Evaluating the Applicability of PERSIANN-CDR Products in Drought Monitoring: A Case Study of Long-Term Droughts over Huaihe River Basin, China

Abstract

In this study, Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks–Climate Data Record (PERSIANN-CDR) was evaluated for the assessment of long-term drought monitoring in Huaihe River Basin using daily gauge observation data for the period from 1983 to 2017. The evaluation results show that the PERSIANN-CDR algorithm has a good detection ability for small precipitation events over the whole basin, but a poor ability for extreme precipitation events (>50 mm/day). Daily PERSIANN-CDR estimates perform relatively better in areas with abundant precipitation, while the monthly and yearly PERSIANN-CDR estimates are highly consistent with gauge observations both in magnitude and space. The Standardized Precipitation Index (SPI) at various time scales (3, 6, and 12 months) was calculated based on PERSIANN-CDR and gauge observation, respectively. Grid-based values of statistics derived from those SPI values demonstrate that PERSIANN-CDR has a good ability to capture drought events of each time scale across the basin. However, caution should be applied when using PERSIANN-CDR estimates for basin-scale drought trend analysis. Furthermore, three drought events with long duration and large extent were selected to test the applicability of PERSIANN-CDR in drought monitoring. The results show that it has a good ability to capture when and where droughts occur and how far they spread. Due to the overestimation of small precipitation events, PERSIANN-CDR tends to overestimate the number of extreme droughts and their extents. This needs to be considered in future algorithm improvement.

1. Introduction

With the impact of climate warming, drought has become one of the most destructive, frequent, and wide-ranging natural disasters in China [1,2]. In addition to water shortage areas, areas with abundant rainfall have experienced sustained drought. Since 2009, the southwest region has experienced four consecutive years of low precipitation and significant drought, causing widespread governmental and public concern. At the beginning of 2022, due to the persistent low precipitation in the Pearl River basin, the most severe drought in 60 years swept through the southern cities of China, causing water outages in Guangzhou, a city with annual precipitation levels of 2000 mm. Therefore, drought has always been a research hotspot in China. Wang et al. [3] used monthly meteorological observation data at 633 sites in China during 1961–2012 and found that drying usually occurred in central China and southwestern China in spring and autumn. Shao et al. [4] confirmed this conclusion and further found that the trends in drought duration, temporally averaged severity, and frequency also became more severe in each region during the past 36 years. Some other studies also revealed the intensification of the severity of regional droughts [5,6,7]. Consequently, improving drought monitoring and forecasting skills is a top priority for China, and this is inseparable from reliable precipitation data and high spatial-temporal accuracy.

Recently, we have witnessed significant achievements in satellite-based precipitation products (SPPs). Various SPPs with continuous temporal availability and global coverage have been produced and applied for various research [8,9,10]. Compared with rain gauge and radar observations, SPPs have the advantage of wide coverage and strong timeliness [11,12,13,14]. They have become reliable alternative sources for hydrological and climate studies [15,16,17,18,19]. However, it is worth noting that the coverage time of most popular SPP data-sets is less than 30 years, such as Tropical Rainfall Measuring Mission (TRMM) Multi-Satellite Precipitation Analysis (TMPA, Austin, TX, USA, from 1998) [20]; Global Precipitation Measurement (GPM, from 2014) [21,22]; the National Oceanic and Atmospheric Administration (NOAA, Washington, DC, USA) Climate Prediction Center (CPC, College Park, MD, USA) morphing technique (CMORPH, Boulder, CO, USA, from 2002) [23,24]; and the Global Satellite Mapping of Precipitation Microwave-IR Combined Product (GSMaP, Osaka, Japan, from 2000) [25,26]. According to a World Meteorological Organization (WMO, Geneva, Switzerland) report, climate studies require at least 30 years of historical data. These SPPs are limited in supporting trends and risk analysis of climate characteristics such as drought and extreme precipitation [27].

The PERSIANN-CDR, developed from the Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks (PERSIANN, Irvine, CA, USA) system, provides long-term historical precipitation estimates from 1983 to the present. It makes it possible to investigate whether the SPPs can accurately capture the regional climate information. Since its release, PERSIANN-CDR has been widely investigated in different parts of the world for its performance in estimating rainfall [28,29,30,31,32,33,34,35,36,37] and has been well received by most scholars. However, it cannot be directly used for specific purposes before testing it for suitability. This is because its appropriateness for specific applications varies from place to place and over time due to many factors [38,39,40]. Guo et al. [41] investigated the ability of PERSIANN-CDR to capture the spatial and temporal patterns of drought events over China and found that the 6-month Standard Precipitation Index (SPI) of PERSIANN-CDR shows the best agreement with ground-based data sets in identifying drought months in eastern China. They also found that the performance of PERSIANN-CDR is influenced by the terrain and the satellite algorithm. Majid et al. [42] evaluated three satellite precipitation products using multiple evaluation metrics and three different drought indices over Iran for the period 1983–2017. The results showed that TRMM and PERSIANN-CDR outperformed CHIRPS for drought event detection when compared to ground-based observations. They performed relatively better in arid and semi-arid climates compared to other climate zones. Francisco et al. [43] also compared three satellite precipitation products, concluding that PERSIANN-CDR showed the lowest fit with gauge observations with high underestimation. However, the SPI series of PERSIANN-CDR presented similar results to the other two SPPs. Wei et al. [44] found that the long-term PERSIANN-CDR performed satisfactorily in drought detection, but caution should be applied when studying the spatial variation of drought using PERSIANN-CDR. These studies demonstrate both the potential applicability and limitations (e.g., terrain, climate conditions, satellite algorithm, etc.), of PERSIANN-CDR in drought monitoring, suggesting the need to evaluate the representativeness and applicability of PERSIANN-CDR precipitation products for specific purposes before actual use [40]. Through literature review, we further found that most research efforts usually focused on testing the ability of PERSIANN-CDR to identify drought intensity and frequency over a certain period. Few involve the satellite’s ability to capture the start–development–end of some specific drought events in time and space, which are more critical for drought warning and prevention. In this study, we use PERSIANN-CDR to investigate the space and time behavior of specific drought events in the Huaihe River Basin (HRB) to gain a more detailed understanding of the performance of satellite products.

To identify and monitor drought, various drought indexes such as Palmer Drought Severity Index (PDSI) [45], Standard Precipitation Index (SPI) [46], Effective Drought Index (EDI) [47], and Surface Water Supply Index (SWSI) [48] have been developed over time at different spatial-temporal scales. Among these, the SPI has been widely applied because of its simplicity and ability to be computed at almost any time scale via a simple calculation procedure to assess drought duration, intensity, frequency, and extent [44,49,50,51]. In contrast to other indices, it is solely based on long-term precipitation data, which would avoid the introduction of impacts of other data sets when only evaluating the usability of satellite precipitation products for drought monitoring [52]. Furthermore, this research work involves the drought development of the Huaihe River Basin, which is located in the transition zone of the north–south climate of China, and the climate within the basin is quite different. Because of its standardization, SPI is easily adapted to local climates, which makes the calculated drought characteristics comparable across climatic zones [28,50,53]. These characteristics make the SPI suitable as a drought index in the study area.

As an important grain producing area, the total grain production in the Huaihe River Basin accounts for 20% of the total grain production in China. In recent years, several long-term droughts swept through parts of the Huaihe River basin, causing large losses to industrial and agricultural production [54]. Drought has become one of the main factors restricting the rapid and sustainable development of the regional social economy. This study aimed to yield more detailed insights on PERSIANN-CDR’s reliability for long-term meteorological drought monitoring during 1983–2017 and to try to answer the following questions: 1. Can the PERSIANN-CDR estimates reliably capture the start–development–end of long-term severe droughts? 2. What are the potential factors influencing the performance of PERSIANN-CDR in drought monitoring? The rest of this paper is organized as follows: Section 2 introduces the study area, the PERSIANN-CDR product, the dataset analysis, and the evaluation metrics. Section 3 presents the precipitation and drought conditions comparison between PERSIANN-CDR and gauge observations. The discussion and conclusions are, respectively, presented in Section 4 and Section 5. Figure 1 shows the methodological flow of this study.

2. Materials and Methods

2.1. Study Area

The Huaihe River originates in the Tongbai Mountain area in Henan Province, with a total length of about 1000 km and annual water resources of 79.4 billion cubic meters. It runs from west to east, with the mainstream entering the Yangtze River in Yangzhou, Jiangsu Province. The basin lies between the Yangtze River Basin and the Yellow River Basin, with complex terrain and a drainage area of 270,000 km². In the west, southwest, and northeast of the basin, there are mountainous and hilly areas, accounting for about one–third of the total area; the vast plains account for the remaining two-thirds of the total area (Figure 2). As an essential source of food production in China, the HRB covers most of the regions of four large agricultural provinces: Shandong (SD), Jiangsu (JS), Anhui (AH), and Henan (HN). It is the transition zone of China’s north–south climate, with a regional average annual temperature of 14 °C and precipitation of 806 mm. Due to the monsoon climate, precipitation in the flood season (June to September) accounts for 70% of the annual total, and it often falls in the form of a Rainstorm within a short time. Affected by topography and climate, flood and drought disasters frequently occur, which has severely impacted the basin’s social economy.

2.2. Data Analysis

To better evaluate the performance of the PERSIANN-CDR products, we collected the precipitation observation data of 192 ground observation stations evenly distributed in the HRB from 1983 to 2017 from the Climate Center of Anhui Province. Data quality issues, such as sensor and measurement errors, were addressed beforehand by the technicians of the climate center. However, the issue of missing values in the datasets of some sites should be taken seriously. Firstly, we counted the data recorded for each site for each year from 1 January 1983 to 31 December 2017. If missing data occurred within one month consecutively at the same site, or if the number of missing value days exceeded 40 days in a year, the data for that site were discarded. Thus, we obtained 168 sites with high-quality data as seen in Figure 2. Then, we interpolated those missing values with data from nearby sites using the bilinear interpolation method.

We downloaded the precipitation product of PERSIANN-CDR from the website of NOAA in the United States (https://www.ncei.noaa.gov/data/, accessed on 2 August 2020). PERSIANN-CDR [29] is developed for the National Climatic Data Center (NCDC) Climate Data Record (CDR) program in the National Oceanic and Atmospheric Administration (NOAA). Currently, PERSIANN-CDR provides 0.25° precipitation estimates at daily temporal resolution with near-global coverage from 60°S to 60°N from 1 January 1983 to 31 December 2021. When the grid-scale satellite products are directly compared with the point-scale gauge data, scale differences between different rainfall data sets will likely contribute to the evaluation errors. To reduce scale errors, we interpolated the gridded PERSIANN-CDR estimates into the locations of the 168 gauges through the bilinear interpolation method. By doing so, the two sets of precipitation values were made comparable.

2.3. Statistical Evaluation and Metrics

Six evaluation statistics were adopted to compare the performance of PERSIANN-CDR products against rain-gauge observations quantitatively: Pearson linear correlation coefficient (CC), Spearman correlation coefficient ($\mathsf{\rho }$), relative bias (Bias), root mean squared error (RMSE), probability of detection (POD) and false alarm ratio (FAR). The CC reflects the degree of linear correlation between satellite precipitation and gauge observations, or SPI series derived from these two datasets. $\mathsf{\rho }$ also reflects the degree of correlation between two datasets, defined as the Pearson correlation coefficient between the rank variables [55]. The rank transformation of data can overcome the limitation of the Pearson correlation when the variables are non-Gaussian [56]. The Bias describes the systematic bias of satellite precipitation estimates. The RMSE measures the average error magnitude. The POD reflects the ability of satellite precipitation to capture actual precipitation events accurately, and the optimum value of the indicator is 1. The FAR reflects false alarms of precipitation events from satellite precipitation data, with an optimum value of 0. It answers the question, “What fraction of the predicted ‘yes’ events actually did not occur?” [57]. These statistics are defined as follows:

$$\mathrm{CC}=\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{G}}_{\mathrm{i}}-\overline{\mathrm{G}}\right)\left({\mathrm{S}}_{\mathrm{i}}-\overline{\mathrm{S}}\right)}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{G}}_{\mathrm{i}}-\overline{\mathrm{G}}\right)}^2}·\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{S}}_{\mathrm{i}}-\overline{\mathrm{S}}\right)}^{2}1}}$$

(1)

$$\mathsf{\rho }\left({\mathrm{rg}}_{\mathrm{G}},{\mathrm{rg}}_{\mathrm{S}}\right)=\frac{\mathrm{cov}\left({\mathrm{rg}}_{\mathrm{G}},{\mathrm{rg}}_{\mathrm{S}}\right)}{{\mathsf{\sigma }}_{{\mathrm{rg}}_{\mathrm{G}}}{\mathsf{\sigma }}_{{\mathrm{rg}}_{\mathrm{S}}}}$$

(2)

$$\mathrm{Bias}=\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{S}}_{\mathrm{i}}-{\mathrm{G}}_{\mathrm{i}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{G}}_{\mathrm{i}}}$$

(3)

$$\mathrm{RMSE}=\sqrt{\frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{S}}_{\mathrm{i}}-{\mathrm{G}}_{\mathrm{i}}\right)}^2}$$

(4)

$$\mathrm{POD}=\frac{\mathrm{H}}{\mathrm{H}+\mathrm{M}}$$

(5)

$$\mathrm{FAR}=\frac{\mathrm{F}}{\mathrm{H}+\mathrm{F}}$$

(6)

where n is the number of samples; S_i represents the satellite precipitation estimates or SPI values derived from the satellite precipitation estimates; G_i represents the gauge observations or SPI values derived from the gauge observations; $\overline{\mathrm{S}}$ is the mean satellite-estimated precipitation or mean satellite-based SPI values; $\overline{\mathrm{G}}$ is the mean gauge precipitation or mean gauge-based SPI values; ${\mathrm{rg}}_{\mathrm{G}}$ and ${\mathrm{rg}}_{\mathrm{S}}$ are the rank variables of the satellite precipitation estimates and gauge observations; $\mathrm{cov}\left({\mathrm{rg}}_{\mathrm{G}},{\mathrm{rg}}_{\mathrm{S}}\right)$ is the covariance of the rank variables; and ${\mathsf{\sigma }}_{{\mathrm{rg}}_{\mathrm{G}}}$ and ${\mathsf{\sigma }}_{{\mathrm{rg}}_{\mathrm{S}}}$ are the standard deviations of the rank variables; H is the number of hits for a given forecast event by satellite precipitation data; F is the number of false alarms for a given forecast event by satellite precipitation data; and M is the number of misses for a given forecast event by satellite precipitation data. CC, POD, and FAR are dimensionless, while RMSE is in millimeters.

2.4. Mann–Kendall Test (M–K)

The rank-based Mann–Kendall test [57,58] was used in the drought trend analysis in this study. The method is highly recommended by the World Meteorological Organization and widely used [59,60,61] to assess the significance of monotonic trends in time series of various variables, for it has the advantage of not assuming any distributional form for the data and has the same power as its parametric competitors.

2.5. The Standardized Precipitation Index (SPI)

The major SPI procedures are to accumulate precipitation records at several timescales (typically from 1 to 24 months) and then to utilize the Gamma distribution or Pearson III frequency conversion, and the standardized normal distribution to calculate the normalized drought index value. More detailed information regarding the algorithm can be found in [45,50]. This study calculated SPI values by selecting the Gamma distribution and employing both the PERSIANN-CDR and monthly observed precipitation data from January 1983 to December 2017, denoted as the PERSIANN-CDR-based SPI and gauge-based SPI, respectively. We selected 3-month, 6-month, and 12-month time scales (SPI-3, SPI-6, and SPI-12) to represent short, medium, and long-term droughts, corresponding to the past 3, 6, or 12 months of precipitation totals, respectively. The SPI values generally range from −3 to +3, with positive values indicating wet conditions and negative values indicating dry conditions. Specific SPI drought categories are shown in

Table 1. Drought categories according to SPI values.

Category	SPI Value
Extremely wet	2.0 or above
Severely wet	1.5 to 1.99
Moderately wet	1.0 to 1.49
Near normal	−0.99 to 0.99
Moderately dry	−1.0 to −1.49
Severely dry	−1.5 to −1.99
Extremely dry	−2.0 or below

.

2.6. Drought Station Ratio P_j

Based on the SPI value, the drought station ratio P_j was adopted to describe the scope of drought. It is calculated as the proportion of stations where drought occurs out of the total number of stations in a certain period. The larger the value, the larger the area affected by drought.

$$P_j=\left(n_j/N\right)\times 100\%$$

(7)

Here, N is the total number of meteorological stations, and n_j is the number of stations where drought occurs in the jth year.

3. Results

3.1. Evaluation of PERSIANN-CDR Precipitation

We evaluated the performance of PERSIANN-CDR in estimating precipitation over the Huaihe River Basin based on the Spearman correlation coefficient ($\mathsf{\rho }$), relative bias (Bias), and root mean squared error (RMSE) [62]. The results demonstrated that PERSIANN-CDR shows poor correspondence with gauge observations at the daily scale but agrees on the magnitude. However, the monthly PERSIANN-CDR estimates perform much better, as indicated by a high value of $\mathsf{\rho }$ (0.97) and low Bias (5.12%). It can well capture the temporal and spatial distribution characteristic of precipitation over the study basin, indicating the feasibility of PERSIANN-CDR as a substitute for rain gauges with high-resolution rainfall sources.

In this study, we further used the probability of detection (POD) and false alarm ratio (FAR) to qualitatively diagnose the precipitation detection ability of the satellite. Figure 2 and Figure 3 show that the estimation performance of PERSIANN-CDR for different magnitudes of daily precipitation is variable. It has a good ability to capture and forecast precipitation events of less than 5 mm/day, indicated by a high POD (>0.86) and low FAR (<0.12) (Figure 3a and Figure 4a). Its performance worsens from north to south in space, with a higher POD and lower FAR in northern HN and SD. When precipitation is larger than 5 mm/day, the detection ability of PERSIANN-CDR for precipitation appears to decrease significantly, with POD lower than 0.5 (Figure 3b–d) and FAR higher than 0.8 over the basin (Figure 4b–d). The spatial distribution of POD and FAR is opposite to that of 0–5 mm/day precipitation, showing lower POD and higher FAR in the northern region of the basin. This indicates that the precipitation magnitude probably influenced the precipitation detection ability of daily PERSIANN-CDR estimates.

According to the definition of the FAR, a high FAR value means that the gridded precipitation data of PERSIANN-CDR misclassify most of the precipitation, such that it does not belong to that level or did not occur. Figure 5a shows that PERSIANN-CDR overestimates the frequencies of precipitation events with daily precipitation less than 25 mm but underestimates the frequencies of those precipitation events greater than 25 mm/day. It implies that the false alarms of 5–25 mm precipitation events are mainly due to the misjudgment of precipitation that does not belong to this level, which should come from the underestimation of precipitation events greater than 25 mm/day by satellites. In turn, it can be inferred that the false alarms of precipitation events greater than 25 mm/day are due to the underestimation of these precipitation events by satellites, which makes it out of this level. This further illustrates the influence of precipitation magnitude on satellite performance.

Figure 5b shows the precipitation changes in four seasons. It is clear that summer precipitation accounts for half of the annual precipitation, spring precipitation is the same as autumn precipitation, and winter precipitation is less than 10% of the year. Overestimation from PERSIANN-CDR is observed in all four seasons and is relatively significant in spring and winter. However, PERSIANN-CDR shows high agreement on the magnitude of monthly precipitation in autumn. This indicates that PERSIANN-CDR tends to overestimate precipitation events on a monthly scale.

Then, we compared the spatial distribution of precipitation derived from PERSIANN-CDR estimates and the gauge observations of the whole basin at the yearly scale (Figure 6). The spatial distribution patterns of annual precipitation derived from PERSIANN-CDR are in good agreement with those from the gauge observations in most precipitation magnitudes, except when the precipitation is greater than 1000 mm. They show similar spatial distribution patterns for each precipitation magnitude ranging from less than 800 mm/year to more than 1200 mm/year, demonstrating the good capture ability of the spatial distribution at the yearly scale. There is more area with more than 1000 mm annual precipitation in the spatial distribution map from PERSIANN-CDR, indicating that the satellite also tends to overestimate the yearly precipitation, especially for big precipitation events. It is interesting to find that the daily PERSIANN-CDR estimates have a higher POD and lower FAR (when precipitation >5 mm/day) in those areas with abundant precipitation. This means that the forecast skill of PERSIANN-CDR in space is also related to the climate condition.

To clarify whether the forecast skills of satellite estimates in space are affected by the spatial distribution characteristics of precipitation, we draw the spatial distribution map of Spearman correlation coefficient (ρ) between PERSIANN-CDR and gauge observations at the daily and monthly scale. Not surprisingly, the correlation coefficient for daily precipitation is not high in the whole space. However, it shows a similar spatial distribution pattern to those for yearly precipitation and has a relatively higher correlation in areas with abundance precipitation (Figure 7a). This suggests that the performance of daily PERSIANN-CDR estimates is strongly dependent on the climate condition, which is probably an influencing factor. Figure 7b shows that the spatial distribution of monthly values does not match that of precipitation, with high values (greater than 0.83) across the entire basin, indicating the reliability of the monthly PERSIANN-CDR estimates.

The above analysis indicates that daily PERSIANN-CDR estimates have a good detection ability for small precipitation events over the whole basin, but a low skill in capturing heavy precipitation events. The ability of daily PERSIANN-CDR estimates shows a strong dependence on the spatial distribution characteristics of precipitation. In contrast, the monthly and yearly PERSIANN-CDR estimates show good agreement both in magnitude and space. The performance of PERSIANN-CDR is influenced by the precipitation amount and climate condition of the basin.

3.2. Evaluation of PERSIANN-CDR in Drought Monitoring

The SPI values derived from the PERSIANN-CDR estimates and gauge observations were calculated, and drought conditions were determined by a grid value of SPI <−1.0. Figure 8 shows that the spatial distribution of the POD for drought event monitoring ranges from 0.51–0.73, demonstrating the good ability of PERSIANN-CDR to capture drought events. In space, the grid-based POD decreases as the time scale increases, with the highest average POD value at the SPI-3 scale, followed by SPI-6 and then SPI-12. For short-term droughts, SPI-3 of PERSIANN-CDR performs better in eastern SD, central HN, and western AH (>0.68), followed by AH (0.64~0.68), and worst in JS (0.6~0.64) (Figure 8a). For the medium-term drought, SPI-6 of PERSIANN-CDR decreases in all four provinces, with better performance in the same areas as for SPI-3 (Figure 8b). For the long-term drought, the POD of SPI-12 derived from PERSIANN-CDR is higher than that of SPI-6 in SD and middle HN, but much lower in western AH and JS (Figure 6c). This indicates that PERSIANN-CDR estimates are more reliable for drought monitoring in south-central HN, western AH, and northern SD.

The information presents by the spatial distribution of the FAR is similar to that of POD (Figure 9). The values of FAR for different time scales of the SPI are between 0.32 and 0.58, further indicating the reliability of PERSIANN-CDR in drought monitoring. We also observed lower FAR in south-central HN and northern SD, which are located in the northern part of the basin with less precipitation and in the southern parts of the basin with more precipitation, respectively. As stated above, monthly PERSIANN-CDR estimates tend to have a large error in areas with abundant precipitation. The results indicate that the impact of the precipitation estimation error is reduced when performing drought level assessment. This is similar to the conclusion in Reference [43].

Long-term droughts, which significantly impact watershed resources and agricultural production, are the primary concern in this study. Figure 10a shows that the time series of SPI-12 estimated from PERSIANN-CDR agrees well with that from gauge observations, as indicated by high CC values (0.937) and low RMSE (0.356). Good correspondence with gauge observations was also observed in drought coverage estimation, with a CC value of 0.935 (Figure 10b). This indicates that PERSIANN-CDR is reliable in capturing long-term drought events over the study basin.

We further calculated the number of drought events at different drought intensity levels according to the SPI-12 series (Figure 11a) to evaluate the ability of PERSIANN-CDR to monitor different levels of drought events. The results show that SPI-12 of PERSIANN-CDR generally overestimates the number of drought occurrences of drought events, with a significant overestimation of extreme drought occurrence, but an underestimation of severe and moderate droughts. Figure 11b shows that SPI-12 of PERSIANN-CDR overestimates the drought events with less than 10% drought area and those larger than 70% drought area. The results demonstrate that the SPI-12 values derived from PERSIANN-CDR estimates tend to overestimate the number of extreme droughts and the coverage of large-scale droughts. It is likely because of the overestimation of small precipitation events and underestimation of big precipitation events as shown in Figure 5a.

Table 2. Summary of trend analysis for drought frequency by Mann–Kendall test.

Category	SD		HN		AH		JS		Basin
	Gauge	Satellite	Gauge	Satellite	Gauge	Satellite	Gauge	Satellite	Gauge	Satellite
Moderate drought	−2.68 *	−1.09	−0.54	−0.81	0.37	−0.42	−0.53	0.16	0.14	−0.61
Severe drought	−0.95	−1.16	−0.18	−0.23	0.26	0.04	−0.42	−0.04	−0.60	−0.18
Extreme drought	−0.11	−0.02	0.42	0.28	0.51	0.21	−0.33	−0.67	0.14	−0.44
All droughts	−2.46 *	−0.81	−0.16	−0.21	0.53	0.11	−0.53	−0.19	0.35	−0.68

* Indicates significant with a confidence level of 95%.

shows the statistical results of trends in drought frequency at different intensities based on the Mann–Kendall test. Negative values indicate a downward trend in drought frequency in the future, while positive values indicate an upward trend. The PERSIANN-CDR is consistent with gauge observations in estimating the trends of droughts with moderate or above intensity in all four provinces. However, it gives an opposite result for moderate drought trends in AH and JS. The basin-based trend analysis from PERSIANN-CDR did not match the results of gauge observations. This suggests that PERSIANN-CDR estimates probably have potential utility for assessing local drought trends, but not for assessing widespread drought trends. Caution should be applied when using PERSIANN-CDR for drought trend analysis over the whole basin. Although the drought frequency in most provinces generally shows a downward trend, upward trends are identified in the AH province and the entire Basin scale, which needs more attention in the future.

3.3. Performance of PERSIANN-CDR in Specific Drought Events

We selected three severe drought events that occurred during 1984–2016 in the study basin according to the criteria of SPI-12 values less than −1 and drought station portion larger than 60%, as well as a drought period of more than three months. The three specific drought events are highlighted with gray background in their respective time series plots in Figure 10.

Table 3. Characteristics of three typical drought events at the 12-month timescale.

Time	Product	Start–End	Duration (Months)
1992	Gauge	June 1992–October 1992	5
	PERSIANN-CDR	June 1992–February 1993	8
1999–2000	Gauge	June 1999–June 2000	13
	PERSIANN-CDR	July 1999–June 2000	13
2001–2002	Gauge	August 2001–May 2002	10
	PERSIANN-CDR	August 2001–May 2002	10

lists their start–end time and duration, and Figure 12 shows the variation process of the three droughts over time. It is clear that the SPI from PERSIANN-CDR is in good agreement with the SPI from gauge observations regarding the start–end time and duration, especially for long-term drought, indicating that the satellite can better grasp the development process of long-term droughts in time. However, influenced by precipitation estimation errors, the droughts intensities in the three cases are misjudged. There is an overestimation for the 1992 drought event and an underestimation for the two drought events with a long history in 1999–2000 and 2001–2002. However, those under/overestimation do not affect the drought class determination much, and the estimation errors for severe and above droughts are basically controlled within the same drought class. This also illustrates that the commonly used SPI time series analysis can only test the ability of satellites to capture drought events in a general way.

Figure 13 shows the spatial distribution of the 1992 drought event at different stages of evolution. In June, a moderate and above-level drought co-occurred in the central and northern parts of the study area, while extreme drought occurred in HN and the junction area of JS and SD (Figure 13a). PERSIANN-CDR captures the occurrence of the drought event in time, but it presents a certain degree of underestimation of the drought intensity and coverage. It only shows the moderate and small-scale drought in HN and northern AH (Figure 13b). In July, drought spread across the basin. The SPI of PERSIANN-CDR clearly identifies the extreme drought area in SD and severe drought in northern JS and HN, with a certain degree of overestimation of extreme drought events in terms of drought intensity and coverage. After August, the drought gradually eased, and the basin-wide drought turned into a partial drought until its end in October 1992, with only a small range of moderate drought still in HN and SD. However, the SPI of PERSIANN-CDR still shows large areas of moderate drought and even severe drought in SD and part of HN and AH, lasting until January 1993. This case indicates that PERSIANN-CDR has a good ability to capture the spatial development and peak of long-term droughts. Still, it overestimates the drought coverage and intensity, as well as the drought duration, especially for extreme drought and severe drought.

Figure 14 compares the spatial distributions of drought events in 1999–2000, which lasted for 13 months and were divided into two main phases: June 1999 to February 2000 and March–June 2000. In June 1999, a moderate drought started in central AH, worsening and spreading northward in July and August. Severe and above-level drought covered almost the whole basin in August, with extreme drought covering 30% of the area and affecting all four provinces. Then, the drought gradually eased, with only the southern part of HN and the western part of AH sustaining moderate drought. During this phase, AH and HN were the two provinces that suffered the most severe and longest-lasting drought. The SPI of PERSIANN-CDR captures the spatial development and peak of the drought in July and August. It also does a good job of identifying the areas where the drought was most severe. During March–May 2000, the drought redeveloped to a second peak, mainly distributed in HN. PERSIANN-CDR captures the beginning and end of the second extreme drought, but the drought coverage expanded northward and was significantly overestimated. Like in the previous case, PERSIANN-CDR overestimates the coverage area of extreme drought and moderate drought.

The drought events from August 2001 to May 2002 started in southern AH and then spread throughout AH and HN (Figure 15). The drought affected all four provinces, with HN and AH being the most severely affected and SD being less affected. Extreme drought was entrenched in parts of Henan and Anhui for as long as seven months, until the month before the end of the drought. In this case, the SPI of PERSIANN-CDR agrees well with that of gauge observations in terms of the temporal and spatial distributions of the drought and accurately locates the drought coverage. This is the best performance among the three typical drought events, with slight underestimation of extreme drought in HN and overestimation of severe drought in SD.

By analyzing the three drought events, we can further understand the long-term drought development in time and space and the reliability of PERSIANN-CDR in drought monitoring. Basically, PERSIANN-CDR can well capture the start and end times of drought and the outbreak periods of droughts. It also shows a good ability to capture where droughts occur and how far they spread. The reliability of PERSIANN-CDR application in drought monitoring in the study basin is thus verified.

4. Discussion

The three selected drought cases all lasted more than five months and had a large impact area. The first two drought cases started and broke out in summer, and the third one started and broke out in autumn. These two seasons account for 70% of the annual precipitation of the HRB. Long-duration drought in these seasons typically affects reservoir re-storage in the flood recession period and exacerbates the adverse effects of drought, e.g., reductions in wheat and corn production. The main crops in the HRB are dry crops such as wheat and corn and a certain area of one-season rice. The growing cycle of wheat is from October to June, and the growing cycle of maize and one-season rice is from April to October. Undoubtedly, the happening of these drought events had a severe impact on the growth of major crops in the basin. We further draw the thermodynamic diagrams of drought events derived from the two precipitation data sets (Figure 16). It shows that spring drought had the highest frequency among the three seasons (except winter, with less than 10% precipitation of the whole year), indicating that the yields of corn and rice were more seriously affected by drought.

Figure 16 also shows that the SPI-12 from PERSIANN-CDR had a higher consistency with those from the gauge observations in autumn than in other seasons. It is likely because of the high agreement of PERSIANN-CDR on autumn monthly precipitation (Figure 5b) and could explain why the ERSIANN-CDR performed best in the third case. It is noteworthy that PERSIANN-CDR consistently tended to overestimate extreme drought coverage in the case study. We also observed that PERSIANN-CDR overestimated the frequency of drought events of all intensities at the basin scale (Figure 16). This indicates that the PERSIANN-CDR estimation error of precipitation will result in an overestimation of drought coverage and frequency. The current PERSIANN-CDR algorithm should undergo future algorithm development to improve its performance in drought monitoring.

Although a series of quality control measures were adopted in this study to obtain objective and reasonable assessment conclusions, some uncertainties are still inevitable, e.g., the uneven spatial distribution of meteorological stations, the type and installation methods of observation instruments, and precipitation observation methods. However, some uncertainties can be reduced in future studies. In this study, we obtained the area rainfall using the Tyson polygon method, widely accepted for calculating area rainfall. It ignores the influence of terrain and assumes that the gauges vary linearly from site to site, while the study basin has complex terrain. This uncertainty may introduce estimation errors into the evaluation conclusion. The inconsistent judgments on drought trend analysis of area rainfall (Table 2) are likely to be caused by this. This can be addressed by averaging multi-method results in future studies. Moreover, we used bilinear interpolation in the precipitation comparison analysis to obtain the point-based data. This method also considers that the precipitation in the grid changes linearly, which results in insufficient representativeness of point-based values due to the large range of the grid. A collection of denser rainfall stations with high-quality data could improve both the estimation accuracy of area rainfall and the comparability between the cell-based data and site-based data. Although these uncertainties exist, in this paper, we minimized the impact of data quality on the performance of remotely sensed precipitation products through strict quality control.

5. Conclusions

By evaluating the utility of PERSIANN-CDR products for drought monitoring in the Huaihe River basin, some main findings of this study can be summarized as follows:

(1)

The spatial performances of POD and FAR show that the precipitation forecasting skill of PERSIANN-CDR decreased as the precipitation magnitude increased. The performance of daily PERSIANN-CDR estimates is highly dependent on the spatial distribution of precipitation and performs relatively better in areas with abundant precipitation. In contrast, the monthly and yearly PERSIANN-CD estimates are highly consistent with gauge observations both in magnitude and space. However, PERSIANN-CDR tends to overestimate small precipitation frequency and underestimate big precipitation frequency, which should be addressed in future algorithm development.

(2)

These two indices (POD and FAR) were also calculated from the SPI values at various time scales (3, 6, and 12 months). The results show that the POD values are within 0.51–0.73 and the FAR values are within 0.32–0.58 in space, indicating the reliability of PERSIANN-CDR as precipitation input data for drought evaluation in the Huaihe River basin. Although its spatial-temporal performance decreases as the time scale of drought increases, the SPI-12 values estimated from PERSIANN-CDR are still satisfactory, as indicated by the high POD and the low FAR, as well as the high CC values (0.93) and low RMSE (0.356). However, when participating in drought trends analysis, the basin-based results from PERSIANN-CDR do not match that of gauge observations. Caution should be applied when using PERSIANN-CDR for drought trend analysis over the whole basin.

(3)

Three specific drought events are selected further evaluate the ability of PERSIANN-CDR in drought monitoring. Basically, PERSIANN-CDR captures the spatial development of long-term drought events, such as the start–end time, the outbreak periods, and the spread range. The reliability of PERSIANN-CDR application for drought monitoring in the study basin is thus well verified. The case study also shows that the SPI values of PERSIANN-CDR perform relatively better in autumn than in other seasons. This is probably related to the strong ability of PERSIANN-CDR to estimate autumn precipitation. However, due to the overestimation of low precipitation events, PERSIANN-CDR tends to overestimate the number of extreme droughts and their coverage.

It is essential to mention that the southern HN and AH provinces, which suffered more severe droughts, are located in the southern part of the basin and are also the areas with the most abundant precipitation. This means that these areas will be affected by both floods and droughts. PERSIANN-CDR data will provide robust data support for drought monitoring in these regions. Still, their performance in storm-related precipitation event monitoring needs to be further verified considering the significant errors in satellite estimation of extreme precipitation events.