Daily rainfall data collected from more than 100 gauges over Thailand for the period 1993–2002 are used to study the climatology and spatial and temporal characteristics of Thailand rainfall variations. Comparison of the Thailand gauge (TG) data binned at 1° × 1° with the Global Precipitation Climatology Centre (GPCC) monitoring product shows a small bias (1.11%), and the differences can be reconciled in terms of the increased number of stations in the TG dataset. Comparison of daily TG with Tropical Rainfall Measuring Mission (TRMM) version 6 (V6) 3B42 rain estimates shows improvements over version 5 (V5) in terms of bias and mean absolute difference (MAD). The V5 is computed from the adjusted Geostationary Operational Environmental Satellite (GOES) precipitation index (AGPI) and V6 is computed using the TRMM Multisatellite Precipitation Analysis (TMPA) algorithm. The V6 histogram is much closer to that of TG than V5 in terms of rain fraction and conditional rain rates. Scatterplots show that both versions of the satellite products are deficient in capturing heavy rain events. In terms of detecting rain events, a critical success index (CSI) shows no difference between V6 and V5 3B42. The CSI for V6 is higher for the rainy season than the dry season. These results are generally insensitive to rain-rate threshold and averaging periods. The temporal and spatial autocorrelation of daily rain rates for TG, V6, and V5 3B42 are computed. Autocorrelation function analyses show improved temporal and spatial autocorrelations for V6 compared to TG over V5 with e-folding times of 1, 1, and 2 days, and isotropic spatial decorrelation distances of 1.14°, 1.87°, and 3.61° for TG, V6, and V5, respectively. Rain event statistics show that the V6 3B42 overestimates the rain event durations and underestimates the rain event separations and the event conditional rain rates when compared to TG. This study points to the need to further improve the 3B42 algorithm to lower the false detection rate and improve the estimation of heavy rainfall events.
Southeast Asia is the rice bowl of the world. As the world’s leading exporter of rice, Thailand’s agriculture is heavily dependent on rain. A better understanding of the spatial and temporal rainfall distributions is therefore essential for the Thai economy.
While rain gauge measurements are usually used to tune hydrologic models, they are limited by their spatial coverage. A network of weather radars provides good spatial and temporal coverage. However, the problem of inter-radar calibration and blockage by mountains still limits its capability. Remote sensing techniques using spaceborne sensors provide an excellent complement to continuous monitoring of rain events both spatially and temporally. The launch of the Tropical Rainfall Measuring Mission (TRMM) satellite in November 1997 by the National Aeronautics and Space Administration (NASA) of United States and the Japanese Aerospace Exploration Agency (JAXA) has provided more than 9 yr of quality rainfall data for tropical areas. The recent decision to extend TRMM operations to 2009 further increases its potential value for application use. To improve the TRMM data quality, TRMM products are periodically reprocessed, as new information acquired by TRMM leads to improved algorithms. The most recent TRMM reprocessing was completed using the version 6 (V6) algorithms. It is necessary to quantify the biases and uncertainties of the TRMM products relative to gauge measurements that have been used historically to calibrate hydrological and agricultural models.
There are a number of efforts to intercompare TRMM rainfall products with other rainfall measurements (Adler et al. 2003; Chiu et al. 2006b, c; Nicholson et al. 2003a, b). However, these studies are usually limited to comparisons at the monthly scale. For hydrological or agricultural applications, shorter periods, such as daily, 5-day (pentad), or 10-day (decade) rain rates are more appropriate. Comparison of satellite and gauge estimates at the submonthly scales have been performed, for example, by Brown (2006) for the India–Sri Lanka area and by Islam and Uyeda (2006) over Bangladesh. More recently, comparison of operational high-resolution rain products at the daily scale have been carried out by the International Precipitation Working Group (IPWG) for the continental United States, Europe, and Australia (Ebert 2002; Ebert et al. 2007; Turk et al. 2006). The comparison over Asia and in Thailand, in particular, is lacking.
Ten years of daily rain gauge data over 105 stations over Thailand have been acquired. A comparison with the Global Precipitation Climatology Centre (GPCC) gauge analysis allows an assessment of the quality of the Thailand gauge (TG) data. The comparison with TRMM products at the submonthly scale enables an evaluation of the applicability of satellite rain estimates for agricultural and hydrological applications in Thailand. Both version 5 (V5) and V6 of TRMM and other satellite merged rainfall (3B42) are included in the comparison. The V5 3B42 is an IR estimated rain rate calibrated to TRMM Combined Instrument (TCI), whereas the V6 is mainly based on microwave rain estimate from all available sensors, all calibrated to the TCI, and then merged with gauge measurements at the submonthly scale.
Because of their design, each methodology has different spatial and temporal characteristics. The comparison of TRMM rainfall products with independent gauge measurement would aid in the refinement of rainfall algorithms for TRMM and the planned Global Precipitation Measurement Mission (GPM).
The purposes of this work are to study the rainfall climatology and to examine the temporal and spatial characteristics of rainfall over Thailand. In addition, the utility of satellite rainfall estimates is evaluated by comparing TRMM rainfall algorithms (V5 and V6) with rain gauge data. The datasets are described in section 2. Section 3a discusses the seasonal and regional variations of TG climatology. Monthly and daily comparison of TRMM data products with TG data are described in sections 3b and 3c, respectively. Conclusions and discussions are contained in section 4.
The data include the TG data, GPCC monitoring product, and V5 and V6 of TRMM 3B42 and 3B43. The data are discussed below.
a. Thailand gauge data
Thailand rain gauge data are obtained from the Thai Meteorological Department (TMD). Ten years (1993–2002) of daily rain rates are collected from 105 rain gauge stations over Thailand. Figure 1 shows the location of the gauge stations separated into five regions: north, central, northeast, east, and south. These five regions are classified by their climatic conditions and geography by the TMD. The data are first binned into 1° × 1° boxes, with each gridded box having at least one gauge station. There are 15 boxes with at least three gauges. The rain rates are computed by calculating the mean of all gauges inside the grid box.
b. Global Precipitation Climatology Centre monitoring product
As part of the Global Precipitation Climatology Project (GPCP), the GPCC is responsible for producing precipitation estimates over land based on gauge analysis. A monitoring product is produced shortly after the month ends based on a subset of about 7000 stations (Rudolf et al. 1994). The product is directly accessible via the GPCC Web site (http://gpcc.dwd.de). The gauges used in the GPCC analysis are not fixed and the number of gauges varies with time. The number of gauges for each month at each grid box is given in the GPCC product. From the number of gauges, we estimated there are approximately 52 gauges over Thailand included in the GPCC analysis during the period 1993–2002.
c. TRMM 3B42 and 3B43 version 5 (V5) algorithm
The major rain measuring sensors for the TRMM are Precipitation Radar (PR), TRMM Microwave Imager (TMI), and Visible and Infrared Scanner (VIRS). A TCI algorithm (TCA; 2B31) was also developed that make use of the attenuation measured by TMI as a constraint for PR rain profiles at the PR swath. Kummerow et al. (2000) provided a summary of the TRMM project and instrument performance. A list of TRMM standard, value-added, and ancillary products is available from the Goddard Distributed Active Archive Center (GDAAC) Precipitation Data and Information Services Center (PDISC) Web site (http://daac.gsfc.nasa.gov/precipitation/data_access.shtml) and they are described in Chiu et al. (2006a).
The V5 3B42, TRMM-calibrated IR rain product, is included in our analysis because it is the only product in this work that has no direct gauge analysis input. Coincident TCA and VIRS (1B01) data are analyzed to establish transfer coefficients between TCA and VIRS. This relation is used to calibrate IR estimates from geosynchronous satellite infrared data to form the V5 3B42 product, following the adjusted Geostationary Operational Environmental Satellite (GOES) precipitation index (AGPI; Adler et al. 1994). The V5 3B42 algorithm provides daily precipitation and root-mean-square (RMS) error estimates at 1° × 1° latitude–longitude grids in the TRMM domain 40°N–40°S (Adler et al. 2000).
The V5 3B43 product is produced by merging the monthly accumulation of daily V5 3B42 with the monthly accumulated Climate Assessment and Monitoring System (CAMS) or GPCC rain gauge analysis (3A45) following Huffman et al. (1995). The monthly accumulation of 3B42 and 3A45 is weighted by the inverse of their respective random error fields (Huffman et al. 1997). The V5 3B43 algorithm provides 1° × 1° gridded monthly precipitation product.
d. TMPA (3B42) and 3B43 version 6 (V6) algorithm
The V6 3B42 is a 3-hourly, 0.25° product based on TRMM Multisatellite Precipitation Analysis (TMPA; Huffman et al. 2007). First, all available TCI-calibrated microwave estimates [from TMI, Special Sensor Microwave Imager (SSM/I), Advanced Microwave Scanning Radiometer (AMSR), and Advanced Microwave Sounding Unit (AMSU)] are put into the appropriate space–time bins. Microwave-calibrated IR rain estimates fill the remaining bins. These instantaneous estimates are summed over a calendar month to create a monthly multisatellite (MS) product. The MS and gauge analysis (GPCC or CAMS) are merged optimally (Huffman et al. 1997) to create a post-real-time satellite-gauge (SG) monthly 0.25° product, which is the TRMM V6 3B43 product. The final V6 3B42 is computed by scaling the intermediate 3B42 by the ratio of monthly MS to SG [the scale factor being limited to a range (0.2, 2)] The gauge analyses (GPCC or CAMS) employed are presented on a 2.5° or 1° grid, so that the finescale spatial and temporal information in the 0.25° 3B42 and 3B43 data are attributed to satellite inputs.
It should be emphasized that the differences between V5 and V6 are 1) V5 is based on IR rain estimates calibrated to the V5 TCA while the V6 is based on microwave rain estimates calibrated to V6 TCA, 2) V6 TCA rain rate is lower than V5 TCA (Chiu et al. 2006c), and 3) V6 includes gauge analysis and V5 does not. In the following discussion, V6 3B42 is synonymous with TMPA.
In this section, the climatology of Thailand rainfall is first described. The comparison between TG and GPCC is included in the monthly comparison section to show the consistency and reliability of the TG data. To compare these rain estimates, the following statistical measures are used—bias and root-mean-square difference (RMSD). Mean absolute difference (MAD) is also included as it is deemed a more appropriate measure for comparison (Willmott and Matsuura 2005). These are defined as follows:
where n is the total number of samples, i = 1, . . . , n and x is the algorithm rain rate, and TG is the Thailand gauge analyses for the grid box. Only data covering the Thailand region are included for comparison. The comparisons are carried out at 1° × 1° latitude–longitude boxes for the periods of overlap (1998–2002). The daily TG, V5, and V6 3B42 data are averaged to monthly 1° × 1° for monthly comparison. Correlation, cumulative distribution function (CDF), scatter diagram, and skill scores are also used for the diagnostics.
a. Thailand rainfall climatology
Figure 2 shows the monthly rainfall climatology for the different regions of Thailand computed from the TG. In the north, northeast, central, and east regions, there are three seasons: winter (November to February), hot (March to May/June), and rainy (May/June to October). Conversely, there are only two seasons in the south region: rainy (May/June to December) and hot (January to May).
There are two rainfall peaks during the rainy season: one in May and the other in August or September, except in the south region, which shows a minimum in February and a maximum in November. The secondary rainfall peak in May is associated with the southwest monsoon, during which the southwesterly wind from the Indian Ocean brings moist air toward Thailand. Some of the moisture is blocked by the mountain ranges in Burma. In August the northeasterly trade winds blow in the reverse direction, causing the August to September maximum rainfall peak. For the onset, the data show the rainfall pattern in the northward progression—the early rainfall peaks first in the east, northeast, central, and then north region. The end of the rainy season occurs from north to south, with the peak occurring in August in north and northeast regions and in September in central and east regions, consistent with the movement of the intertropical convergence zone (ITCZ) rainbands.
b. Monthly comparison
The climatology of Thailand rainfall from all rain estimates is shown in Fig. 3. All datasets are collected from 1998 to 2002. The 5-yr monthly averages of the TG, GPCC, V5 3B42, V6 3B42 are 143.00, 141.41, 170.89, and 139.51 mm month−1, respectively; V5 3B42 tends to be higher, especially in the rainy season. All satellite estimates show double peaks in May and August. The GPCC and V6 3B43 show low biases compared to V5 3B42, which is consistent with the design of V6 3B42. Except for the V5 3B42, all these products use the GPCC and/or CAMS gauge analyses in their production. The V5 3B42 is the highest and overestimates the TG measurements throughout most of the year (except in February). TG estimates are within 1.1% of GPCC estimates. The correlation between TG and GPCC at each grid box (using monthly 1993–2002 data) are significant at the 95% confident level. Since the GPCC monitoring product is not available at the daily scale, the daily TG data will be used in the daily comparison.
c. Daily comparison
In this section, the V5 is included to gauge the improvement of V6 over V5. Furthermore, it is the dataset without the gauge analysis for merging. The 5-yr (1998–2002) daily averages for TG, V6, and V5 3B42 are 4.73, 4.58, and 5.62 mm day−1, respectively. The bias, RMSD, and MAD are 0.89, 11.81, and 5.71 for V5 and −0.12, 11.89, and 5.02 for V6. The V6 3B42 shows better correspondence to TG with a smaller bias compared to V5. The regional biases, RMSDs, and MADs for all five regions indicate that the V6 3B42 is consistently lower than the V5 estimates, and hence shows smaller biases (Table 1). Except for the east region, the differences between V6 and TG are less than 0.5 mm day−1.
To examine the effect of the gridding on the statistics, bias statistics of 15 grid boxes that have at least three gauges are computed. The bias is 1.4% versus −2.6% for all the gauges. A Student’s t test of the difference between V6 and TG with at least three gauge boxes shows a t score of 1.08, which is not significant at the 95% level. Hence the hypothesis that there is significant difference in the bias between grid boxes with at least three gauges and all the other grid boxes is rejected. A similar t test of all grid boxes with two or more gauges show a t score of 1.2, again not significant at the 95% level.
Figure 4 shows the cumulative distribution functions (CDFs) of the daily rain rates binned at 0.5 mm day−1 interval. The zero rain rate (RR = 0) fractions are calculated to be 0.53, 0.40, and 0.36 and the conditional rain rates (rain rate when raining) are 10.06, 7.60, and 8.75 mm day−1 for TG, V6, and V5, respectively. Overall, both V5 and V6 overestimate the rain fraction and underestimate the conditional rain rates. The curves for V6 and V5 cross the TG curve at ∼10 and ∼20 mm day−1, respectively. In fact, both the V6 and V5 3B42 CDF have higher rain fraction (45% and 42%, respectively) than TG (33%) at rain rates between >0 and 10 mm day−1. It is concluded that the V6 algorithm shows improvements in rain probability over the V5 when compared with TG data.
The spatial and temporal structures of these rainfall products are investigated using the autocorrelation function of TG, V5, and V6. The generalized space–time lag correlation function is defined as
where X(s, t) and Y(s, t) are the rain rates of the different algorithms, s and t are the spatial and temporal coordinates, and δ and τ are the spatial and temporal lags, respectively;  denotes ensemble averaging over all space and time samples and σX and σY are the standard deviations of X and Y, respectively. The autocorrelation refers to the same variable, that is, X = Y, or in our case, X = TG or V5 or V6 3B42. The temporal autocorrelations are computed by setting δ = 0. Likewise, by fixing τ = 0 and varying δ, the spatial autocorrelations are computed. When both δ and τ are zero, the cross-correlation coefficient between X and Y (at zero lag) is computed.
The temporal autocorrelation functions are shown in Fig. 5. Both the TG and V6 have decorrelation times (time lag at which the autocorrelation function drop to 1/e) of about 1 day. The decorrelation time for V5 is almost twice as long (∼2 days). The spatial autocorrelations are shown in Fig. 6. The spatial autocorrelation function shows that the Thailand rainfall is nonisotropic. Such nonisotropy is more pronounced for the satellite products than TG. The rain systems are slightly elongated in the NE–SW direction, which is consistent with the rainbands during the monsoon season. The isotropic spatial decorrelation distances are 1.14°, 1.87°, and 3.61° for TG, V6, and V5, respectively.
Other rain event statistics such as event duration, event separation, and event-conditional rain rate are also considered. A rain event starts when the rain rate goes from zero to above 0 mm day−1 and ends when the rain stops. Rain-rate statistics are computed according to rainy and dry seasons. The winter and hot seasons in the north, northeast, central, and east are combined together as the dry season based on the amount of rainfall. Since the V6 performs better than the V5, rain event statistics (separation, duration, and conditional rain rate) for different regions are shown only for V6. Figure 7 shows mean event separations, durations, and conditional rain rates for the different regions computed for the rainy and dry seasons. There is little difference in the separation between TG and V6 in the rainy season, but the separation is longer for TG for the dry season, especially in the north, northeast, and central regions. On the other hand, the duration is nearly identical between TG and V6 in the dry season, but is about twice as long for V6 in the rainy season. Even though the V6 3B42 has more rain events, the conditional rain rate is lower. The results are not unexpected due to the design of the V6 algorithm, which is a gauge-merged analysis. The biases with gauge are expected to be minimal for V6. Hence if the event durations are longer, the conditional rain rates are almost guaranteed lower for V6.
Skill measures such as the probability of detection (POD), false alarm rate (FAR), and critical success index (CSI) are used to examine rain events, and the gauge measurements are used for verification (Schaefer 1990) as defined below:
where A is the number of hits (daily grid boxes that correctly estimate rain by the algorithm), B is the number of grid boxes that TG classifies as rain but the algorithm shows no rain, and C is the number of grid boxes that the algorithm incorrectly estimated to be rainy but rain did not occur. The POD considers the fraction of correctly estimated rain grids. The FAR considers the fraction of grids that are falsely estimated by the algorithm. For the perfect algorithm, POD = 1, FAR = 0, and CSI = 1. The POD, FAR, and CSI of the V5 and V6 algorithms for the whole year and for the different seasons are calculated and shown in Table 2. The results show that the POD is higher for V5 than for V6. However, the FAR is lower for V6. As a result, the CSI is about the same (0.62) for both V5 and V6. All statistics (POD, FAR, and CSI) are in general better for the rainy season [June–August (JJA) to September–November (SON)] than for the dry season [December–February (DJF) to March–May (MAM)]. Even though the biases for all the regions are substantially reduced for V6, there is little change in the CSI.
For many hydrological or agricultural applications, 5- or 10-day average rain rates are used. The correlations with TG at daily 1° grid scale are calculated to be 0.37 and 0.44 for V5 and V6, respectively. A series of correlation coefficients are also computed to investigate optimal averaging time (Fig. 8). The correlation increases drastically for averaging periods of 1 to 5 days and steadily increases beyond 5 days. The V6 is perceptibly better correlated with TG than V5.
Figure 9 shows the CSI as a function of the averaging period. Again, there is no difference in CSI for V5 and V6. The CSI increased drastically from 0.62 to 0.82 for averaging periods of 1 to 5 days. After that, the CSI gradually increased to about 0.94 when the average days are increased to 30 days. This suggests that the 5-day average for 1° × 1° would be appropriate for hydrological applications.
Uncertainty in the rain-rate threshold may affect the definition of a rain event. Figure 10 shows the sensitivity of the skill measures as a function of the rain-rate threshold. The POD and CSI decrease when the rain-rate threshold is increased, but the FAR increases. This problem is probably due to the lack of heavy rain in the 3B42 algorithms.
To further examine the relation between TG and TRMM daily rain rates, Fig. 11 shows the scatterplot of TG versus 3B42 V5 and V6 rain rates. The V5 (satellite measurements only) rain rates are mostly limited to below 100 mm day−1; in fact the rain rates mostly appear in 0–50 mm day−1 range. The results are consistent with decreases in the CSI when the rain-rate thresholds are increased.
4. Discussion and conclusions
In this study, satellite rainfall estimates are compared with point gauge measurements binned at 1°. Errors in the binned rain gauge measurement include gauge undercatch due to wind effects (Sevruk 1982) and inadequate spatial sampling within the grid box. In addition, the inaccuracy in spatial (gridded) interpolation is also significant. Willmott and Johnson (2005) estimated resolution errors associated with gridded monthly precipitation fields in the Amazon basin to be 11% for the summer and 15% for the winter season for 1° grid boxes.
An examination of the Thailand gauges shows two distinct seasons: dry and rainy. The rainy season starts in May and progresses southeast to northwest. The maximum rainfall, however, occurs in August for the north and northeast regions, in September for the central and east regions and in November for the south region. The retreat of rainfall follows the movement of the ITCZ and goes from north to south.
Analyses of monthly data yield very low bias, RMSD and MAD among TG, 3B43, and V6 3B42. Since these products use gauge analysis (GPCC) either as an input or a constraint, the results are consistent with the design of these algorithms. The V5 3B42, which involves no direct gauge input, is much higher than TG. Since the number of Thai gauges is ∼100 whereas that for GPCC is only ∼50, the TG provides better sampling than GPCC and hence is better suited for comparing daily statistics.
The histogram of the daily rain rates for V6 shows better agreement with TG than the V5 version. The scatterplots show that while the V5 data have a larger bias, it misses most of the heavy rain (>100 mm day−1). Since the V6 is constrained by the monthly 3B43 product, with a scaling factor between 0.2 and 2, it would be interesting to examine the contribution of the scaling factor to the existence of heavy rain in V6.
The temporal and spatial autocorrelation function of V6 is closer to TG than V5. This can be attributed to the use of microwave sensor estimates from multiple satellites in V6 while V5 data are based on calibrated geosynchronous IR data. The CSI, which is an index to compare algorithm rainfall detection, is basically unchanged between V5 and V6. Even though the FAR for the V6 is slightly decreased, the POD is also decreased. As the rain-rate threshold is increased, the CSI is decreased, which is again consistent with the lack of heavy rain rates in the satellite products.
We have computed rain-rate statistics that are useful for hydrologic applications. The analyses of the rain event statistics for V6 3B42 reveal that the rain event separation (∼5 days) is much shorter than the TG (∼10 days) in the dry season and the rain event duration (∼9 days) is much longer than TG (∼5 days) in the rainy season. Since the V6 is weakly constrained to match the monthly gauge rain rate, the V6 conditional rain rate is in general lower than TG. The results from the event statistics are consistent with the analysis of the histograms and scatterplots of TG and V6 rain rates.
Huffman et al. (2007) examined the biases associated with TMPA and a real-time product (3B42RT) with gauge analysis. They showed that TMPA has a lower bias than the 3B42RT compared to gauge analysis. Chiu et al. (2006c) showed that V6 3B42 and TCA are lower than the V5 3B42 and TCA by 20% and 13% over land, respectively. The 3B42RT has been evaluated by a number of investigators. Katsanos et al. (2004) evaluated the performance of the 3B42RT over central and eastern Mediterranean using one year (June 2002–May 2003) of 12-hourly data over 73 stations (June 2002–May 2003). They found that the detection statistics (POD and FAR) tends to be better in the rainy season than in the dry season. Similar results are also found in Ebert et al. (2007) in evaluating the operational satellite rain algorithms.
From these statistics, it can be concluded that although the TMPA algorithm improves significantly in the bias, the satellite-only algorithm still overestimates low rain rates and underestimates heavy rain rates. The next-generation satellite rainfall products should address these issues since extreme rain events are useful for natural hazard (flood and drought) monitoring.
In light of the proposed Pilot Evaluation of High Resolution Precipitation Products (PEHRPP; Arkin et al. 2005), we examined the performance of TMPA over Australia by compiling the skill scores available from the IPWG Web site (http://www.bom.gov.au/bmrc/SatRainVal/monthlysum.html). Preliminary results show that the TMPA, Climate Prediction Center (CPC) morphed precipitation (CMORPH), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN), and GPCP show higher skill in the rainy season than in the dry season, consistent with our findings. Furthermore, TMPA outperforms GPCP and PERSIANN and is comparable to CMORPH. A full assessment must await a comprehensive comparison of these operational algorithms. We plan to perform the comparison over Thailand in our future work.
We thank the Thailand Meteorological Department for supplying the rain gauge data. The TRMM data are processed by the TRMM Science Data and Information System (TSDIS) and distributed by the NASA GSFC Distributed Active Archive Center (GDAAC). LSC acknowledges support from NASA TRMM and REASoN programs. RC is supported by a GRA from GMU/CEOSR and ESGS.
Corresponding author address: Long S. Chiu, Department of Earth Systems and GeoInformation Sciences, MSN 6A2, College of Science, George Mason University, Fairfax, VA 22030-4444. Email: firstname.lastname@example.org