Evaluation of CMPA precipitation estimate in the evolution of typhoon-related storm rainfall in Guangdong, China

The merged precipitation data of Climate Prediction Center Morphing Technique and gauge observations (CMPA) generated for continental China has relatively high spatial and temporal resolution (hourly and 0.1 W ), while few studies have applied it to investigate the typhoon-related extreme rainfall. This study evaluates the CMPA estimate in quantifying the typhoon-related extreme rainfall using a dense rain gauge network in Guangdong Province, China. The results show that the event-total precipitation from CMPA is generally in agreement with gauges by relative bias (RB) of 2.62, 10.74 and 0.63% and correlation coef ﬁ cients (CCs) of 0.76, 0.86 and 0.91 for typhoon Utor, Usagi and Linfa events, respectively. At the hourly scale, CMPA underestimates the occurrence of light rain ( < 1 mm/h) and heavy rain ( > 16 mm/h), while overestimates the occurrence of moderate rain. CMPA shows high probability of detection (POD ¼ 0.93), relatively large false alarm ratio (FAR ¼ 0.22) and small missing ratio (0.07). CMPA captures the spatial patterns of typhoon-related rain depth, and is in agreement with the spatiotemporal evolution of hourly gauge observations by CC from 0.93 to 0.99. In addition, cautiousness should be taken when applying it in hydrologic modeling for ﬂ ooding forecasting since CMPA underestimates heavy rain ( > 16 mm/h).


INTRODUCTION
Accurate measurement of precipitation with fine resolution and reasonable accuracy is essential but challenging (Li et al. ). Automatic rain gauge networks can provide near-real time measurements, while the representativeness of a sparse and uneven distribution of gauges is worthy of consideration (Li & Shao ). Weather radar is capable of monitoring local precipitation under fine spatial resolution and reliable quality, although it is affected by atmospheric conditions and high terrain in mountainous areas (Zhang et al. a). With the development of satellite precipitation retrievals, high resolution and multi-sensor precipitation pro- Satellite precipitation estimates have considerable importance in many hydrological and meteorological applications (Tian et al. ). The satellite-based precipitation products usually need verification before being applied in hydrologic modeling (Turk et al. ). The comprehensive evaluation to satellite precipitation products has significant implications. On the one hand, the knowledge of error characteristics is beneficial to the improvement of retrieval

Study area
The study area is in the Guangdong Province of southern

Typhoon events
There were a total of seven typhoon events which made landfall in Guangdong Province from 2013 to 2015 (Table 1) (Table 1).
Typhoon Rammasun was a long-lived storm, which origi- It brought an average of 59 mm of rain to the western areas of Guangdong Province (Table 1).

Gauge data
The dense rain gauge network is maintained by the Depart-

METHODOLOGY Normal comparisons
The typhoon-affected rainfall is first defined according to the If the distance of a rain gauge from the typhoon center during any hour is less than 400 km, that particular hour of that gauge is defined as a 'typhoon hour'. The typhoon-related storms normally last more than 24 hours (Table 1) where n is the number of gauges; G and M are the gauge observations and CMPA estimates, respectively.
The positive bias and negative bias will be offset in RB, which is used to describe the accumulative bias of the

Contingency statistics
Two contingency statistics, probability of detection (POD) and false alarm ratio (FAR), are used to measure the accuracy between CMPA and gauge observations. Each pair of hourly rainfall records from gauges and CMPA grids are classified as a Hit, Miss, or False ( where N hit is the total count of hours when both gauge and CMPA observe rain, which is equal or larger than the rain rate threshold (P 0 ¼ 0.1 mm/h). N miss is the total count of hours when CMPA misses rain that is detected by the gauge. N false is the total count of hours when CMPA detects rain that is not detected by the gauge. PDFc and PDFv are computed using the 'Hit' pairs of gauges and CMPA grids in Table 3.

K-means cluster analysis
The K-means cluster algorithm has been applied to charac-

Events total rainfall
The event total precipitation from gauges and CMPA for the three events of Utor, Usagi and Linfa are illustrated in Figure 2.    Hourly gauge-CMPA comparison The spatial distributions of CC, RB and RMSD reveal error characteristics of the hourly CMPA estimate in different    In the storm center (Type 1) for the Utor event, the peak rain rate occurs in the 9th-11th hours, when the Utor landed. CMPA greatly underestimates the peak rain and leads to a smaller event total rain by À15.54%, although it has a similar estimate prior to and post the peak time  They are slightly   larger than the gauge observations to Type 2 and Type 3 for the Usagi event and Type 3 for the Linfa event. There were similar patterns for the three types of hyetographs of the auxiliary four events (Table 4 and Figure A6). In other words, CMPA underestimate the intense rain and overestimate the light rain (Table 4).

SUMMARY AND REMARKS
The   large FAR and small missing ratio. The miss count has a larger proportion in the low rain rate ranges, leading to a smaller mean rain rate for the miss count than that of the false count.
All gauges are further classified into three types of hyetograph, which form a spatial pattern that corresponds to the three stages of the storm evolution, the initial landfall (storm center), advance and the dissipation. Even though the performance of CMPA may vary from case to case, CMPA generally underestimates rainfall of Type 1 and overesti- to total rainfall volume that might not be properly represented by gauges, especially in mountainous areas.