Performance of Two Long-Term Satellite-Based and GPCC 8.0 Precipitation Products for Drought Monitoring over the Yellow River Basin in China

Abstract

This study investigated the accuracy and drought monitoring application of two newly-released long-term satellite precipitation products (i.e., the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record, PERSIANN-CDR and the Climate Hazards Group Infrared Precipitation with Station data version 2.0 CHIRPS) and the latest reanalysis precipitation product (i.e., the Global Precipitation Climatology Centre full data monthly version 2018, GPCC 8.0). Satellite- and reanalysis-based precipitation sequences and standardized precipitation indices (SPIs) were compared comprehensively with background estimates of the China Gauge-based Daily Precipitation Analysis (CGDPA) dataset at spatial and multiple temporal scales over the Yellow River Basin (YRB) in China during 1983–2016. Results indicated the PERSIANN-CDR, CHIRPS and GPCC 8.0 precipitation products generally had good consistency with CGDPA (correlation coefficient, CC > 0.78). At spatial, monthly and seasonal scales, the consistency between GPCC 8.0 and CGDPA precipitation was found to be better than that of the two satellite products. Due to their good performance at the spatiotemporal scale, the satellite with long-time record and GPCC 8.0 products were evaluated and compared with CGDPA to derive SPI-1 (1-month SPI), SPI-3 (3-month SPI), and SPI-12 (12-month SPI) for drought monitoring in the YRB. The results showed that they had good application in monitoring droughts (CC > 0.65 at spatial scale, CC > 0.84 at temporal scale). The historical drought years (i.e., 1997, 1999, and 2006) and the spatial distribution of drought area in August 1997 were captured successfully, but the performance of GPCC 8.0 was found to be the best. Overall, GPCC 8.0 is considered best suited to complement precipitation datasets for long-term hydrometeorological research in the YRB.

1. Introduction

Drought can be the most frequent, longest duration, widespread, and serious natural disasters in the world [1]. Because of the spatiotemporal variability and complexity of drought [2], the general understanding of its evolutionary characteristics is limited and thus it is difficult to avoid serious drought-related losses to agriculture, ecology and society. To describe and monitor drought events quantitatively, some studies have used statistical methods to establish a common drought index according to the mechanism of drought occurrence [3]. For example, the Palmer Drought Index, which was proposed by Palmer in 1965, has become the semi-official drought index of the United States [4]. The Standardized Precipitation Index (SPI) established by McKee in 1993 is based on long-term precipitation records and has been used widely to analyze the characteristics of drought in various regions [5,6,7,8]. For monitoring meteorological drought, relevant indices include the relative humidity index [9] and the Standardized Precipitation Evapotranspiration Index [10]. Among these various indices, the SPI has the advantage of simple calculation and applicability on multiple timescales, and it can be applied to comparative analysis of droughts and floods in different areas [11,12,13]. Because of the above advantages, the SPI has been used widely in the study of meteorological drought characteristics under the recommendation of the World Meteorological Organization (WMO) [14]. Therefore, this study adopted the SPI as the meteorological drought monitoring index.

The initial cause of drought is inadequate precipitation over a certain period [15]. Therefore, precipitation has become a key meteorological variable to study drought and as input data to calculate drought indices (including the Palmer Drought Index, the Standardized Precipitation Evapotranspiration Index and the SPI) [1,16]. Because of the great differences in natural factors such as geographical location, topography, climate, meteorology, and the ecological environment, precipitation has considerable differences over various temporal and spatial scales [17]. Meteorological stations have inevitable limitations regarding monitoring of precipitation at large spatial scales [18]. Moreover, large uncertainty can be introduced when using spatial interpolation methods to derive precipitation data over a complex terrain [19]. To a certain extent, satellite-derived and reanalysis precipitation datasets can overcome the shortcomings of data obtained at meteorological stations, and they can represent powerful datasets for various hydrometeorological applications, especially in regions with minimal or no gauge-based data [20,21].

Since the 1980s, the rapid development of computer and advanced remote sensing technology has led to a series of precipitation products becoming available to the public. Short-term (below 30 years) satellite-derived precipitation products with high-resolution include the Tropical Rainfall Measuring Mission Multi-Satellite Precipitation Analysis products [22] and the Climate Prediction Center Morphing Technique estimates [20]. Long-term continuous historical records (over 30 years) include the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR) [23] and the Climate Hazards Group Infrared Precipitation with Station data version 2.0 (CHIRPS) [24], produced in accordance with the recommendations of the WMO [25]. These satellite precipitation products can be used for precipitation evaluation and drought monitoring. Conversely, meteorological station data have been used to produce surface reanalysis precipitation products such as the Global Precipitation Climatology Centre (GPCC) full data [26]. Such reanalysis precipitation estimates provide basic data for studying spatiotemporal variations of precipitation patterns and drought monitoring. However, both satellite-derived precipitation products and reanalysis precipitation estimates have their own limitations. The performance of satellite-derived precipitation products might be affected by factors such as location and height [27], and by the high uncertainty of satellite-based pre-encrypted datasets used for drought monitoring [15]. Precipitation estimates based on meteorological station data might also be affected by the estimation method adopted and the long update period, which cannot provide timely information for extreme drought events.

Many recent studies have examined the evaluation and application of satellite-derived precipitation products [18,25,28]. Some results have indicated that the discrete nature of meteorological observations, dry climate, complex terrain and few meteorological stations for the correction of satellite errors might affect the accuracy of satellite precipitation monitoring [1]. Consequently, some studies have proposed measures intended to address the uncertainties of satellite-derived precipitation through improving the algorithms used for satellite precipitation retrieval or increasing the numbers of meteorological stations used for correction [29]. In 2018, the full data monthly version of the GPCC gridded precipitation product (GPCC 8.0) incorporated data from more than 80,000 meteorological stations. However, further study is required to establish whether the GPCC 8.0 can provide useful precipitation estimates for the correction of bias in satellite-derived precipitation products. Therefore, intercomparison and practical application of satellite precipitation products and the GPCC 8.0 are urgently needed to develop a profound understanding of the error characteristics of different precipitation products, which could also provide a reference for future correction of bias in satellite precipitation products and hydrometeorological applications.

The Yellow River Basin (YRB) in China is a region in which floods have occurred frequently from historical times to the present, but it is also the major river basin in China affected most seriously by drought [30]. This study evaluated the applicability of the PERSIANN-CDR, CHIRPS and GPCC 8.0 precipitation datasets over the YRB in China, from the period 1983–2016, based on the China Gauge-based Daily Precipitation Analysis (CGDPA) product [31]. The applicability of the three precipitation products to drought monitoring in the region was analyzed using the SPI. Comparative evaluation of the performance of these precipitation products is important regarding hydrometeorological applications in relation to the studied basin and could also provide a reference for further general improvement of the applicability of the PERSIANN-CDR, CHIRPS and GPCC 8.0 datasets.

2. Study Region, Datasets and Methods

2.1. Study Area

The Yellow River, which is the second longest river in China, has a total length of 5464 km and flows through nine provinces. The study area in central China (32°9′–41°51′ N, 95°53′–119°14′ E) encompasses a watershed area of 752,000 km² (Figure 1). The elevation of the study area terrain is high in the west (6199 m) and low in the east (−9 m), and the average slope is about 2.4‰. There are considerable geomorphological differences between different districts. The basin has a temperate monsoon climate with abundant sunshine. Seasonally, it is wet and rainy in summer, and cold and dry in winter. The spatial distribution of precipitation in the area decreases from the southeast toward the west, and the average annual precipitation is 466 mm. The average annual temperature is between −4 °C and 14 °C [32]. Currently, the China Meteorological Data Network (http://data.cma.cn) constitutes approximately 405 meteorological stations that are used to form the CGDPA product in and near the boundary of the YRB. The spatial distribution of these stations is very dense but irregular, as shown in Figure 1, and the data acquired are also of importance temporally.

2.2. Datasets

2.2.1. CGDPA as Basic Precipitation Data

The 0.25° spatial resolution CGDPA dataset, which is a daily precipitation product based on data from more than 2400 discrete meteorological stations distributed across the country, was produced by the Meteorological Data Room of the National Meteorological Information Center of China [33]. Based on the climatological background field, the CGDPA interpolates daily precipitation using the optimal interpolation method in combination with topographic correction and strict quality control [31]. Digital elevation data are introduced to eliminate the influence of elevation on the accuracy of the spatial interpolation of precipitation under the unique topographic conditions in China. Therefore, the precipitation observations for the region have high quality and reliability. The monthly CGDPA dataset for the study period was calculated, and it is considered reliable and suitable as a basic precipitation dataset with which to assess the performance of satellite-derived precipitation and reanalysis precipitation products. For this study, precipitation observations of the daily CGDPA product of 1983–2016 were downloaded from the website of the Meteorological Data Center of the China Meteorological Administration. At present, the CGDPA product with 0.25° spatial resolution, provided by China Meteorological Network after April 2008, can be used directly. However, before that time, the spatial resolution of CGDPA product is 0.5°. In order to match the spatial resolution of satellite and reanalysis products, the CGDPA with 0.5° spatial resolution from 1983 to 2009 was resampled to obtain the CGDPA dataset with 0.25° spatial resolution through cubic function of ArcGIS software.

2.2.2. Satellite Precipitation Products

The PERSIANN-CDR product was developed by the organization, which is Center for Hydrometeorology and Remote Sensing of the University of California (USA), using an artificial neural network model in combination with an algorithm for the classification of meteorological elements [34]. It constitutes a daily precipitation data with 0.25° spatial resolution for the region 60° S–60° N over the quasi-global range from 1983 to the present [35]. It can be used for long-term (>30 years) precipitation assessment and drought monitoring and investigation of other extreme events attributable to climate change. Further details regarding this product are available on the Internet (http://chrs.web.uci.edu/persiann/). In this study, monthly PERSIANN-CDR precipitation data with 0.25° spatial resolution from 1983 to 2016 were used and compared with the CHIRPS and GPCC 8.0 datasets.

Currently, there are two versions of the CHIRPS product, which have characteristics of multiple timescales (day, month, season, year), multiple scales of spatial resolution, and wide area coverage (50° S–50° N). The latest CHIRPS products, which can be downloaded from the Internet (http://chg.geog.ucsb.edu/data/chirps/), have 0.05° × 0.05° spatial resolution, and the monthly precipitation products cover the period from 1981 to the present [36]. The CHIRPS product represents the combination of global climatology, in-situ measurements and high-resolution satellite estimation technology designed to develop and produce three different types of dataset that can be used for precipitation assessment, drought monitoring and early warning, as well as for the study of climate evolution and future trends. At the earliest stage, it was used for both future research on long-term trend analysis and support of the Famine Early Warning System Network of the United States Agency for International Development [8]. In this study, the CHIRPS v2.0 monthly precipitation data of 1983–2016 with 0.25° spatial resolution were evaluated and used for drought monitoring.

2.2.3. Reanalysis Precipitation Product

There are many different versions of the GPCC product, the latest of which is GPCC 8.0 [28]. The GPCC 8.0 product provides data with multiple spatial resolutions (i.e., 0.25° × 0.25°, 0.5° × 0.5°, 1.0° × 1.0° and 2.5° × 2.5°) for latitudes −90° S to 90° N and longitudes −180.0° E to 180.0° E from January 1981 to December 2016. Further details concerning this product can be found on the Internet (ftp://ftp.dwd.de/pub/data/gpcc/html/fulldata-monthly_v2018_doi_download.html). Based on rain gauge data from a network of approximately 80,000 meteorological stations around the world, which includes approximately 160 (i.e., less than in Figure 1) International Ground Meteorological Exchange Stations in China, monthly precipitation is estimated. Therefore, compared with the CGDPA product, GPCC 8.0 incorporates fewer meteorological stations in the YRB. The accuracy of GPCC 8.0 in this area remains uncertain and it needs to be evaluated and studied. In this study, GPCC 8.0 products with 0.25° spatial resolution from the period 1983–2016, were used. Because the precipitation product is estimated based entirely on meteorological station data and the renewal cycle is often long, it constituted only one of the evaluation objects used in this study.

2.2.4. Differences of the Inversion Process among the Precipitation Products

Preliminary flowcharts of the satellite (PERSIANN-CDR, CHIRPS), GPCC 8.0 and CGDPA precipitation estimation algorithms are shown in Figure 2. Their workflows are markedly different. The CGDPA product, which is based on 0.5° × 0.5° resolution GTOP30 data from 1961 over Mainland China, is combined with data (i.e., elevation, longitude and latitude) from more than 2400 meteorological stations to correct deviations [33]. Therefore, it has characteristics of a high-density observation network and high-precision precipitation data. To a certain extent, CGDPA is used as the benchmark dataset for comparison of satellite-derived precipitation products with gauge-based precipitation products in China. The error correction method of GPCC 8.0, which is based on quality control, is highly dependent on the gauge stations [26,37] and uncertainty in the product renewal cycle. The PERSIANN-CDR and CHIRPS products have similar precipitation estimation processes, which rely on satellite image data and merged meteorological station datasets to correct product deviation; however, the number of gauges used in each varies by approximately 40,000. Therefore, for the same region, the accuracy of the precipitation estimates of PERSIANN-CDR based on the Global Precipitation Climatology Project [38] and CHIRPS differs.

2.3. Methods

2.3.1. Evaluation Method

The research objectives of this study were to evaluate precisely the performance of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products based on the CGDPA product during 1983–2016 and to assess the validity of their application to drought monitoring. Therefore, this analysis is divided into two parts. The first part uses five statistical (Table 1) and classification measures to evaluate the product performance spatially and on multiple timescales. The correlation coefficient (CC) is used to assess the consistency between the CGDPA product and the PERSIANN-CDR, CHIRPS and GPCC 8.0 products. The root mean-squared error (RMSE) and mean error (ME) often can be used to describe the errors, and the relative bias (BIAS) is defined to estimate the deviations [39,40]. In the second part, the applicability of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products to drought monitoring is tested based on the SPI meteorological drought index.

2.3.2. Standardized Precipitation Index

The SPI, which is based on a calculation of precipitation data, is used to describe the multitemporal scale of drought. The gamma distribution probability is used to eliminate the spatiotemporal difference in precipitation. The SPI is sensitive to the change of drought, and it has the advantages of simple calculation and applicability on multiple timescales. It can be used to compare and analyze drought and flood in different regions, and it has better stability than other indices. Consequently, the WMO has recommended the SPI be used as an indicator of drought [41]. Different temporal scales reflect the soil water status (hourly timescale) or groundwater, and river and lake water levels (large timescales) [8]. In this study, periods of 1, 3 and 12 months were used as the timescales for grid computing SPI, namely SPI-1, SPI-3 and SPI-12, to reflect the spatiotemporal situation of drought. The SPI is obtained by calculating the cumulative gamma distribution probability of precipitation over a certain period and then standardizing the cumulative probability. Further details regarding the specific calculation process and the categorization of drought and flood grades can be found in the literature [8,41]. The approximate solutions of the SPI [8] are as follows:

(1)$SPI=\{\begin{array}{c}-(t-\frac{c_0+c_{1}t+c_2{t}^2}{1+d_{1}t+d_2{t}^2+d_3{t}^3}),t=\sqrt{\ln \frac{1}{H(x)}},0<H(x)\le 0.5\\ (t-\frac{c_0+c_{1}t+c_2{t}^2}{1+d_{1}t+d_2{t}^2+d_3{t}^3}),t=\sqrt{\ln \left\{\frac{1}{1-H(x)}\right\}},0.5<H(x)\le 1\end{array}$

where c₀ = 2.515517, c₁ = 0.802853, c₂ = 0.010328, d₁ = 1.432788, d₂ = 0.189269 and d₃ = 0.001308 and H(x) denotes the cumulative probability of precipitation x over a certain period of time.

3. Results

3.1. Evaluation and Comparison of the Accuracy of Satellite Precipitation and Reanalysis Precipitation

3.1.1. Evaluation at Spatial Scale

Comparison of the accuracy of the monthly precipitation data of PERSIANN-CDR, CHIRPS and GPCC 8.0 in relation to CGDPA (1983–2016) based on the spatial distribution of the CC, ME, BIAS and RMSE values is presented in Figure 3. The regional-averaged CC of PERSIANN-CDR, CHIRPS and GPCC 8.0 with CGDPA in the YRB are 0.926, 0.929 and 0.97 respectively; the regional-averaged BIAS are 0.016, −0.004 and 0.049 respectively; the regional-averaged RMSE are 17.06, 15.92 and 11.07 mm/month respectively; and the regional-averaged ME are 0.62, −0.64 and 2.04 mm/month respectively. As can be seen from Figure 3a–c, the CC values of GPCC 8.0 at most grids vary in the range 0.96–1. However, the CC values of CHIRPS at most grids vary in the range 0.9–0.96 and few grids vary in the range 0.96–1. The lowest range of CC values of 0.8–0.86 is the largest relative to the PERSIANN-CDR and GPCC 8.0 products, and it is distributed in the northwestern boundary area of the YRB, i.e., in relation to the Loess Belt in the north of the region. The CC values of PERSIANN-CDR are more than 0.86 but almost no grid has a high CC value (>0.96). It can be seen in Figure 3d–f that most values of the BIAS range from −0.4 to 0.4, and that the values in most areas are in the range −0.1 to 0.1. The RMSEs of the satellite-derived products in the region to the northwest of the Yellow River are in the range 10–20 mm/month, while for GPCC 8.0, the values across a comparatively large area are in the range 0–10 mm/month. The MEs are found mainly in the range –5 to 5 mm/month, and the performance of the CHIRPS product is found to be better than that of the others. In summary, the performance of the GPCC 8.0 precipitation product is found to be better than that of either CHIRPS or PERSIANN-CDR, although it is impossible to determine which satellite product performed best. This result is related to the conclusion that the discreteness of meteorological stations (Figure 1), complex terrain (Figure 1) and minimal correction of satellite data errors using meteorological station observations (Figure 2) might affect the accuracy of satellite-derived precipitation products [1].

A boxplot was adopted for further comparison of the performance of the PERSIANN-CDR, CHIRPS and GPCC 8.0 precipitation products because it represents a good aggregate calculation comprising the minimum, median, maximum, and other statistical values, as shown in Figure 4. It can be seen from Figure 4 that PERSIANN-CDR, CHIRPS (CC: 0.9–0.95, CC_CHIRPS > CC_PERSIANN-CDR) and GPCC 8.0 (CC > 0.95) all have good consistency with CGDPA. The BIAS of CHIRPS is close to zero, which is better than the other products (Figure 4b). It can be seen from Figure 4c that the RMSE of GPCC 8.0 is close to zero relative to both PERSIANN-CDR and CHIRPS. Moreover, it is evident that GPCC 8.0 has better performance than either PERSIANN-CDR or CHIRPS. In addition, the MEs also show that GPCC 8.0 has the best performance. Based on the analysis of the spatial distributions and boxplots of the statistical indicators, it can be concluded that the GPCC 8.0 precipitation product has strong spatial correlation with CGDPA, and that CHIRPS is more consistent than PERSIANN-CDR with CGDPA at the spatial scale. This result is related to Figure 2, and it indicates that the GPCC 8.0 product could have a positive correction effect on satellite-based precipitation retrievals.

3.1.2. Evaluation at Monthly Scale

For the three evaluated products, the results of monthly statistical indices are presented in Figure 5. They show that the level of consistency of GPCC 8.0 with CGDPA is higher than that of CHIPRS and PERSIANN-CDR. For example, Figure 5a indicates that the CC value of GPCC 8.0 (approximately 0.9) is higher than that of the other products over the 12 months. The CC values of the CHIRPS product fluctuate considerably from month to month with the lowest values of 0.54 and 0.76 in January and December, respectively. Conversely, the lowest CC values for PERSIANN-CDR are in June–September when precipitation is more frequent. The reasons are CHIRPS is insensitive to solid precipitation in winter and the regional distribution of the meteorological stations used for the inversion of the product is sparse. In terms of the BIAS, the trend of GPCC 8.0 is reasonably flat and close to zero, whereas that of both CHIRPS and PERSIANN-CDR changes considerably in the months with less precipitation. Overall, the GPCC 8.0 product has best performance in terms of RMSE and ME; however, in July and August, when precipitation is high, both metrics for all products deviate from their ideal values. For example, the RMSE of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products in July is 14, 15 and 18.5 mm/month, respectively. The ME of PERSIANN-CDR fluctuates greatly, with a negative value in August and a maximum value (ME > 7 mm/month) in October. The ME of CHIRPS is in the range 2–7 mm/month, which is higher (ME > 6 mm/month) in July–September. The ME of GPCC 8.0 is lower than PERSIANN-CDR and CHIRPS in most months.

3.1.3. Evaluation at Seasonal Scale

In this study, the values of the CC, BIAS, RMSE, and ME of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products were calculated for March–May as spring, June–August as summer, September–November as autumn, and December–February (next year) as winter (Figure 6). The CC values of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products in each season are in the range 0.839–0.924, 0.784–0.965 and 0.851–0.96, respectively, i.e., they all show reasonable consistency. The CC values of PERSIANN-CDR and CHIRPS are highest in spring and lowest in summer and winter, respectively, while the CC values of GPCC 8.0 are lowest in spring. For BIAS, the errors in winter are significantly larger than in the other seasons, except for PERSIANN-CDR in summer. The RMSE, which is less than 46 mm/season in each season in the YRB, is largest in summer and smallest in winter. The maximum ME values of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products are achieved in summer. Compared with the other products, the BIAS and RMSE of CHIRPS measured against CGDPA are smaller, except in winter. Generally, the GPCC 8.0 product performs best in all four seasons, followed by CHIRPS. This is because precipitation in the YRB in winter occurs mainly in the form of snow or solid precipitation, whereas in summer, it is characterized by short-duration heavy precipitation produced by severe convective weather conditions. In addition, the capability of an inversion of satellite-derived data to estimate both solid and short-term precipitation is limited; consequently, the CC values are low in winter and the deviation is large in summer.

3.1.4. Evaluation at Annual Scale

In terms of the annual average precipitation of the PERSIANN-CDR, CHIRPS, GPCC 8.0 and CGDPA products, we collected and analyzed annual average precipitation data from 1983 to 2016. The grid-based comparison for the 34-year period comprised 1285 grid points. The statistical indicators used in this part to reflect consistency at the annual scale are shown in Figure 7. It can be seen that the performance of CHIRPS is best (highest CC values), followed by GPCC 8.0. This finding is inconsistent with the results of the previous analyses at the monthly and seasonal scales, i.e., the RMSE of the three products is over 48 mm/year, and the standard deviation is over 130 mm/year. This is attributable to the large annual precipitation with cumulative effect and the relatively large proportion of the total error. In summary, the findings of Section 3.1 show that satellite and reanalysis products have better performance in estimating precipitation in the YRB at spatial and multi-temporal scales (CC > 0.78, RMSE < 20 mm/month in most of the region and at temporal scale). They have some deviations, which are still within acceptable accuracy. It indicates that the PERSIANN-CDR, CHIRPS and GPCC 8.0 products are all adequate in reflecting precipitation patterns and that they all present very serviceable precipitation data for drought monitoring.

3.2. Evaluation and Comparison of Drought Monitoring Based on Satellite Precipitation and Reanalysis Precipitation

3.2.1. Evaluation at Spatial Scale

The PERSIANN-CDR, CHIRPS and GPCC 8.0 products have high spatial resolution and relatively satisfactory correlation (CC > 0.8) in comparison with the CGDPA product in terms of the estimation of precipitation on multiple timescales (Section 3.1) over the YRB. Given the reasonable performance of these precipitation products, this section evaluates their applicability to drought monitoring based on the SPI. The SPI-1, SPI-3, and SPI-12 were calculated for the PERSIANN-CDR, CHIRPS, GPCC 8.0 versus CGDPA products over the 34-year period.

The SPI calculated by the satellite-derived and reanalysis precipitation products is in good agreement (regional-averaged CC > 0.75) with the SPI calculated by CGDPA. Specifically, on multiple time scales, The CC values of PERSIANN-CDR, CHIRPS and GPCC 8.0 in most areas of the YRB are between 0.65–0.85, 0.70–0.85 and 0.90–0.99, respectively (Figure 8) and they all exhibit the same characteristics, i.e., the CC value of SPI-3 is higher than that of both SPI-1 and SPI-12 across the entire basin. Satellite-derived precipitation has poor performance in relation to drought monitoring in the areas of high elevation of the western YRB, especially PERSIANN-CDR. Combined with the results of precipitation product evaluation (Section 3.1) and the differences of these precipitation products (Figure 2), results indicate that high altitude topography, minimal correction of satellite data errors using meteorological station observations, and spatial-temporal distribution of precipitation might affect the accuracy of satellite drought monitoring. Overall, the GPCC 8.0 product is considered the most suitable dataset for drought monitoring in the region.

3.2.2. Evaluation at Temporal Scale

To evaluate the accuracy of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products for drought simulation at the temporal scale, we compared their regional-averaged SPI-1, SPI-3 and SPI-12 over a long time series with those of the CGDPA product, and we compared their CCs and RMSEs calculated based on regional-averaged SPI over the entire YRB. It can be seen from Figure 9 that GPCC 8.0 and CGDPA have the best consistency at the three timescales (CC > 0.97, RMSE: 0.092–0.108), while the performance of CHIRPS is the worst (CC < 0.9, RMSE > 0.24). In addition, the simulation results of the three precipitation products are best for SPI-3. Previous studies have shown that SPI-1 does not consider the impact of precipitation in the earlier period, and it has strong randomness and weak temporal persistence. Conversely, SPI-3 has the best correlation with farmland areas and can be used for agricultural drought monitoring [8]. It is found that the satellite-derived and GPCC 8.0 precipitation can reflect the occurrence of historical drought years (1997, 1999 and 2006) in a changing environment. This finding is consistent with that based on the NMSDI, SPI and SSFI series spanning 1993–2012 in subbasin 31, as shown in Figure 2 of Xu et al. [3]. Although the performance of the satellite-derived and GPCC 8.0 precipitation products is very good at most time points, there are some errors that cannot be ignored. For example, Figure 9 shows that they often overestimate the extent of drought at specific times (e.g., 2003 and 2013) over the YRB, which is related to the spatial distribution of the SPI in the YRB. Generally, the GPCC 8.0 product has good consistency with the SPI of the CGDPA product at the various timescales, which indicates that GPCC 8.0 has good applicability for drought monitoring, followed by PERSIANN-CDR. This result is contrary to the assessment of the performance of the products in terms of spatial precipitation detected by satellites, as discussed in Section 3.1.1.

3.2.3. Application of the Three Precipitation Products to a Typical Drought Event

As can be seen from Figure 9, a drought occurred in 1997. Therefore, we discussed and analyzed that typical drought event in this section. First, the times of onset and cessation of the drought, its duration, and the minimum drought SPI value of this event were established (Table 2). From Table 2, it can be seen that GPCC 8.0 captures this drought event better than the satellite products at the 1- and 12-month scales, irrespective of the times of onset and cessation of the drought, its duration, and the drought SPI value. At all monthly scales, the time range and duration of the drought based on PERSIANN-CDR coincide with those of CGDPA, although there is considerable difference in the minimum SPI value. The characteristics of the CHIRPS product do not correspond to the CGDPA product, except at the 1-month scale.

To discuss the performance of the PERSIANN-CDR, CHIRPS and GPCC 8.0 products in relation to spatial drought monitoring, the spatial distribution of their SPI values over the YRB in August 1997, based on the periods shown in Table 2 and Figure 9, is shown in Figure 10. The values of four grids with 0.25° spatial resolution are the same as the value of the corresponding one grid with 0.5° spatial resolution, due to the cubic resampling for the CGDPA with 0.25° spatial resolution during 1983–2009. So that the pixels in CGDPA seem larger than in PERSIANN-CDR, CHIRPS, GPCC 8.0 in Figure 10. This has little effect on the results obtained from the evaluation and comparison of the satellite and reanalysis precipitation products. The results show that SPI-1, SPI-3 and SPI-12 all indicate drought in the southeastern part of the YRB and that the drought area accounts for about half the overall area. However, the CGDPA product indicates drought in the western part of the YRB, whereas PERSIANN-CDR and CHIRPS indicate only a few grids with drought, and GPCC 8.0 indicates only a slightly larger drought area. In addition, the SPI values of GPCC 8.0 and CGDPA are similar in terms of their spatial distribution, consistent with the results of Figure 8. Overall, the GPCC 8.0 product has best agreement in terms of the SPI of multiple timescales, followed by PERSIANN-CDR, indicating their applicability for drought monitoring.

4. Discussion

4.1. The Statistical Test of PERSIANN-CDR, CHIRPS, GPCC 8.0

Statistical indicators are commonly used to evaluate the accuracy of satellite and reanalysis precipitation products. However, there are data differences among the evaluated products, and it is worth discussing and studying whether the differences among PERSIANN-CDR, CHIRPS, and GPCC 8.0 are statistically significant or not. Figure 11 shows that the p value is obtained by Friedman test and any two products of them are significantly different by the multcompare function under the condition of the corresponding significance test. From Figure 11a, it can be seen that the data difference of the three products passes the 0.01 significance test in most areas of the YRB, and 0.05, 0.1 and 0.5 significance test in a small part of the region, but does not pass the 0.5 significance test in the central part of the YRB. The data difference of CHIRPS and GPCC 8.0 is very significant, covering most area of the YRB, while the data difference of PERSIANN-CDR and GPCC 8.0 is significant in a smaller area, and the area with significant difference between the two satellite products is the smallest, as shown in Figure 11b. This may be because the datasets of the satellite are monitored by remote sensing, while GPCC 8.0 obtains data through meteorological stations. In a word, the data difference between satellite and reanalysis precipitation products is significant, but the data difference between satellite products is not significant. This is consistent with the difference in the accuracy of these products monitoring precipitation and drought (Figure 3, Figure 4 and Figure 8).

4.2. The Error Analysis against the Background of Morphometric Parameters

The accuracy of the precipitation estimation by satellite and reanalysis precipitation products at time and space was analyzed through statistical indicators in the YRB of China with a complex underlying surface, which fully reflected the difference of the ability of satellite and reanalysis products to capture precipitation in spatial-temporal distribution. In addition, the elevation, slope and aspect of underlying surface may also affect the accuracy of these products [1,19]. In the future, the study area should be divided into multiple subregions, according to the elevation, main mountains and spatial-temporal distribution characteristics of precipitation. The error analysis of precipitation estimation for these products in the different subregions could be made by comparing the accuracy of different products in the same subregion and the same product in different subregions, which could fully reflect the impact of underlying surface elements on the accuracy of these precipitation products [1].

5. Conclusions

In this study, the accuracy and drought monitoring applicability of the long-term PERSIANN-CDR, CHIRPS and GPCC 8.0 precipitation products were evaluated comprehensively based on the CGDPA product benchmark data for 1983–2016. Five statistical and categorical metrics were used to evaluate and compare the PERSIANN-CDR, CHIRPS and GPCC 8.0 products with CGDPA gauge measurements from throughout the YRB in China. The consistency of GPCC 8.0 precipitation in most parts of the YRB was found slightly higher than that of the PERSIANN-CDR and CHIRPS products, especially in western parts of the YRB. However, all products showed good consistency in the southeastern region of the YRB (CC > 0.8), and they could all reflect the precipitation distribution on multiple timescales. Based on a multiple timescale SPI accuracy assessment, the performance of the satellite-derived precipitation products was found to be poor in the areas of high elevation of the western YRB, especially PERSIANN-CDR; however, they could reflect the characteristics of drought monitoring and distribution in the YRB. In terms of comprehensive precipitation and drought evaluation, the GPCC 8.0 product was found to be most suitable for application to water resource management and drought monitoring over the YRB. Although the CHIRPS product was found to be more suitable for detecting precipitation, its drought-monitoring performance was weaker than that of PERSIANN-CDR, which is related to the SPI.

In summary, in comparison with satellite-derived precipitation products, the GPCC 8.0 product incorporates data from a greater number of meteorological stations and can produce a more accurate spatial distribution of precipitation; thus, it represents a precipitation dataset suitable for drought assessment and is valuable for hydrometeorological research. Although the updating of the CHIRPS satellite product has been delayed, it is likely to be achieved before the GPCC 8.0 and PERSIANN-CDR products are updated [26]. Therefore, it is more likely that the CHIRPS product will be used in the near future to study the recent drought characteristics in the YRB. However, considering the accuracy of precipitation detection, the GPCC 8.0 product could be considered a valuable long-term precipitation dataset for the study of historical drought in the YRB. Therefore, we suggest that the GPCC 8.0 product be used as data for the correction of the inversion of satellite-derived remote sensing precipitation in the future.

Author Contributions

L.W., S.J., and L.R. designed the framework of this study; L.Z. collected the data; L.W. processed and analyzed the data and wrote the paper; F.Y. provided significant suggestions on the structure of the manuscript.

Funding

The current study was jointly supported by the National Key Research and Development Program approved by Ministry of Science and Technology, China (2016YFA0601504); the National Natural Science Foundation of China (51979069); the Program of Introducing Talents of Discipline to Universities by the Ministry of Education and the State Administration of Foreign Experts Affairs, China (B08048); the Postgraduate Research & Practice Innovation Program of Jiangsu Province (Grant No. 2019B72614/SJKY19_0477); the Fundamental Research Funds for the Central Universities (2019B10414).

Acknowledgments

We thank the anonymous reviewers for their valuable comments. In addition, we would like to express our appreciations to the editor of our paper, Lacey Cao, for his thoughtful comments and efforts in handling our paper. Finally, we are grateful to thanks for the producers of PERSIANN-CDR, CHIRPS, GPCC 8.0 and CGDPA.

Conflicts of Interest

The authors declare no conflict of interest.

Figures and Tables

Figure 1. Geographical location of the YRB and the distribution of the 405 meteorological stations.

Figure 2. Preliminary flowcharts of precipitation estimation algorithms of the PERSIANN-CDR, CHIRPS, GPCC 8.0, and CGDPA products.

Figure 3. The spatial distribution of CC, BIAS, RMSE (mm/month), ME (mm/month) of PERSIANN-CDR, CHIRPS, GPCC 8.0 versus CGDPA over the YRB.

Figure 4. Boxplot of the statistical indicators of PERSIANN-CDR, CHIRPS, GPCC 8.0 versus CGDPA: (a) correlation coefficient (CC); (b) relative bias (BIAS); (c) root mean-squared error (RMSE); (d) mean error (ME).

Figure 5. The statistical indicators of PERSIANN-CDR, CHIRPS, GPCC 8.0 versus CGDPA at the monthly scale: (a) correlation coefficient (CC); (b) relative bias (BIAS); (c) root mean-squared error (RMSE); (d) mean error (ME).

Figure 6. The CC, BIAS, RMSE (mm/season), and ME (mm/season) of PERSIANN-CDR, CHIRPS, GPCC 8.0 versus CGDPA at the seasonal scale.

Figure 7. Taylor diagram showing statistical comparison of the CGDPA with the estimates of annual mean precipitation during 1983–2016 of the PERSIANN-CDR, CHIRPS and GPCC 8.0.

Figure 8. The spatial distribution of the CC value of SPI-1, SPI-3, and SPI-12 based on PERSIANN-CDR, CHIRPS, GPCC 8.0 versus CGDPA over the YRB.

Figure 9. Time series of regional-averaged SPI-1, SPI-3, and SPI-12 based on PERSIANN-CDR, CHIRPS, and GPCC 8.0 versus CGDPA during 1983–2016.

Figure 10. The spatial distribution of SPI-1, SPI-3, and SPI-12 based on PERSIANN-CDR, CHIRPS, GPCC 8.0, and CGDPA over the YRB in August 1997.

Figure 11. The statistical test between PERSIANN-CDR, CHIRPS, GPCC 8.0 in the YRB: (a) the p value of the returned test by friedman function; (b) the significant difference of two products by multcompare function.

Table 1

Statistical indicators for satellite, reanalysis products and drought assessment.

Statistical Metrics	Meaning	Equation	Value Range	Perfect Value
CC	Correlation coefficient	$\mathrm{CC}=\frac{{\displaystyle \sum _{i=1}^n\left(C_i-\overline{C}\right)\left(S_i-\overline{S}\right)}}{\sqrt{{\displaystyle \sum _{i=1}^n{\left(C_i-\overline{C}\right)}^2}}\sqrt{{\displaystyle \sum _{i=1}^n{\left(S_i-\overline{S}\right)}^2}}}$	0 to 1	1
BIAS	Relative bias	$\mathrm{BIAS}=\frac{{\displaystyle \sum _{i=1}^n\left(S_i-C_i\right)}}{{\displaystyle \sum _{i=1}^n{C}_i}}$	−1 to 1	0
RMSE	Root mean-squared error	$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left(S_i-C_i\right)}^2}}$	0 to ∞	0
ME	Mean error	$ME=\frac{1}{n}{\displaystyle \sum _{i=1}^n\left(S_i-C_i\right)}$	−∞ to ∞	0
Std	Standard deviation	$\mathrm{Std}=\sqrt{\frac{1}{n-1}{\displaystyle \sum _{i=1}^n{\left(X_i-\overline{X}\right)}^2}}$	0 to ∞	0

Note: Ci means CGDPA precipitation or SPI; Si means PERSIANN-CDR, CHIRPS and GPCC 8.0 precipitation or SPI; n means the number of samples (34 × 12); $\overline{X}$ represents the mean of the sample Xi used.

Table 2

The characteristics of a typical drought event in the region.

SPI Temporal Scale (Month)	Product	Time Range (Year, Month)	Duration (Month)	Minimum of SPI
1-month	CGDPA	August 1997	1	−1.530
	PERSIANN-CDR	August 1997	1	−1.665
	CHIRPS	August 1997	1	−1.691
	GPCC 8.0	August 1997	1	−1.574
3-month	CGDPA	June 1997 to October 1997	5	−1.390
	PERSIANN-CDR	June 1997 to October 1997	5	−1.481
	CHIRPS	May 1997 to October 1997	6	−1.497
	GPCC 8.0	May 1997 to December 1997	8	−1.452
12-month	CGDPA	August 1997 to May 1998	10	−1.212
	PERSIANN-CDR	August 1997 to May 1998	10	−1.325
	CHIRPS	July 1997 to May 1998	11	−1.410
	GPCC 8.0	August 1997 to May 1998	10	−1.251