km apart. The CV for rainfall affected by the cold vortex depended on the timescale. The Cl event on 28—29 June included 12 showers and lasted for 10—20 min, with rain areas of 6—10 km; the hourly CV in the 9-km box was very high but was small on a daily scale. This may have been caused by a small rainfall cloud passing randomly across the 9-km pixel area on an hourly scale, but with a more equal probability for this 9-km box on a daily scale.

The mean CV in the arid region of China was 0.3—0.5, with a maximum CV up to 1.1 ( Jiao and Wang, 2001 ). High CVs have also been reported for rainfall events with low total rainfall in Denmark( Pedersen et al., 2010 ). Similar results were found in our study, with the mean and maximum CV in the 9-km box being 0.33 and 1.20,respectively; a higher CV occurred with small amounts of rainfall( ≤ 5 mm).

The local movement of rainstorms can affect the variability of the subpixel rainfall ( Huffand Shipp, 1969 ). High-resolution (1.25-km)cloud maps in the visible channel were used to investigate the impact of the movement of raining clouds on the spatial variability of rainfall for these rain types ( Fig. 3 and Fig. S3).

Two C events, on 29 July and 17 August, with relatively small raining clouds had higher CVs ( Fig. 3 (a, b) and Fig. S3). Both C events had a number of similar characteristics: (1) the rain systems moved from southwest to northeast; (2) the horizontal shape of the raining systems was a long strip; and (3) the measurement field was located at the margin of the raining cloud (Fig. S3). The C event on 18 July, with a large(mesoscale) cloud producing rain, had a lower CV ( Fig. 3 (c)). The convection cells for rainfall events affected by the cold vortex with a high hourly CV ( Fig. 3 (d, e)) could be distinguished in the western region of the cold vortexes in the 1.25-km visible-channel cloud map, but were difficult to find in the 5-km IR cloud map as a result of smaller clouds and surface radiation (Fig. S1). The stratiform event on 16 July had a lower CV due to the long timescale over which the cloud produced rain( Fig. 3 (f)). These results suggest that (1) the CV may be affected by the area of the cloud producing rain and its movement; and (2) a larger subpixel variability in rainfall tends to occur at the edge of small convective clouds producing rain.

Fig. 3. Images from the FY-2C satellite in the visible channel at 1.25-km resolution. The black point denotes the location of the 9 km ×9 km pixel, and the white box denotes the Xilin River catchment. Panels (a—e) show convective rainfall events and (f) a stratiform rainfall event. The distribution of rainfall for the 9 km ×9 km pixel is highly inhomogeneous in panels (a) and (b). Convective showers in (d) and (e) were affected by the cold vortex over Mongolia.

Fig. 4. Total rainfall estimated by the (a) CMORPH, (b) PERSIANN, (c) TRMM, (d) AE, and (e) AEW products, and (f) that observed by RGs in summer 2009.

3.2. Evaluation of satellite rainfall estimation

The evaluation of daily rainfall estimated by satellite is presented in Table S4 and Fig. 4 . The interannual variation of total summer rainfall followed the order 2007

<

2009

<

2008.

The CMORPH, PERSIANN, and TRMM products had similar amounts of rainfall to the observations, but only the TRMM product captured the interannual variation in rainfall. The TRMM product gave the highest COR and the smallest FAR; the PERSIANN product gave the highest POD, the lowest COR, and the largest FAR; the CMORPH product gave an intermediate performance (Table S4). Our results are similar to those reported by Ebert et al. (2007) and Tong et al. (2014) . The good performance of the TRMM product was probably due to its incorporation of RG observation data, which are not included in the CMORPH and PERSIANN products ( Tian et al., 2009 ; Tong et al., 2014 ).

The methods of the AE (auto-estimator) function (Table S3) using the

R

—

T

relationship were better than those methods of the GMSRA(GOES multispectral rainfall algorithm) function (Table S3) based on total summer rainfall, which was mainly caused by the difficulty to detect the raining area by the four methods of GMSRA (Fig. S4 and related discussion in the Supplementary Material). Note that the AE and GMSRA functions fit deep and shallow convective clouds based on our observation, respectively (Fig. S5), implying that it was difficult to capture all rainfall using only one

R

—

T

relationship, due to the complex nature of the rainfall in this semi-arid area. The

R

—

T

IR1 relationship of the AE function gave the best results in July because shallow convective clouds that were producing rain (e.g., a rain event caused by cold vortexes)occurred less often in July than in June and August (Fig. S1 and Fig.S5).

Although the AE and AEW (auto-estimator weighted) methods gave poorer results than the PERSIANN product for the total summer rainfall,they had a similar COR, POD, and FAR to this product. The relative humidity—revised method, AEW, improved the skill of the AE method,increasing the POD and decreasing the FAR (Table S4). The estimation of monthly rainfall by satellite was improved after revising the relative humidity, especially in western China during the dry season ( Liu et al.,2005 ). Similar results were also found in this study, where the relativehumidity revision performed better in dry years (2007 and 2009) than in a rainy year (2008).

The spatial distribution of summer rainfall decreased from south to north for the five products and methods, and the TRMM product was the closest to RG observations ( Fig. 4 ). However, TRMM overestimated the rainfall over Dali Lake, considering the small difference in rainfall between Dali Lake and Xilinhot observed by Bao et al. (2006) . Note that overestimation of rainfall by the TRMM product was also apparent over a lake area in Africa ( Kimani et al., 2017 ), although there was a smaller difference between lakes and land areas in a humid area of China ( Zhang et al. 2016 ). These results may have been caused by a disturbance in the form of a strong dry—wet contrast over land—lake areas(Dali Lake) in this semi-arid region by the TRMM product.

4. Summary

The subpixel (4 and 9 km) variability of summer rainfall was characterized for different rainfall types and total rainfall in a semi-arid grassland based on MRR observation. The effects of the size and movement of clouds producing rain on the spatial variability of different types of rainfall were analyzed using cloud maps. Satellite rainfall estimation products and methods were mainly evaluated based on RG observations. The primary findings can be summarized as follows:

Rainfall events with a high CV were usually convective with a small amount of rainfall. The mean CV for convective, mixed, and stratiform rainfall within a 9-km (4-km) grid box were 0.36 (0.26), 0.27 (0.11),and 0.20 (0.05), respectively. The mean CV for ≤ 5 and

>

5 mm of rainfall within a 9-km (4-km) grid box were 0.36 (0.22) and 0.28 (0.18),respectively. For shallow convective rainfall affected by the cold vortex,the hourly CV was high, while the daily CV was small. The spatial size and movement of small clouds producing rain affected the subpixel CV,and spatially inhomogeneous rainfall was most likely to occur at the edge of small clouds producing rain.

The TRMM 3B42V7 product was better than the CMORPH and PERSIANN products for estimating rainfall in the Xilin River catchment in terms of its correlation with the RG measurements and other verification metrics for daily rainfall; whereas, the TRMM 3B42V7 product tended to overestimate rainfall in the lake area in the grassland.

Based on the relatively complex rainfall characteristics of midlatitude semi-arid areas found in our results, accurate satellite rainfall estimation for this area still needs to consider the impacts from both rainfall subpixel variability and local rainfall synoptic situations in the future. Note that, corresponding to the IR spatial resolution of the sub-satellite point in the latest-generation geostationary meteorological satellites (e.g., 2 and 4 km in Himawari-8 and FY-4A ( Bessho et al.,2016 ; Zhang et al., 2019 )), the IR spatial resolution of satellite pixels in the midlatitude area was still at ～4 and ～9 km because of its long distance away from the equator. Therefore, our results for subpixel rainfall variability with a 4- and 9-km grid box could be used as a reference for the rainfall estimations of these satellites in midlatitude semi-arid areas.

Funding

This work was funded by the National Key R&D Program of China[grant number 2017YFC1501404], the German Research Foundation[Research Unit 536], and the National Natural Science Foundation of China [grant number 41675137].

Acknowledgments

We thank Dr Wanbiao Li from Peking University for providing help in FY-2C data processing. The comments from the two anonymous reviewers were greatly appreciated.

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.aosl.2021.100055 .