Evaluation of Three Long-Term Remotely Sensed Precipitation Estimates for Meteorological Drought Monitoring over China

Abstract

Remotely sensed precipitation estimates (RSPEs) play an essential role in monitoring drought, especially in ungauged or sparsely gauged areas. In this study, we evaluated the ability of three popular long-term RSPEs (PERSIANN, CHIRPS, and MSWEP) in capturing the meteorological drought variations over the 10 first-level water resource basins of China, based on the standardized precipitation index (SPI). Drought events were identified by run theory, and the drought characteristics (i.e., duration, severity, and intensity) were also evaluated and compared with a gridded in situ observational precipitation dataset (CMA). The results showed that the three RSPEs could generally capture the spatial patterns and trends of the CMA and showed better performance in the wetter basins. MSWEP had the best performance for the categorical skill of POD, followed by CHIRPS and PERSIANN for the four timescales. SPI6 was the optimal timescale for identifying meteorological drought events. There were large skill divergences in the 10 first-level basins for capturing the drought characteristics. CHIRPS can efficiently reproduce the spatial distribution of drought characteristics, with similar metrics of MDS, MDI, and MDP, followed by MSWEP and PERSIANN. Overall, no single product always outperformed the other products in capturing drought characteristics, underscoring the necessity of multiproduct ensemble applications. Our study’s findings may provide useful information for drought monitoring in areas with complex terrain and sparse rain-gauge networks.

1. Introduction

Drought is a stochastic climate phenomenon triggered by persistent precipitation deficits [1,2]. The potential threats of drought to society, the economy, and the environment are devastating due to its close association with water resource shortages [3], food insecurity [4], and ecosystem degradation [5]. Hence, monitoring the evolution of drought and quantifying its characteristics (i.e., duration, severity, and intensity) [6,7] are crucially important for sustainable development and water resource management.

Droughts are generally classified into four types [8]: meteorological, hydrological, agricultural, and socioeconomic droughts, according to the specific sectors affected and the physical processes involved. To quantify the characteristics of the four drought types, substantial efforts have been devoted to developing drought analysis and monitoring techniques, of which the development of an applicable drought index is one of the most widespread approaches. Several drought indices have been derived to characterize the duration, severity, and intensity of droughts [9]. Among these indices, several typical indices—such as the Palmer drought severity index (PDSI) [10], standardized precipitation–evapotranspiration index (SPEI) [11], and standardized precipitation index (SPI) [12]—are commonly used to represent drought evolutions. The PDSI is calculated by a physical water balance model [10], taking into account the effects of factors other than precipitation (P) on drought evolution. Numerous studies have reported that the PDSI can be used as a good indicator for characterizing long-term droughts [13,14]. However, the PDSI also has some shortcomings, including a fixed temporal scale, high sensitivity to soil information, and autoregressive characteristics [15]. The SPEI overcomes the shortcomings of PDSI with a fixed timescale, and it depicts drought characteristics at flexible timescales by normalizing the difference between monthly potential evapotranspiration (PET) and P. As the main input for the SPEI, the PET has different algorithms, and the choice of PET algorithms can considerably affect the drought evolution [16,17]. Compared with the PDSI and SPEI, the SPI has more wide applications because of the advantages of simple calculation, flexible timescales, and low data requirements (i.e., only precipitation) [12,18]. Therefore, the SPI has been recommended by the World Meteorological Organization (WMO) as the preferred meteorological drought index [19]. The SPI has been widely used for drought monitoring and forecasting around the world, such as in China [20], the USA [21], Turkey [22], Malaysia [23], Ethiopia [24], and Spain [25].

Reliable precipitation records are crucial for calculating drought indices [26]. However, long-term gauge-based precipitation observations are often unavailable in developing countries or rural areas, limiting the application of drought indices in these areas [27,28]. Fortunately, the remotely sensed precipitation estimates (RSPEs)—such as Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks—Climate Data Record (PERSIANN-CDR) [29], Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) [30], and Multi-Source Weighted-Ensemble Precipitation (MSWEP) [31]—can provide long-term spatiotemporal continuous precipitation inputs for the calculation of drought indices, overcoming the shortcoming of insufficient ground precipitation observations in ungauged or sparsely gauged areas. However, existing RSPEs often have varying degrees of uncertainty due to the uncertainties of the precipitation retrieval algorithms and the complexities of the precipitation occurrence mechanisms [32,33]. It is necessary to carry out reliability assessments of different RSPEs in drought monitoring [34,35].

Numerous studies have evaluated the performance of RSPEs in capturing meteorological drought on a regional or global scale [36,37,38]. For example, using gauge-based observations as a reference, Zhang [37] reported that CHIRPS could reasonably depict drought characteristics at different timescales in the Hai River Basin of China. Guo [39] also found the outstanding performance of CHIRPS in drought monitoring in the Mekong Basin. Zhang [40] and Brito [41] evaluated the drought monitoring capabilities of PERSIANN-CDR and CHIRPS in two basins of East Africa and Brazil, respectively, and the results showed that CHIRPS performed better in the basin of East Africa, while PERSIANN-CDR performed better in the basin of Brazil. MSWEP, compared with the reanalysis product, performed better in reproducing the drought characteristics in a basin of Colombia [42]. In summary, most previous studies have adopted a regional focus or used a single RSPE product for drought evaluation, which may result in doubt about the generalizability of their findings. Therefore, it is urgent to evaluate the capability of different RSPEs in drought monitoring under a wide range of climatic conditions.

This study evaluated the capability of three popular long-term RSPEs in drought monitoring over China using a gauge-based gridded precipitation dataset (CMA) as the reference. The SPI and run theory were used to characterize the drought characteristics (i.e., duration, severity, and intensity). The objectives of this study were as follows: (1) to investigate spatiotemporal changes in drought in first-level water resource basins located in various climatic regions from 1983 to 2019; (2) to evaluate the performance of three RSPEs (PERSIANN-CDR, CHIRPS, and MSWEP) in capturing drought characteristics at both gird and basin scales; and (3) to illustrate the potential ability of RSPEs in depicting the drought characteristics of specific drought events. The novelties of this paper mainly include (1) the use of grid-based precipitation data interpolated from ∼2400 in situ gauged observations, (2) the selection of three typical long-term (>30 years) RSPEs, and (3) a study area covering different climatic regions. The results of this study are expected to provide an important reference for selecting reasonable RSPEs for drought monitoring over different climatic regions or similar basins.

2. Study Area and Precipitation Datasets

2.1. Study Area

China is covered by complex and diverse geography, with elevations spanning from −127 to 8848 m above sea level (Figure 1a). China spans multiple climatic zones, from hyper-arid to humid (Figure 1b), with an aridity index (AI, defined as the ratio of P to PET) ranging from 0.016 to 3.7. Diverse climates and complex underlying surface conditions make China an ideal place to evaluate the performance of RSPEs in drought monitoring. To further investigate the regional differences among RSPEs in capturing drought, 10 first-level water resource zones were selected over China, and their spatial distribution is shown in Figure 1a. Five climate zones were classified based on the AI [43]: hyper-arid (AI < 0.05), arid (0.05 ≤ AI < 0.2), semi-arid (0.2 ≤ AI < 0.5), dry semi-humid (0.5 ≤ AI < 0.65), and humid (AI ≥ 0.65).

2.2. Precipitation Products

2.2.1. Grid-Based Observational Precipitation

The observation-based daily precipitation dataset released by the China Meteorological Administration (CMA) was used as the reference dataset to evaluate the performance of three RSPEs in drought monitoring. The CMA has a spatial resolution of 0.25°and covers the period 1983–2019. It was generated based on ∼2400 in situ gauged observations (Figure 1a) and a climatology-based optimal interpolation algorithm [44]. This algorithm considers the effect of topography on precipitation interpolation and removes the high-frequency noise of daily precipitation. These ∼2400 gauged data were strictly quality-controlled, including outlier identification, internal consistency checks, and spatiotemporal consistency checks [44]. The CMA dataset has been widely used as a benchmark to assess the reliability of other precipitation products [45,46]. These daily precipitation data were obtained from the National Meteorological Information Center (http://data.cma.cn/) (accessed on 1 October 2022), and they were accumulated into monthly precipitation.

2.2.2. Long-Term Remotely Sensed Precipitation Estimates

PERSIANN-CDR (PERSIANN for short hereafter) was released by the National Climatic Data Center (NCDC) [29]. It provides long-term (>30 years), high-spatiotemporal-resolution precipitation estimates for hydrological and climate research [27,33]. This product uses an archive of gridded satellite (GridSatB1) infrared radiation (IR) data to force the artificial neural network model. To reduce bias in precipitation estimates, PERSIANN was corrected by the Global Precipitation Climatology Project (GPCP) monthly at 2.5°, and it was then downscaled to 0.25°. Monthly PERSIANN has a spatial coverage of from 60°S to 60°N, and a temporal span from 1 January 1983 to the present.

CHIRPS (version v2.0) is provided by the Climate Hazards Group, and it is explicitly designed for agricultural drought monitoring and global change investigation [30]. This dataset was processed in three steps: (i) generating a global 0.05° monthly precipitation climatology (GHPclim), (ii) obtaining CHIRPS by blending various datasets into GHPclim, and (iii) releasing the CHIRPS by incorporating station observations from five public data streams and several private archives into CHIRPS. CHIRPS is a quasi-global (50°S–50°N), high-resolution (0.05°), long-term (1981 to present) precipitation dataset with multiple timescales (from 6-hourly to 3-monthly aggregates).

MSWEP (version 2.0) is one of the recently released RSPEs, and it is the first remote sensing precipitation product with full global coverage [31]. The product merges numerous state-of-the-art data sources—such as in situ observation-, satellite-, analysis-, and reanalysis-based datasets—to greatly improve the accuracy of precipitation estimations. A feature of MSWEP is that it corrects the biases in precipitation estimates using a Budyko-based framework and global coverage runoff observations. In addition, MSWEP also considers the wind-induced precipitation undercatch, which is critical for improving the accuracy of solid precipitation estimates. For a fair comparison, all RSPEs were focused on the monthly timescale (1983–2019). CHIRPS and MSWEP were resampled to 0.25° using ArcGIS 10.8. More detailed information about these three RSPEs is listed in

Table 1. Basic information of precipitation datasets.

Name	Temporal Range	Temporal Resolution	Spatial Coverage	Spatial Resolution	Data Sources	Access Website
CMA	196001-Now	Daily	China	0.25	in-situ	http://data.cma.cn/ (accessed on 1 October 2022)
PERSIANN-CDR	198301-Now	Daily	60°N~60°S	0.25	G, S	https://www.ncei.noaa.gov/data/precipitation-persiann/ (accessed on 1 October 2022)
CHIRPS v2.0	198101-Now	Daily	50°N~50°S	0.05	G, S, R, A	https://data.chc.ucsb.edu/products/CHIRPS-2.0/ (accessed on 1 October 2022)
MSWEP v2.0	197901-Now	3 h	Global	0.1	G, S, R, A	http://www.gloh2o.org/mswep/ (accessed on 1 October 2022)

Note: the letters in the table represent the following: G-ground observation, S-satellite retrieval, R-reanalysis data, A-analysis data.

.

3. Methodology

3.1. Methodological Framework

To fully evaluate the drought monitoring performance of the three long-term RSPEs over the largest river basins of China, four major parts are included in this paper. The schematic flowchart of the methodological framework is shown in Figure 2, and the detailed method is described below.

3.2. Performance Metrics

Three basic performance metrics were selected for evaluating the drought monitoring of the three RSPEs, including absolute bias (aBIAS, −∞ < aBIAS < +∞), relative bias (rBIAS, −∞ < rBIAS < +∞), and Kling–Gupta efficiency (KGE, −∞ < KGE ≤ 1). Their equations are given as follows:

$$aBIAS=\frac{{\displaystyle \sum _{i=1}^N{R}_i-G_i}}{N}$$

(1)

$$rBIAS=\frac{{\displaystyle \sum _{i=1}^N{R}_i-G_i}}{{\displaystyle \sum _{i=1}^N{G}_i}}\times 100\%$$

(2)

$$KGE=1-\sqrt{{(1-r)}^2+{(1-\frac{{\sigma }_R}{{\sigma }_G})}^2+{(1-\frac{{\mu }_R}{{\mu }_G})}^2}$$

(3)

where R_i and G_i represent the monthly remotely sensed data and ground-based gridded data for the ith month, respectively; ${\sigma }_R$ and ${\mu }_R$ represent the standard deviation and mean of the remotely sensed values, respectively; ${\sigma }_R$ and ${\mu }_R$ are the standard deviation and mean of the ground-based data, respectively.

In addition to the basic performance metrics, one categorical metric—the probability of detection (POD, 0 ≤ POD ≤ 100)—was also used to evaluate the performance of the monthly RSPEs in capturing the drought categories. The POD represents the ratio of the number of drought months detected by the RSPEs to the actual number of drought months detected by the CMA. This metric can be calculated as follows:

$$POD=\frac{N(R_i=1\&G_i=1)}{N(G_i=1)}\times 100$$

(4)

where R_i = 1 (R_i = 0) and G_i = 1 (G_i = 0) denote the drought events in the ith month correctly (or not) detected by the RSPEs and ground-based precipitation, respectively.

3.3. SPI Definition and Classification

The SPI is one of the most robust drought indices for depicting meteorological drought. It has been widely recommended by the WMO due to its computational flexibility, requiring only precipitation as its input [12]—for example, 1 month for meteorological drought monitoring (SPI1), 3 months and 6 months for agricultural drought applications (SPI3 and SPI6), and 12 months (SPI12) for monitoring the basin’s water resources [27]. These four typical timescales (SPI1, SPI3, SPI6, and SPI12) were selected to evaluate the performance of the RSPEs in monitoring the drought evolution from short to long term. The SPI values span from −3 to +3, and the larger positive SPI values represent wetter conditions, while smaller negative ones indicate dryer conditions. According to the SPI values, drought can be divided into seven categories, spanning from extreme drought to extreme wet. Specific details are listed in

Table 2. Categories of drought calculated by the SPI.

Drought Classification	SPI Value
Extreme wet	(2, +∞)
Very wet	(1.5, 2.0]
Moderate wet	(1.0, 1.5]
Near normal	(−1, 1)
Moderate drought	(−1.5, −1.0]
Severe drought	(−2.0, −1.5]
Extreme drought	(−∞, −2]

.

3.4. Run Theory and Drought Events’ Characteristics

The run theory is used to identify drought events [47,48], and the detailed identification process by the run theory is shown in Figure 3. Generally, three typical drought events were identified in this schematic diagram. Then, the drought event characteristics estimated from the three RSPEs were compared with those of the CMA estimates. The duration (DD), severity (DS), intensity (DI), and peak (DP) of the drought event characteristics are depicted by the mean drought duration (MDD), mean drought severity (MDS), mean drought intensity (MDI), and mean drought peak (MDP), respectively. These expressions of these metrics are given as follows:

Figure 3. Illustration of drought characteristics based on the SPI categories and run theory.

Figure 3. Illustration of drought characteristics based on the SPI categories and run theory.

$$MDD={\displaystyle \sum _{i=1}^{n}D{D}_i}$$

(5)

$$MDS=\frac{{\displaystyle \sum _{i=1}^{n}D{S}_i}}{N},DS={\displaystyle \sum _{j=1}^{DD}SP{I}_j}$$

(6)

$$MDI=\frac{{\displaystyle \sum _{i=1}^{N}D{I}_i}}{N},DI=\frac{DS}{DD}$$

(7)

$$MDP=\frac{{\displaystyle \sum _{i=1}^{N}D{P}_i}}{N},DP=\underset{1\le j\le DD}{max(SP{I}_{\mathrm{j}}})$$

(8)

where i represents the index of the ith drought event, N is the total number of drought events for the study period, and j is the index of the drought month in a specific drought event.

4. Results

4.1. Basic Statistical Skill of the RSPEs against the CMA

4.1.1. Spatial Distribution and Trend of RSPEs

Figure 4 and Figure 5 present the spatial patterns and the statistical information of annual precipitation in the 10 first-level river basins. All three long-term RSPEs could generally capture the spatial pattern of the annual mean CMA—that is, the pattern of annual precipitation decreasing from the southeast to the northwest (Figure 4a–d and Figure 5a). Across mainland China, PERSIANN (632.9 mm/y) had a closer multiyear mean value to the CMA data (629.6 mm/y), while CHIRPS and MSWEP underestimated the mean annual precipitation. However, there were significant differences for each first-level basin. For example, all products overestimated the annual precipitation in the northeastern areas (Basin ID 1–3) and the Yangtze River, along with the southwestern rivers, but underestimated it in the northwestern rivers (Figure 5a). CHIRPS and MSWEP showed better performance in reproducing the spatial distribution of the CMA data than PERSIANN, and such outperformance regions were mainly located in the northwest.

For the spatial pattern of annual precipitation trends, 73.3% of mainland China showed an increasing trend, of which 39.5% was statistically significant (p < 0.05). The significantly increased areas were mainly located in the northwestern rivers (Figure 4e). The RSPEs had difficulties in reasonably capturing the spatial trends of monthly precipitation. For example, compared with the CMA data (1.08 mm/yr²), higher trend values of PERSIANN were found in the western areas of the northwestern rivers, the upper and downstream areas of the Yangtze River, and the southern regions of the Pearl River (Figure 4f). Those divergences of PERSIANN caused the annual trend to be overestimated by 10.2%. However, the precipitation trend was overestimated by CHIRPS by 106.5%, particularly in the southeastern rivers and the Pearl River (Figure 5b). For MSWEP, the annual precipitation trend was largely underestimated in the Liao, Hai, Yellow, Huai, Yangtze, and southeastern rivers but overestimated in the Pearl River and most western areas. Such divergence in the precipitation trends of MSWEP resulted in a slight underestimation of 5.5% (Figure 4h and Figure 5b).

4.1.2. Basic Performance of RSPEs

Figure 6 shows the performance ability of the three RSPEs at the grid and basic scales. Generally, the RSPEs showed better performance in eastern China and poor performance in the northwest regions (Figure 6). PERSIANN generally outperformed CHIRPS and MSWEP, and MSWEP showed large spatial inconsistencies in the northwestern rivers. Specifically, PERSIANN had lower rBIAS (−0.05%) and generally outperformed CHIRPS and MSWEP over mainland China. The large discrepancies of rBIAS were mainly located in the northwestern rivers (Figure 5a–c,g), underestimated by approximately 27.2%. The rBIAS in eastern China had good performance, with most grids being less than 10%. For the comprehensive metric of KGE, PERSIANN and CHIRPS showed similar performance (0.62) and were slightly better than MSWEP (0.60) over mainland China. While MSWEP presented large spatial heterogeneity in the northwestern area (Figure 6f), it outperformed PERSIANN and CHIRPS for most eastern basins (Figure 6h).

4.2. Performance of RSPEs in Capturing Spatiotemporal Variations in the SPI

To examine the performance of the three RSPEs in capturing the SPI, the correlation coefficient (r) and absolute bias (aBIAS) of the SPI estimated from the three RSPEs and the CMA data were evaluated. Figure 7 and Figure 8 present the spatial distribution of r and aBIAS in four timescales. Generally, MSWEP performed the best, followed by PERSIANN, while CHIRPS was poor in capturing the CMA’s SPI variability. At the national scale, the three RSPEs had reasonable r scores, with average values above 0.69. Relatively better performance was found in eastern China, while poorer performances were mainly located in western regions (Figure 7). Taking SPI3 of MSWEP as an example, it could excellently capture the variation in the CMA data in eastern basins (Basins 1–8; Figure 7 and Figure 8), with minimum, maximum, and mean values of 0.83, 0.92, and 0.86, respectively. However, the r sharply reduced to approximately 0.65 in the western basins (Basin 9–10), indicating that this poses a great challenge for RSPEs in accurately capturing the CMA data’s SPI fluctuations in the complex terrain and arid climatic regions.

With the increase in the SPI timescales, PERSIANN and CHIRPS improved their r values on a national scale, while MSWEP showed a decreasing performance. For example, r increased from 0.69 to 0.71 and from 0.67 to 0.72 for PERSIANN and CHIRPS, respectively. The improved r in PERSIANN mainly benefited from its higher performance in eastern China, as the r value in western regions became worse as the timescale increased. In contrast, the improved r in CHIRPS may have benefited from the increasing r in the northwestern rivers. However, r in MSWEP became poor with the increase in the timescale, decreasing from 0.74 for SPI1 to 0.72 for SPI12, which was mainly caused by the decreasing SPI in the basins of central–eastern China, such as the Liao River, Hai River, Yellow River, and Huai River (Basins 2–5). It should be noted that the reason for evaluating aBIAS instead of rBIAS was mainly because rBIAS is too sensitive to small values, which can result in inconsistencies in spatial distribution patterns. In Figure 8, aBIAS in both PERSIANN and CHIRPS becomes better as the timescale increases. Such similar rBIAS scores in these two products were mainly derived from the fact that both of them used the infrared retrieval of precipitation as their main input data source [29,30]. SPI1 was significantly overestimated (p < 0.001) by these two products in the northwestern rivers, resulting in underestimated drought severity. Thus, it is not recommended to use SPI1 in this region.

Generally, SPI12 was the best timescale in capturing the CMA’s SPI variations for the three RSPEs. In contrast, it was strongly recommended to apply PERSIANN and CHIRPS over the eastern basins of China. CHIRPS generally outperformed PERSIANN over most basins (Figure 8), and this may be due to CHIRPS assimilating more remote sensing data sources as well as ground observations [30].

4.3. Performance of RSPEs in Depicting the Drought Characteristics

4.3.1. Categorical Performance of RSPEs in Detecting Drought Months

Figure 9 presents the spatial distribution of POD for the three RSPEs in four timescales over mainland China. From this spatial pattern of POD, three conclusions can be drawn: First of all, for spatial performance, all of the RSPEs performed best in the relatively humid basins, including in the Huai, Yangtze, southeastern, and Pearl rivers, followed by the climatic transition basins, such as the Songhua, Liao, Hai, and Yellow rivers. The poor values of POD were located in the western arid basins. This indicates that further improvements in drought monitoring skills over the arid areas are needed, and much more caution should be taken in selecting precipitation products in relatively dry regions. Second, for the performance of the RSPEs, MSWEP generally performed best over mainland China, with a mean POD value of 59.1 for all timescales, followed by CHIRPS (POD = 53.9), while PERSIANN performed poorly (POD = 52.8). Significant regional performance divergences were detected over the 10 basins between the three RSPEs. For example, MSWEP showed a reasonable performance over most basins except for the northwestern and southwestern rivers. CHIRPS performed well in the Hai, Huai, and southeastern rivers, but PERSIANN performed poorly over all basins except the Songhua River Basin. Finally, for the performance at different timescales, the POD of all RSPEs generally increased with the increase in the timescale, and the best timescale for POD was SPI6. Therefore, SPI6 was selected for further analysis of drought event characteristics.

4.3.2. Performance of RSPEs in Capturing Drought Characteristics

Four aspects of drought characteristics—including the drought duration, severity, intensity, and peak—were used to evaluate the performance of the RSPEs in monitoring drought events. These four characteristics were quantified by the metrics MDD, MDS, MDI, and MDP, respectively. Figure 10 and Figure 11 exhibit the spatial distribution and statistical information of these four metrics in SPI6, respectively, between 1983 and 2019.

The MDD over mainland China was 4.24 months, and the smaller values were mainly located in the northeastern and central plains of China (Basins 1–5), while the higher values were found in southeastern and western China (Basins 6–10; Figure 10a and Figure 11a). The three RSPEs generally overestimated the MDD, especially for PERSIANN, with an 8.7% larger MDD than the CMA data (Figure 10b). Although CHIRPS and MSWEP could reasonably capture the mean value of MDD, they also exhibited evident spatial divergences. For example, many overestimated areas were found in the northwest regions (Figure 10). At the basin scale, the largest MDD values of PERSIANN and CHIRPS were found in the northwestern and southwestern rivers, respectively (p < 0.001). The best performances of PERSIANN, CHIRPS, and MSWEP were found in the southwestern, Hai, and southeastern rivers, respectively (Figure 11a).

As the drought severity (DS) is accumulated from the SPI6 of all drought months, a longer drought duration (DD) can cause greater severity. Therefore, the MDS of each precipitation dataset shared similar spatial distribution and basin statistical values with their corresponding MDD (Figure 10e–h and Figure 11b). For example, PERSIANN overestimated drought severity by approximately 13.4% over mainland China, with an especially obvious overestimation in the northwestern rivers (24.8%) (Figure 10f and Figure 11b). CHIRPS and MSWEP could reasonably capture the mean MDS of the CMA data, but they generally overestimated the drought severity. CHIRPS performed well in the Hai, Huai, and Pearl rivers (Basins 3,5, and 8, respectively; Figure 10g and Figure 11b).

In terms of the mean drought intensity (MDI), CHIRPS outperformed the other two products in reproducing the spatial pattern. Specifically, CHIRPS could perfectly reproduce the patterns of high drought intensity in the semi-humid and semi-arid climatic transition areas (Figure 1b and Figure 10i,k). On the other hand, both PERSIANN and MSWEP had difficulties in reasonably capturing the spatial distribution of drought intensity over the CMA data. At the basin scale, PERSIANN performed best in the Huai and Pearl rivers (Basins 5 and 8, respectively), CHIRPS in the Hai and southeastern rivers (Basins 3 and 7, respectively), and MSWEP in the southwestern rivers (Basin 9; Figure 11c). Therefore, the optimal product in each basin can be recommended for the detection and early warning of drought intensity.

Mean drought peak (MDP) can reflect the extremity of drought events. It plays a crucial role in crop growth and water resource management, because whether the crops can survive or the residents’ drinking water is available largely depends on the drought’s extremity. The MDP over China was −1.79, which falls into the severe drought category, indicating China that experienced severe, extreme droughts during 1983–2019 (Figure 10m). PERSIANN overestimated the drought peak over most parts of China (MDP = −1.87), while both CHIRPS and MSWEP were able to adequately capture the mean CMA value. At the regional scale, PERSIANN performed best in the Songhua, Liao, and Pearl rivers (Basins 1, 2, and 8, respectively), CHIRPS in the southeastern and northwestern rivers (Basins 7 and 10, respectively), and MSWEP in the Pearl River (Basin 8). Above all, there was significant divergence in capturing the spatial distribution of drought characteristics estimated from SPI6. Generally, CHIRPS had the best performance, followed by MSWEP, while PERSIANN was poor due to overestimating the drought events for most areas.

5. Discussion

5.1. Potential Reasons for Influencing RSPEs’ Performance

RSPEs have great advantages in capturing the spatial distribution and temporal variations of drought events [34]. However, the performance of the three selected RSPEs was poor in western China, which has dry climatic conditions (Figure 6, Figure 7 and Figure 8). Some inherent factors may contribute to this poor performance, such as (1) the data source or algorithm used for retrieving precipitation records, (2) terrain complexity, and (3) station density. Here, their roles in determining the monitoring of drought events are quantitatively analyzed.

For the first factor—the data source or algorithm—PERSIANN is mainly derived from the GridSat-B1 infrared (IR) data based on the PERSIANN model. Although PERSIANN performed well in many basins, it always presented significant bias in monitoring drought characteristics. This was mainly due to the mechanism by which PERSIANN evaluates the precipitation, by building a statistical relationship between the infrared brightness temperature of geostationary satellites and rainfall intensity. This mechanism may be not correct, because the actual precipitation process occurs in the lower part of the cloud body, while the radiative temperature at the top of the cloud and the precipitation intensity under the cloud do not have a simple statistical relationship [49]. Therefore, the precipitation estimates retrieved from infrared–visible images alone have great uncertainty. In addition, the parameters of PERSIANN were only adequately optimized in the United States, without training the ANN data in China [50]. The bias correction was conducted using the coarse spatial resolution of GPCP (2.5°), without station data, which may also undermine the performance of PERSIANN [29]. CHIRPS was initially supported for agricultural drought monitoring, and it also estimates precipitation based on infrared cold duration observations. However, it generally outperforms PERSIANN in monitoring drought events (Figure 8 and Figure 9). This is mainly because CHIRPS incorporates more accurate satellite information (such as TMPA 3B42 v7), more in situ observation datasets (such as the best available monthly and pentannual station data), and a novel blending algorithm (i.e., an improved inverse distance weighting method) [30]. MSWEP is regarded as part of the new generation of precipitation datasets, with the characteristics of high spatiotemporal resolution and incorporated with a range of the state-of-the-art gauge, satellite, and reanalysis estimates. Thus, it has excellent performance and has been widely reported in previous studies [32,38]. MSWEP was skillful enough in reproducing the spatial patterns and monthly trends of the CMA data (Figure 4), as well as the four temporal scales of the SPI (Figure 7) and the categorical metrics (Figure 10). Outstanding performance was found in the relatively wet areas of China. However, it still needs to be noted that MSWEP also exhibited unsatisfactory performance in terms of absolute bias (Figure 8) and the four drought characteristic metrics (Figure 10). Numerous mosaics with unexpected poor performance were found in northwestern and northeastern China. These unsatisfactory mosaics may be derived from the heterogeneous weight coefficient of MSWEP, when it merges numerous data sources. Multiple sources may significantly disturb the continuity of each incorporated dataset [31]. Thus, the algorithm of MSWEP needs to be paid more attention for improving drought monitoring in those regions.

The negative effect of terrain complexity on precipitation estimates has been widely discussed in previous research [32,33,34,35]. Here, we analyzed the influence of terrain complexity on RSPEs. The terrain complexity is defined as the difference between the maximum and minimum elevation within the 0.25° grid. Figure 12a shows the correlation coefficients (r) between the three RSPEs and the CMA data as the terrain complexity increases. The results reveal that the r of MSWEP shows a slight upward trend from 0.8 to 0.88, when the terrain complexity is below 1000. However, r drops rapidly after 1000 (p < 0.001) and, finally, falls below 0.7. CHIRPS shares a similar variation to MSWEP. Significantly different from MSWEP and CHIRPS, the r of PERSIANN consistently declines as the terrain complexity increases. PERSIANN generally outperforms the other two products under different terrain conditions. Our findings solidly confirm that complex topography can pose a great challenge for accurately estimating precipitation using remote sensing technology.

The third reason is the density of precipitation observational stations. In China, most stations are constructed in low-altitude plains and developed cities (Figure 1a). Observational precipitation from the high-density stations can potentially provide sufficient reference data for bias correction for the RSPEs. In contrast, the station density is always low in mountain or remote areas, such as in western China. It is difficult to optimize the parameters of RSPEs for those regions with limited observational precipitation data. Here, to reveal the reason for this, we calculate the station density of China and present the r scores of the three RSPEs as the station density changes (Figure 12b). The r values of all RSPEs remain relatively constant at an approximate value of 0.9 below 10,000 km²/station. Then, the r values in PERSIANN, CHIRPS, and MSWEP decrease significantly (p < 0.01)—from 0.9 to 0.8, 0.7, and 0.65, respectively—when the station density increases from 10,000 to 30,000 km²/station. Finally, the r values of the three estimates remain stable when the station density is greater than 30,000 km²/station. Our findings indicate the importance of building additional in situ observational stations in sparse and remote areas, as well as the crucial urgency of improving the algorithms for precipitation estimates.

5.2. Uncertainties and Prospects

Strict quality control was implemented for objectively evaluating the performance of the three RSPEs, but some inherent limitations still exist. Firstly, the in-situ stations in the northwest and Tibetan Plateau areas are sparse (Figure 1), which can weaken the confidence of the CMA data, posing some uncertainties for their use as a reference benchmark for evaluating the precipitation estimates. Secondly, gamma distribution, as proposed by McKee [12], was used for fitting the monthly cumulative precipitation in calculating the SPI. Many researchers have reported some other theoretical distributions-such as the Pearson type III, Weibull-type, and log-normal distribution-as being optimal for specific regions, rather than the gamma distribution [51]. Thus, further investigations are needed for future works.

Regardless of the few abovementioned limitations of our study, our preliminary findings present a valuable reference for selecting the optimal RSPEs for basin drought monitoring in different climate regions. Based on the results of this paper, we intend to merge various RSPEs into one outstanding estimate of China, by using state-of-the-art technologies such as ensemble model output statistics [52], Bayesian models [46], and machine learning [53]. In addition, meteorological drought is considered the first stage of drought’s propagation to hydrological drought (streamflow deficit), agricultural drought (soil moisture deficit), and socioeconomic drought (social water supply deficit) [8]. Therefore, how meteorological drought evolves into the other three types across different climate zones requires comprehensive analysis. Finally, more evaluation work over mountainous or complex orography should be conducted to further clarify the limitations of RSPEs in drought monitoring applications over those areas, which always contain the water sources of China and even the whole of Asia, such as the Sanjiangyuan region in the Tibetan Plateau [54].

6. Conclusions

The performance of three long-term remotely sensed precipitation estimates (RSPEs; PERSIANN, CHIRPS, and MSWEP) was comprehensively evaluated in monitoring meteorological drought over mainland China, based on gridded in situ observational precipitation data and the run theory.

Our main conclusions are as follows: (1) All of the RSPEs can reasonably capture the annual spatial distribution and trends of the CMA data over China, with better performance in the humid areas and poor performance over the dry areas. (2) With the increase in the SPI timescales, the performance of PERSIANN and CHIRPS became better in reproducing the SPI’s variation in the CMA data, but MSWEP became worse. Generally, the best performance of r in MSWEP and aBIAS in CHIRPS was detected over most basins. (3) In terms of categorical performance, MSWEP showed superior performance to PERSIANN and CHIRPS in four timescales over China. SPI6 was the best timescale for all RSPEs in identifying meteorological drought. For the four drought characteristics estimated from SPI6, CHIRPS had the best performance, followed by MSWEP and PERSIANN. Our findings confirmed that these three long-term RSPEs can be efficiently used in drought monitoring over a large basin scale with different climatic conditions. However, they should be applied with caution to specific regions, due to the impact of the product retrieval algorithms, terrain complexity, and station density on drought-detecting performance. Therefore, our findings can provide a reference for choosing optimal precipitation estimates for drought monitoring over specific climatic regions.