Development of APHRO _ JP , the first Japanese high-resolution daily precipitation product for more than 100 years

We constructed historical (1900–) high-resolution (0.05° × 0.05°) daily precipitation data over the Japanese land area as part of the product of the “Asian Precipitation – HighlyResolved Observational Data Integration Towards Evaluation of the Water Resources” (APHRODITE) project. This product APHRO_JP is derived from rain gauge observations and is intended to accurately represent both mean and extreme values. Due to new interpolation techniques developed in APHRODITE, estimation accuracy for orographic precipitation is improved, and bias for long-term amount is reduced, even for the early 20th century in which the observation network was sparse in space. Moreover, the product can be used for statistical analysis of heavy precipitation up to about 150 mm/day, over a long term period (≥ 100 years). APHRO_JP enables diverse research, including validation of meso-scale models and analysis of the longterm extreme precipitation trend in Japan.


INTRODUCTION
Long-term high-resolution daily precipitation data are useful for evaluating meso-scale models, driving river-flow models, and analyzing temporal and spatial variations of precipitation fields.Observational records longer than 50-100 years enable one to investigate precipitation changes caused by global warming.In addition, gridded data make model evaluation and water budget analysis easier, when compared with point observation data.
In Japan, the Radar-AMeDAS (R/A) developed by the Japan Meteorological Agency (JMA) is a well-known highresolution grid precipitation product.Since its observation period is not long enough for statistical studies, it has often been used in studies that do not require long-term data.For example, Mizuta et al. (2005) evaluated simulated extreme precipitation using the 20-km-mesh atmospheric general circulation model (AGCM) with 10-year R/A data, while Iida et al. (2006) investigated sampling errors caused by rainfall observations sparse in time and space for five lowearth-orbiting satellites including TRMM, using 3-year R/A composite data.R/A is based on radar observations and calibrated by the dense rain gauge observations termed "Automated Meteorological Data Acquisition System" (AMeDAS).As shown in Figure 1, more than 20 JMA-radar sites and 1300 AMeDAS rain gauge stations are deployed across the Japanese land area, which are used in R/A.R/A has been available since 1988 with 1-hour temporal and 5 × 5 km spatial resolutions, and was upgraded to 30-minutes temporal and 1 × 1 km spatial resolutions.Because R/A uses a combination of dense spatial radar and rain gauge observations, it is naturally expected to be of excellent quality.While the short-term quality may be good because the data are designed for disaster prevention activities, R/A does not guarantee long-term (monthly, annual) precipitation amount.Wakazuki et al. (2007) reported that R/A overestimated accumulated precipitation by 14% from 1995 to 2004, when compared with neighboring rain gauges.Sasaki and Kurihara (2008)  2006) of mean monthly precipitation (June and July) between R/A and rain gauges including AMeDAS around Kanto and Koshinetsu areas, and found that R/A underestimated precipitation by 6-7%.
As mentioned above, R/A is of insufficient observation period for trend analyses, and therefore, its long-term (monthly, annual) mean may also be of insufficient quantity.We therefore constructed historical (1900-) and highresolution (0.05° × 0.05°) daily precipitation product for the Japanese land area based on rain gauge observations with new interpolation techniques considering precipitation climatology and local topographical features.Owing to the interpolation techniques and elaborate quality control (QC) procedures, this product, APHRO_JP, enables many types of studies including extreme event analysis.
APHRO_JP is developed under the framework of the project "Asian Precipitation -Highly-Resolved Observational Data Integration Towards Evaluation of the Water Resources" (APHRODITE) (Yatagai et al., 2009).This paper describes briefly the production method and statistical characteristics of APHRO_JP.Information on the application limits of the product is also provided.
As a basic interpolation scheme, we used Spheremap (Willmott et al., 1985) which is categorized as an Angular-Distance-Weighting (ADW) method.The weighting function and influence radius were optimized to Japanese precipitation.In the calculation of the weighting function, Spheremap only considers horizontal locations of interpolated grid cells and neighboring rain gauges, which often leads to an overly smoothed horizontal distribution pattern.To improve this problem, we introduced a new scheme to alter the weight function calculated by Spheremap by using local topographical features, such as elevation and slope direction.For example, if a high crest exists between the interpolated grid cell and the rain gauge, the weight function is reduced.
In addition, to reduce estimation bias of long-term precipitation in mountainous areas in which rain gauge density is often sparse, Mountain Mapper (MM) (Schaake, 2004) was utilized.MM is a convenient method as it interpolates the ratio of the precipitation to its climatological mean, not the value of the precipitation itself.MM is also useful for improvement of the heavy precipitation frequency.In the original Spheremap, the interpolated value is limited to the range of input data, which leads to underestimation of heavy precipitation.However, MM eliminates this limitation for precipitation itself.The mesh climatology 2000 (MESHCLIM) produced by JMA is applied as the base climatology, whereas APHRO_PR_V0902 uses Worldclim (Hijmans et al., 2005).The MESHCLIM is 30-year (1971 to 2000) monthly climatology data with 1 × 1 km grid resolution.The data is produced by multiple regression with geographical features as explanation factors (Kitamura, 1990).The most suitable explanation factors are chosen automatically from tens of candidates (e.g.latitude, longitude, elevation, slope direction, distance from coastline) using the stepwise regression method.The daily climatology was calculated from monthly MESHCLIM data using Fourier interpolation.For the base climatology, spatial distribution is the most important factor and seasonal progress is enough even if its rough tendency is expressed.
We exclusively used rain gauge data observed by JMA in this version of APHRO_JP, as our main object is to develop a precipitation data with homogeneous quality over the long-term.The JMA data are available for a long period, and rain gauge location and instrumentation have not changed significantly.The following three kinds of data are used here.
The first is hourly data from JMA historical surface observations from 1900 to 1960, which is the same data used in Fujibe et al. (2005).This data was digitized manually from paper records by JMA since observation commenced.Up to 144 surface observatory stations are included in this data.The daily amount was calculated by summing hourly data from 01 JST to 00 JST of the next day.This is because the boundary time for daily averaging is different with different periods and with different stations (some daily data are bounded on 09 JST (00 UTC)).The second source data set is daily data from JMA surface observations from 1961 to 2004, and this is also comprised of surface observatory data from 158 stations in the maximum year.The stations are almost the same as for the first source, but data was digitized initially.The third and the last is that observed by AMeDAS, at more than 1300 stations.Suspicious data in these data sets were eliminated by the quality control (QC) system developed in APHRODITE.The first dataset is not quality-controlled data, while the second and the third are quality-controlled by JMA.
To investigate the effect of the number of rain gauges, we prepared the following two versions: APHRO_JP and APHRO_JP_60STN (60STN).The former is produced with all available rain gauge data, and the latter uses only 60 stations which have existed since 1900 (Figure 1).The historical changes in the number of available rain gauges is shown in Figure 2. Before 1977, less than 200 rain gauges were available but these increased greatly in number up to more than 1300 after 1977 due to installation of the AMeDAS observation network.

ERROR ESTIMATION AND STATISTICAL CHARACTERISTICS
Figure 2 shows mean annual precipitation and maximum annual daily precipitation from 1900 to 2008.The values are averaged over the entire Japanese land area.The R/A grid was converted to 0.05° × 0.05° before calculating statistical values.To investigate the effect of the MM method, the interpolated data without MM (NO-MM) are also shown.When APHRO_JP and 60STN are compared with each other, the annual precipitations do not show a large change around the year of 1978, when the number of input rain gauges was greatly increased.The bias which arises from the spatially sparse observations is corrected using MM.We can therefore conclude that APHRO_JP and MM do not contain spurious trends forged by the change in the number of rain gauges.On the other hand, NO-MM is less than APHRO_JP by about 1 mm/day before 1977, although the number of rain gauges is the same as APHRO_JP.This indicates that if the interpolation is performed with only 60 rain gauges, most of which are located in plains regions, large precipitation over highelevation can be underestimated without MM method.Unusual behavior exhibited in R/A.Until 2000, R/A is larger than APHRO_JP but it becomes abruptly smaller after 2001.In April 2001, the spatial resolution of R/A changed from 5 km to 2.5 km, and the analysis method was also changed.The algorithm change is considered to be the major cause of the reason for this temporal inconsistency.
If we compare APHRO_JP and 60STN in the annual maximum of daily precipitation after 1977, 60STN is smaller than APHRO_JP.This denotes that MM is insufficient as a bias correction method for such extreme analysis.The annual maximum of daily precipitation is a very rare event in time and space, and therefore its observed value is greatly influenced by spatial sampling density.If we make a trend analysis of the annual maximum daily precipitation using APHRO_JP, the generation of a false trend is possible.The ratio of annual maximum precipitation to the mean precipitation is shown in Figure 2. The trend of this ratio is larger in APHRO_JP than in 60STN (APHRO_JP: 0.057 yr −1 , 60STN: 0.036 yr −1 ).
Figure 4 shows distributions of annual precipitation (1989-2007) calculated from APHRO_JP, 60STN, NO-MM and R/A.When APHRO_JP and R/A are compared with each other, the detailed distribution patterns corresponding to geographical features are similar, except for northeast of Hokkaido.R/A can underestimate precipitation over this region because the radar site is far away (Figure 1).An anomalous grid pattern is found in R/A, which results from transformation from oblique Lambert to Cartesian projections.As we mentioned before, R/A contains quantitative discontinuity before/after 2001.In addition to this, R/A contains unusual distribution patterns which are not seen in APHRO_JP.These results indicate that APHRO_JP is of better quality than R/A, however we do not know the real precipitation features.
The precipitation pattern of NO-MM is too smooth whereas the number of input rain gauges is the same as that in APHRO_JP.While 60STN shows good agreement with APHRO_JP, orographic precipitation is captured very well by a limited number of rain gauges due to the introduction of the MM method.This result confirms that APHRO_JP's spatial distribution of annual precipitation is of sufficiently good quality, even in the earlier years of the 20th century in which the surface observation network was immature.
We also have checked statistical characteristics of APHRO_JP other than the means, as APHRO_JP is intended for extreme event analyses.Figure 5 shows the probability distribution function (PDF) of daily precipitation in JJA (June-July-August).The analysis was conducted for the period between 1989 and 2007 in which R/A is available.Frequency is calculated for every 1 mm/day bin.In addition, two analysis periods (1989-2000 and 2001-2007) were set for R/A, as R/A changed its characteristics after April 2001.The frequency of heavy precipitation of R/A reduces after 2001.APHRO_JP is in good agreement with R/A (2001)(2002)(2003)(2004)(2005)(2006)(2007).60STN is less than APHRO_JP for heavy precipitation, but similar to APHRO_JP and R/A (2001)(2002)(2003)(2004)(2005)(2006)(2007) for precipitation intensities of the order of 100 to 150 mm/day.This means that APHRO_JP can be used for statistical heavy precipitation analysis up to this value, even for the early 20th century.Meanwhile, frequency of weaker precipitation is similar to each other for all products.

CONCLUSION
We have developed APHRO_JP, a new historical highresolution daily precipitation product for the Japanese land area.This is the first grid product that can be used to investigate daily precipitation over a time scale longer than 100 years.Owing to state-of-the-art interpolation techniques developed in the APHRODITE project, the quality of mean precipitation is considered to be high even for the early 20th century, and is consistent through the period despite an abrupt change in the number of available rain gauges during the period.For statistical extremes analysis, APHRO_JP can be used to determine heavy precipitation up to approximately 150 mm/day.However, it should be noted that extreme analysis is influenced strongly by the number of rain gauges which have changed during the period.The user can estimate this effect in their analysis by using 60STN, which was produced from continuous rain gauge observations since 1900.
Some improvements to APHRO_JP are planned for the next version.The major one is to introduce an adjustment scheme known as "wind-induced under-catchment".When snowing, wind prevents rain gauges from capturing solid precipitation (snowflakes), which causes underestimation of winter precipitation.This effect seems very large especially for Japan due to the existence of large areas with heavy snow.Adam and Lettenmaier (2003) produced a global dataset of adjustment ratios that can be used for correcting this underestimation.In their result, annual global precipitation increases about 11.7%.Utsumi et al. (2008) developed a gauge-based analysis of daily precipitation over Japan with wind-induced undercatch correction.Then they showed that annual precipitation over Japan is approximately 10% larger than commonly thought.
We have already made a preliminary adjustment scheme Figure 1.Location of rain gauge and JMA-Radar sites (blue square).Black dots are AMeDAS rain gauges deployed after 1977, whereas red dots show JMA surface observatory rain gauge which have existed since 1900.

Figure 2 .
Figure 2. Historical changes in the number of available rain gauges (bottom), annual mean and maximum daily precipitation (middle), and the ratio of annual maximum daily precipitation to annual mean daily precipitation (top).

Figure 3 .
Figure 3. Mean monthly precipitation over Japanese land area (mm/day).

Figure 5 .
Figure 5. Probability distribution function of daily precipitation.