1. Introduction
Strong rainfall processes— rainstorms, snowstorms, thunderstorms, and floods—and other abrupt natural disasters have the characteristics of rapid variation, short life cycle, and high intensity. Currently, the monitoring and forecasting of the occurrence, development, and evolution of these processes timely and accurately have become a crucial issue of meteorological studies. On one hand, however, the conventional meteorological observations are far from sufficient because of their limited temporal and spatial resolutions. On the other hand, although current operational numerical forecast has been greatly improved, it is still difficult to improve the forecast accuracy of the sudden disastrous weathers, especially for rainfall start time, area distribution, and intensity. The satellite observation has the advantages of wide observation range and high spatial resolution, which can be applied to real-time observation of a relatively large area. Therefore, it has become an effective tool for monitoring and short-term forecast of weather systems such as rainstorms (snowstorms) and intense thunderstorms.
For early geostationary satellite rainfall retrieval algorithms such as the infrared (IR) channel rainfall estimation (Stout et al. 1979; Adler and Mack 1984; Scofield 1987), the Geostationary Operational Environmental Satellite (GOES) precipitation index (GPI; Arkin and Meisner 1987), the convective–stratiform technique (CST; Adler and Negri 1988), and some simple IR–visible (VIS) algorithms (Lethbridge 1967; Griffith et al. 1978; Lovejoy and Austin 1979), the utilization of IR channel in the rainfall estimation was mainly based on the assumption that high cloud tops (i.e., the low brightness temperature) indicates strong convection and heavy precipitation (Miller et al. 2001). However, the physical properties of water droplets and/or ice particles within cloud were hardly considered in these algorithms. Therefore, for the retrieval of the warm cloud rainfall with relatively low cloud-top altitude, especially of the stratiform cloud rainfalls like nimbostratus rainfall, obvious defects are shown in these methods (Kidd et al. 2003).
Satellite passive microwave (PMW) remote sensing provides direct information of precipitation fields, and thus the instantaneous rain rate can be retrieved reliably (Ferraro 1997; Ferraro and Li 2002; Kummerow et al. 1996, 2001; McCollum and Ferraro 2003; Tapiador et al. 2004; Olson et al. 2006). However, microwave detectors such as Special Sensor Microwave Imager (SSM/I) and space-borne precipitation radar (PR) are always aboard polar orbiting satellites with low temporal resolution, and thus their applications are limited in time and space and are difficult to apply to operation retrievals (Morrissey and Janowiak 1996; Soman et al. 1995). However, the improvement is significant using the PMW observations to adjust the IR rainfall retrieval results (Adler et al. 1993; Kummerow and Giglio 1995; Xie and Arkin 1996; Xu et al. 1999; Todd et al. 2001; Huffman et al. 1995, 2001).
In recent years, with the rapid increasing of the temporal and spatial resolutions of satellite observation, the IR- and VIS-based rainfall retrieval techniques are developed rapidly. The GOES Multispectral Rainfall Algorithm (GMSRA) proposed by Ba and Gruber (2001) identifies the rainfall area using the thresholds of VIS albedo and IR brightness temperature or cloud droplet effective radius, and effectively improves the identification accuracy of the daytime rainfall with warm cloud top. On this basis, the rain rate is estimated according to the cloud-top temperature and adjusted by an empirical moisture factor. An operational GOES IR rainfall estimation technique named Auto-Estimator (AE) was developed by Vicente et al. (1998). The rainfall areas are identified using a set of threshold parameters including the cloud-top temperature gradient and cloud growth rate, and then, the rainfall intensity is estimated by using IR brightness temperature and adjusted according to convective equilibrium layer, atmospheric humidity, and vertical motion field, etc. The self-calibration multivariate precipitation retrieval (SCaMPR) algorithm (Kuligowski 2002) took advantage of the related algorithms of GMSRA and AE. The temperature value, temperature difference, temperature gradient, and the slope are considered as the threshold to judge whether rainfall occurs and the criteria to estimate rainfall intensity. Yan and Yang (2007) applied two Moderate Resolution Imaging Spectroradiometer (MODIS) channel (0.65 and 1.38 μm) data to form multiregression curves to estimate the daytime rainfall of eastern China. The results are relatively consistent with Advanced Microwave Scanning Radiometer for Earth Observing System (EOS) (AMSR-E) rainfall. Yu (1998, 2003) and Wang et al. (2008) used IR and VIS multispectral information from geostationary meteorological satellites (GMS) and multifunctional transport satellites (MTSAT) to generate two- and three-dimensional rainfall probability and rainfall intensity category matrices, and achieved relatively favorable results on the estimation of rainfall intensity distribution with five grades—namely, no rain, light rain, moderate rain, heavy rain, and rainstorm.
During the retrieving process, a critical problem is that a “true” value of rain rate needs to be specified for comparing with its counterpart (i.e., the satellite-retrieved value). On one hand, it directly determines whether the rainfall retrieval and the test results are reliable or not; on the other hand, it is the key factor to whether the retrieval method is meaningful for the synoptic and climatic analysis, and whether it is applicable in operational application.
What is usually used as the true values of the satellite retrieval is either the radar rain rate or the rain gauge rain rate. Radar rain rate is more physically significant; however, problems exist if it is directly used as the true value of the retrieval. It is an indirect estimation based on a Z–R relationship (Woodley et al. 1975) and, thus, the calibration, the cloud phase, and attenuation problems during the observation may result in the unstable Z–R relationship (Crosson et al. 1996; Zhang et al. 2001). Inevitably, it is bound to influence the accuracy of radar rain rate, or even result in great errors. The most difficult problem commonly seen is that it is often accompanied by a greater false alarm rate (FAR) when relatively good probability of detection (POD) is obtained. The gauge rain rate should be the closest to the true value after quality control processing. However, in the previous studies on satellite rainfall retrieval, gauge rain rate of one hour or an even longer period was usually used. The satellite observation is instantaneously completed, while the rain gauge measurements are accumulative rainfall during a period. The time inconsistency between both observation methods may inevitably bring errors to the satellite rainfall retrieval (Wei et al. 2006). The longer the accumulated period for the rain gauge measurements is, the greater the caused errors will be (this problem will be described in detail in section 3a). The same problem also exists when the radar rain rate is used as the true value of the satellite rainfall retrieval. Radar scanning is completed every six minutes, while the one-hour radar rain rate could be made up after scanning 10 times. If a half-hour or one-hour radar rain rate is taken as the true value for the satellite rainfall retrieval, the nephanalysis obtained by the instantaneous satellite observation has difficulty in fully characterizing the rainfall fields in this interval, thus the accuracy of rainfall retrieval is inevitably affected. This could be attributed to the obvious movement and development of the rainfall cloud, especially for the cumulonimbus inducing heavy rainfall, occurring in a half hour or one hour. However, if a 6-min radar rain rate is taken as the true value for satellite rainfall retrieval, the satellite observation can only retrieve a 6-min rain rate each time, but at present the interval of geostationary satellite observation is half an hour or even an hour, so that the complete continuous retrieval of rain rate cannot be implemented directly.
To solve this problem, this study is divided into two parts. The first part aims at improving the accuracy of rainfall retrieval. The study on satellite rainfall retrieval is carried out by using 10-min gauge rain rate as the true value. In the second part, the monitoring and nowcasting method of a half-hour rain rate matched with the satellite observations is further discussed on the basis of the study of the first part.
This paper is the first part of this study, consisting of five sections. Section 2 introduces the satellite and conventional observation data used in this study, section 3 describes the retrieval test of 10-min rain rate with daytime satellite images at a pixel-level resolution by combining the rainfall probability identification matrix (RPIM) and the multispectral segmented curve-fitting rainfall algorithm (MSCFRA), in section 4 an evaluation and case study for the algorithm is conducted, and section 5 is the discussion and conclusions.
2. Data
This study mainly focuses on daytime rainfall.
The spectral data adopted in this study come from the Visible and Infrared Spin Scan Radiometer (VISSR) from China’s second-generation geostationary meteorological satellite FengYun-2C (FY-2C), which has five channels: IR channel 1 (IR1 or IR; 10.3–11.3 μm), IR channel 2 (IR2; 11.5–12.5 μm), water vapor channel (WV; 6.3–7.6 μm), middle IR channel (MIR; 3.5–4.0 μm), and VIS channel (VIS; 0.55–0.90 μm). Except for the VIS channel providing information of albedo, other channels provide brightness temperatures. Existing studies indicate that the daytime rainfall is most closely related to the IR brightness temperature and VIS albedo (Cheng et al. 1993; Hsu et al. 1999; Behrangi et al. 2009; Ba and Gruber 2001; Yu 1998, 2003; Wang et al. 2008); and they are finally selected for the retrieval algorithm in this study.
The normalization has been carried out on the VIS albedo. The specific algorithm is shown in another paper (Zhuge et al. 2011, manuscript submitted to Atmos. Res.).
The equal longitude–latitude projection method is used for all the satellite images. For the projection center, the latitude is 29°42′N, the longitude is 111°12′E, and the spatial resolution is 0.05°. The half-hour-interval satellite images during a period of 180 days from June to August of 2007 and 2008 are used for study, of which the images in 2007 are used for the formation of the algorithm, while those in 2008 for the assessment of the algorithm.
As shown in Fig. 1, the 10-min rain rate data recorded in the same period with the satellite observation are selected as the rainfall data, which were acquired by more than 300 automatic rain gauges in Anhui, China (with the latitude range of 29°18′–34°52′N and the longitude range of 114°56′–119°37′E). Located in eastern China, Anhui is a critical area under the mei-yu front influence. In the late spring and early summer of each year, the Anhui area is greatly influenced by the mei-yu-front-mixed rainfall (rainfall derived from nimbostratus concomitant with cumulonimbus). Therefore, the rainfall data of Anhui has the representativeness.
Schematic diagram of the location of Anhui. The light gray area is China and the dark gray area is Anhui. The rectangular box with a dashed frame line is the primary area of mei-yu front occurrence. Black dots stand for the distribution of rain gauge stations used in this study.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
Anhui is located in the time zone of UTC + 08 h. Therefore, the time in this study refers to Beijing time (BT).
3. Satellite image rainfall method
a. Comparing satellite images with rain gauge rainfall
Only satellite grids corresponding to the location of the rain gauges are used to collect the training samples and derive the precipitation probabilities.
The satellite observation is instantaneous, while the rain gauge rainfalls are accumulative data during a period. Therefore, a suitable matching scheme should be considered when comparing the satellite with the actual ground-based observation.
The methods commonly used in satellites retrieving rainfall were to establish a statistical relationship between the satellite-instantaneous measurements and one-hour or longer period ground-measured rainfall, which usually caused significant biases in sampling. Because of the rapid movement and development of the cumulonimbus clusters inducing heavy rains, in the collocation, some heavy rainfall is often located in the cirrus zone at the edge of the convective cloud cluster, or even in the cloud-free zone ahead of the convective cloud, rather than in the cumulonimbus with deep convection. Obviously, the closer the accumulated time of measured rainfall to satellite-instantaneous observation time is, the smaller such errors will be. The collocation processes are carried out between the 10-min gauge rain rate and the satellite measurements in this study. The movement and development of cloud clusters are confined to 10 min, so the 10-min gauge rain rate matched up well with the location and development intensity of the cloud cluster.
The FY-2C satellite observations are started at every full hour and every half hour, and a full disk image is completed every half hour. Considering that the selected study area is located in the Northern Hemisphere around the latitude of 30°N, the first 10-min satellite observation has covered this area. A large number of case studies have proved that the 10-min gauge rainfall records started synchronously with satellite scanning (e.g., if the satellite observation starts at 1000 BT, the ground-measured rain rate of 1000–1010 BT is used) have the best matching relationship with the satellite measurements. Other circumstances used for contrast—that is, the records started 10 min prior (rain rate of 0950–1000 BT), 10 min lagging (rain rate of 1010–1020 BT), or 20 min lagging (rain rate of 1020–1030 BT)—have been demonstrated to be poorer (for further discussion, see section 4c). Therefore, the 10-min period after every full hour and half hour is defined as the best comparison period. The ground-measured rain rate in the best comparison period and the FY-2C satellite multispectral information is used for cooperative analysis in this study.
Figure 2 shows a case of contrast tests comparing the measured 1 h and 10 min rain rates with its corresponding cumulonimbus cloud. It can be seen from the figure that the 1-h heavy rainfalls occur frequently in the cirrus zones at the edge of cumulonimbus and even in the cloud-free zones outside the cloud as can be seen on the instantaneous satellite image; however, this is significantly decreased for heavy 10-min rainfalls.
Comparison between satellite images and heavy-rain gauge rainfalls. (a),(b) One-hour rainfall [(a) 1130–1230 and (b) 1200–1300 BT] and (c),(d) 10-min rainfall [(c) 1130–1140 and (d) 1200–1210 BT] in the best comparison period; the date is 23 Jun 2008.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
In Fig. 3, the comparison status between the typical heavy rainfall records and the IR–VIS-measured values is investigated in the selected study area from June to August 2008. The comparison of the one-hour measured rainfall and the IR–VIS satellite measurements [Fig. 3 (top)] is not good enough. The heavy rainfall points (>16 mm h−1) induced by the severe convective clouds are distributed even in the area with the IR brightness temperature (Tb) greater than 250 K and the albedo (A) lower than 0.5 (thin stratus or cirrus, or even cloudlessness usually appear over this region). This result is obviously inconsistent with the actual situations. If such measured dataset is used for the establishment of the satellite statistical rainfall retrieval, it is difficult to get reliable results. However, the comparison between the 10-min measured rainfall and the IR–VIS satellite measurements [Fig. 3 (bottom)] is relatively satisfactory in the best comparison period. The heavy rainfalls are obviously located in the area with low brightness temperature Tb and high albedo A. In particular, the heavy rainfall points with a rainfall greater than 8 mm (10 min)−1 are mainly located in the area with the IR brightness temperature Tb lower than 220 K and the VIS albedo A greater than 0.63.
The scatterplot of the comparison between the typical rainfall intensity center and the IR–VIS data from June to August 2008. (a) One-hour rainfall and (b) 10-min rainfall.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
b. Rain-area identification method
The key of the satellite rainfall retrieval is to identify rainfall clouds and nonrainfall clouds accurately and reasonably at first. On this basis, the rainfalls with different rainfall intensity grades can be further identified. The methods using a single threshold, such as VIS albedo (Ba and Gruber 2001), IR brightness temperature (Adler and Negri 1988; Arkin and Meisner 1987; Vicente et al. 1998), or effective radius of cloud droplets calculated from MIR brightness temperature (Rosenfeld and Lensky 1998; Rosenfeld 2000) usually appeared in the previous literature. The result of the rainfall area screened by double thresholds is better than that by a single threshold. For example, the threshold combination of IR brightness temperature and effective radius of cloud droplets was used by Rosenfeld and Gutman (1994), and that of VIS albedo and effective radius was used by Ba and Gruber (2001). The two-dimensional RPIM determined by the unit-feature-space classification method is a more suitable scheme to identify rainfall areas (Cheng et al. 1993; Yu 1998, 2003; Wang et al. 2008; Behrangi et al. 2009; Nauss and Kokhanovsky 2006).
In this study, the rainfall area identification is achieved by adopting RPIM. To establish RPIM, firstly the IR–VIS two-dimensional spectral space is divided into 64 × 64 basic units, named unit feature space. While reading the ground-based measured rainfall data, the corresponding IR and VIS measurement values are read from the dual-spectral images according to the latitude and longitude of the station, and the measured values must belong to a unit feature space (Wang et al. 2008). The number of rainfall samples and nonrainfall samples in each unit feature space is calculated, the probability of the rainfall occurrence of each unit feature space is determined, and 64 × 64 RPIMs are ultimately established. The final RPIM used in this study (Fig. 4) is determined by investigating 160 000 sets of rainfall data during June to August 2007.
Because of the sufficient and relatively reliable samples, this matrix well reflects the rainfall probability distribution under various combinations of brightness temperature Tb (vertical axis) and albedo A (horizontal axis). As viewed from RPIM, even if the relatively low brightness temperature of the cloud top indicates the existence of high cloud top, the area with low rainfall probability still may exist. The VIS albedos of these areas are relatively small, usually corresponding to the thick cirrus inducing no rain or light rain. However, relatively high rainfall probability often exists in the areas with higher cloud-top temperature but greater albedo, which often corresponds to nimbostratus and cumulonimbus with warm cloud top. Those areas with low cloud-top temperature and high albedo generally correspond to the cumulonimbus with high cloud tops, large vertical thickness, and vigorous convection. Therefore, this RPIM can be used for the analysis of rainfall possibility under the conditions of various brightness temperatures and albedo, and then rainfall area and nonrainfall area can be distinguished. With different rainfall probabilities used as the identification threshold, the rainfall areas to be analyzed will be different in size. If the threshold is too low, many areas without rain may be misjudged as rainfall areas; if the threshold is too high, the rainfall area may be too narrow, thus the rainfall areas may be judged as nonrainfall areas. Therefore, an appropriate rainfall probability threshold should be determined for rainfall area identification.
Contrast among evaluation indices with different rainfall probabilities used as the thresholds for screening rainfall areas.
As shown in Table 1, with the increase of threshold values, although FAR reduces significantly, the POD also declines, while skill scores (HSS, CSI, and ETS) show a parabola. When either the very high or very low threshold value is selected, the skill scores would all be rather low. With a comprehensive consideration, the threshold value set as 50% is relatively appropriate, and in this case, HSS, CSI, and ETS can reach 0.473, 0.419, and 0.309, respectively, while FAR is only 0.469, with POD reaching 0.665.
Table 2 shows the comparisons between the method of distinguishing rainfall and nonrainfall areas with RPIM and other conventional threshold methods. (As noted in Table 1, HSS, CSI, and ETS almost have the same denotation. Therefore, CSI and ETS are not specifically listed in Table 2.)
Comparison among different methods of screening rainfall areas.
In previous literature, the VIS albedo A threshold without correction was usually set to 0.4. Here A is elevated to 0.5 as a result of the normalization in this study. When considering clouds with IR brightness temperature Tb lower than 230 K and A greater than 0.50 (previous IR–VIS algorithm), 37.1% of the ground rainfalls can be detected by satellite, and the corresponding FAR value is 0.763 and the HSS is 0.061. Such low HSS indicates that quite a lot of rainfall samples are related to the warm clouds with cloud-top temperatures higher than 230 K. However, if all clouds with A over 0.50 are considered to be rainfall clouds, the POD value would be up to 0.943, while the FAR would also be significantly increased up to 0.674, and the HSS would be only 0.269. It demonstrates that the A threshold of 0.50 is not very satisfactory in sifting out nonrainfall warm clouds. When the threshold combination of A and effective radius (Re) proposed by Ba and Gruber (2001) is used, HSS can reach as high as 0.350, and POD is 0.832, while FAR is still as high as 0.614. However, if the threshold of Re is increased to 20 μm, POD would be reduced rapidly, while FAR remains almost unchanged, resulting in no increase of HSS. With the RPIM used, the rainfall probability is set at 50%, and the HSS reaches up to 0.473, the FAR is only 0.469, while the POD reached up to 0.665. Comparing the previous screening method with the one using a rainfall probability threshold of 50% adopted by RPIM, it is found that the latter significantly improves the correct detection rate of the rainfall areas without significant change of the FAR value.
c. Satellite retrieval test of 10-min rain rate
The rain rates are estimated only when the cloudy pixels are determined as rainfall cases.
The study mainly focuses on mixed-type rainfall and deep convective cloud rainfall inducing floods and rainstorms in the central and eastern part of mainland China from May to August, with an emphasis placed on mei-yu front rainfall. Rain rate of such type is significantly correlated with cloud-top height and its thickness. Therefore, IR brightness temperature representing cloud-top height and VIS albedo representing cloud thickness are used as important factors for estimating such types of rainfall. The algorithm proposed in this study is based on the following observed facts: the strong cumulonimbus to induce heavy rains and rainstorms has the cloud-top brightness temperature often less than 220 K, accompanied by intermittent ground rainfalls with large instantaneous intensity, while the stratiform clouds like stratus and nimbostratus to induce continuous rainfalls often have a cloud-top brightness temperature between 240 and 270 K. Those with a brightness temperature between that of the above two types are mainly weak convective clouds and multilayer clouds.
Earlier rainfall algorithms generally estimated rain rate through IR brightness temperature (Adler and Negri 1988). On this basis, rain rate was adjusted appropriately through cloud-top brightness temperature growth factor, relative humidity, and other parameters (Ba and Gruber 2001; Kuligowski 2002; Vicente et al. 1998). The recent typical multicurve-fitting scheme was to divide precipitation clouds on satellite cloud images into six categories by means of clustering analysis, and rainfall was fitted with brightness temperature of each type of cloud (e.g., Hong et al. 2004). IR–VIS two-dimensional rainfall intensity category matrix based on a large number of statistical samples reveals the statistical fact that the lower the brightness temperature and the greater the albedo is, the greater the rain rate will be, which can be used to distinguish rainfall with different intensities (Yu 1998, 2003; Luque et al. 2006; Wang et al. 2008). For different VIS albedo, rainfall estimation could also be conducted by fitting the reflectance of 1.38 μm (R1.38), thus forming a dual-spectrum multicurve-fitting algorithm (Yan and Yang 2007). This algorithm indicates that for the same R1.38, the larger the albedo is, the greater will be the rainfall intensity, which also proves that it is feasible to directly fit rain rate with VIS albedo.
To take full account of the differences between the convective rainfall and the stratiform rainfall, the MSCFRA for the 10-min satellite rain rate retrieval is proposed in this study. This algorithm significantly differs from the methods described in previous literature. The rain rate (RR) is not directly fitted by IR brightness temperature Tb. However, according to various types of clouds, we generate a nonlinear function inferred from VIS albedo A, IR brightness temperature Tb, and 10-min rain gauge RR to fit the rainfall intensity. The specific steps are listed as follows: cloud-top brightness temperature Tb is first divided into several ranges belonging to convective cloud rainfall, stratiform cloud rainfall, as well as weak convective or multilayer cloud rainfall lying in between. The Tb range width is 20 K in this algorithm. Considering the smooth transition between different fitting functions, the adjacent ranges are overlapped. Thus, seven Tb ranges are determined finally. In various Tb ranges, different curves are used to fit RR with VIS albedo A, and a cubic function is fitted in Fig. 5. Table 3 lists the corresponding fitted equation.
The functional relationship between RR and A for different Tb ranges.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
The fitting curves of RR [mm (10 min)−1] with A in different ranges of Tb.
4. Test and case analysis
In this study, rainfall grades are involved in the 10-min satellite retrieval rainfall test. As no grade division was referred to for the 10-min rainfall in the operational forecast, it is defined here as follows (Table 4).
Rainfall grade scale defined in this paper (applicable for 10-min rainfall).
a. Overall evaluation
Taking into account the characteristics of instantaneous satellite observations and potential cloud nonlinear movement in 10 min, during the accuracy test of the satellite retrieval rain rate under the pixel resolution, the comparison between point-to-point values (the rain rate of each individual rain gauge station to that of satellite retrieval at that station) is less suitable. Therefore, the comparison between a point (the rain rate of each individual rain gauge station) and an area [retrieval rainfall area with that station as the center and a radius of the “comparison radius (Rc)”] is conducted here. The satellite retrieval rain rate most approximating the rain gauge rain rate within the “area” is taken to represent the estimate at that station. Rc is determined by the cloud movement speed. Obviously, different values of Rc will lead to significantly different statistical results. To evaluate retrieval accuracy objectively, Rc is set as 1, 3, and 5 pixels to assess the 10-min-interval rainfall estimation algorithm.
A total of 11 rainfall events from June to August 2008 are selected for assessment with samples totaling 11 007.
Figure 6 (top) is a scatterplot of satellite-estimated rain rate and the rain gauge counterparts under different Rc. When the Rc is set as 1, the estimation effect of heavy rainfall [>12 mm (10 min)−1] is relatively good, although a small number of rainfalls are significantly underestimated. The weak rainfall [<5 mm (10 min)−1] is likely to be overestimated as well as underestimated. If Rc = 3, the majority of rainfalls are estimated successfully, and a large numbers of rainfall samples are concentrated on the diagonal line. The overestimation of slight rainfall has been significantly reduced, and no rainfalls of <1 mm (10 min)−1 are overestimated as greater than 8 mm (10 min)−1. If Rc = 5, the accuracy of rainfall estimation is further improved accordingly, and the underestimation and overestimation of weak rainfall are effectively controlled. Most estimated rain rates and rain gauge rain rates correspond to each other very well.
The (top) scatterplot and (bottom) error bar plot of estimated RR and rain gauge RR, where the abscissa represents the rain gauge–measured RR, and the ordinate represents the satellite-estimated RR. The symbol + represents samples; seven error bars represent seven grades of rainfall (Table 4), and × represents the mean value of estimated rainfall of each grade. The length of each error bar represents the rmsd of each grade.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
Figure 6 (bottom) is the error bar plot revealing the error distribution of rainfalls of seven grades quantitatively. It can be seen from the figure that when Rc = 1, the rainfalls of the first two grades are overestimated as a whole, rmsd also exceeding 1 mm (10 min)−1. The rainfall mean values of grades 5, 6, and 7 are distant from the diagonal line. When Rc = 3, mean values of estimated rainfalls of the first three grades are located on the diagonal line, indicating that the mean values of estimated rainfalls are consistent with the measurements. The mean values of estimated rainfalls of grades 4, 5, and 6 are very close to the diagonal line with rmsd no greater than 3 mm (10 min)−1. That of the rainfall of grade 7 is farther away from the diagonal line with greater rmsd. If Rc is widened to 5, rmsd of the first three grades will be less than 0.5 mm (10 min)−1, and those of grades 4, 5, and 6 will be reduced further. However, the estimated value of rainfall of grade 7 does not improve significantly compared with the case when Rc = 3, indicating that some defects exist in the algorithm proposed in this study on the estimation of heavy rainfall with the rain rate greater than 12 mm (10 min)−1.
The statistical indicators are used for estimation, and 
It should be noted that this is the statistics of nearly all the rainfall samples from June to August 2008. Statistical results prove that the algorithm proposed in this study is credible and practical.
b. Case study
Since the temporal–spatial inconsistency in the satellite observations and ground measurements is reduced by this algorithm, 10-min rainfall can be retrieved relatively accurately. As for both the division of rainfall and nonrainfall areas and the identification of different rainfall grades, retrieval results are relatively consistent with rain gauge observations. The identification of heavy rainfall centers also has a small bias from the rain gauge observations.
1) 10 June 2008
It was a nimbostratus rainfall process. Instantaneous rainfall was not great throughout Anhui Province, but the range of rainfall area was relatively large with longer duration. On the satellite images and rainfall distribution maps, the rainfall showed significant regionalization characteristics, among which cirrus dominated in northern Anhui with no rainfall; influenced by rainfall cloud clusters, a wide range of rainfall occurred in southern Anhui; continuous light rain occurred in central Anhui in the mid–low cloud area.
The continuous tracking results of retrieval tests of this rainfall cloud cluster are shown in Fig. 7. In sequence, the times of the rainfall retrieval effect map are (a) 1130–1140, (b) 1200–1210, (c) 1230–1240, and (d) 1300–1310 BT. The rainfall and nonrainfall areas (including cloud-free zones) are divided using RPIM. The nonrainfall area is shown by the grayscale IR brightness temperature. The rain rates are estimated with the decision function listed in Table 3. On the map, dots stand for the ground-measured rainfall, and the different colors corresponding to different color bars denote rainfalls of various grades.
The effect map of the 10-min rainfall retrieval for 10 Jun 2008: the corresponding times are (a) 1130–1140, (b) 1200–1210, (c) 1230–1240, and (d) 1300–1310 BT. The division of rainfall grades is shown in Table 4.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
The estimation statistics of rainfalls of various grades are shown in Fig. 8, where Rc = 3. The ordinate represents the satellite-estimated rainfall grade, and the abscissa represents the measured rain gauge rainfall grade. Grade 0 denotes no rain, and the sample values of the diagonal indicate that the rainfall grade is correctly estimated.
Frequency distribution chart of 10-min rainfall retrieval results (10 Jun 2008): the horizontal axis is the 10-min rain gauge rainfall grade, and the vertical axis is the 10-min satellite-estimated rainfall grade. The division of rainfall grades is shown in Table 4.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
The accuracy rates of rainfall estimation of grades 2, 3, and 4 are higher, as 70%, 79%, and 71%, respectively. That of grade 1 is 50%, but there are still 40% of rainfalls estimated as no rain. All the rainfalls of grade 5 are underestimated only by one or two grades. A total of two samples are used for the rainfall of grade 6, and both are underestimated (station A in Fig. 7b and station B in Fig. 7c), indicating that some defects exist in the algorithm proposed in this study in the estimation of heavy rainfall induced by nimbostratus.
It should be noted that the rainfall of grade 6 is measured at station A in Fig. 7b [with the specific rain rate of 9.6 mm (10 min)−1] but is underestimated as grade 4. The 10-min rainfall measured at station B in Fig. 7c is 9.1 mm. However, the IR brightness temperature is about 242 K, and VIS albedo is 0.54 at this pixel with no spectral characteristics of strong cumulonimbus, while the estimated rainfall is only 0.26 mm. Thus, it can be judged that low resolution of satellite observation is likely the reason for underestimation or even serious underestimation. After all, a pixel on a satellite image represents the spectral average situation of an area of 0.05°, while the rainfall data collected by ground observation stations are the rain gauge–measured values at a single station. When the diameter of convection cloud cell is less than a pixel resolution or cannot cover the pixel, such situation will happen inevitably. However, some instrumental errors might also occur to this rain gauge measurement at this station.
2) 15 August 2008
It was a convective rainfall process (Fig. 9). Cloud clusters C and D gradually invaded Anhui Province from west to east, and disappeared entirely at around 1500 BT. Primary convection occurred at 1000 BT induced by cloud cluster E, and small convective cloud cells were generated at 1100 BT, reaching their maximum intensity at 1400–1430 BT. Because of the influence of three convective clouds, slight rains [<0.5 mm (10 min)−1] happened at some of the stations in northern Anhui. This algorithm mainly focuses on the development process of cloud cluster E and the estimation on the induced rainfalls. In sequence, the times of rainfall retrieval effect diagram are (a) 1300–1310, (b) 1330–1340, (c) 1400–1410, and (d) 1430–1440 BT.
As in Fig. 7, but for 15 Aug 2008: the corresponding times are (a) 1300–1310, (b) 1330–1340, (c) 1400–1410, and (d) 1430–1440 BT.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
It can be seen from the satellite image that the gradient of cloud-top brightness temperature is very great, resulting in the large gradient of estimated rainfall. It is relatively difficult to divide rainfall areas in northern Anhui Province. Nonrainfalls at many stations are estimated as slight rain [<0.5 mm (10 min)−1], but a number of slight rains are estimated as no rain. The transitional state between slight rain and no rain is still difficult to grasp with this algorithm. However, the correct or wrong identification of such rainfall has little impact on the algorithm assessment. The statistics of rainfall estimation of various grades in this rainfall process are shown in Fig. 10. It can be seen from the figure that the accuracy rates of rainfall estimations ranging from grades 2 to 7 all exceed 60%. It should be especially pointed out that a total of 11 samples are used in the estimation of heavy rainfall (>grade 6), and nine estimations are absolutely correct, while the two samples with estimation errors are only underestimated by one or two intensity grades. It indicates that the algorithm proposed in this study can well locate the rainfall intensity center of convective cloud clusters. Because of the large brightness temperature gradient of the cloud cluster, the estimation effect of moderate rainfall is not as good as that of nimbostratus rainfall (a case on 10 June), but the accuracy rates of grades 2 and 3 rainfall still reach 0.63 and 0.70, respectively.
As in Fig. 8, except for 15 Aug 2008.
Citation: Journal of Hydrometeorology 12, 6; 10.1175/2011JHM1373.1
c. Further discussion on time comparisons between gauge rainfall and satellite images
As mentioned in section 3a, the 10-min gauge rainfall records started synchronously with satellite scanning have the best matching relationship with the satellite measurements. It will be further explained in this section.
Table 5 shows the statistics of the comparison test results between satellite-estimated rainfall and ground-measured rainfall during different periods. The rainfall estimation algorithm used is still the method proposed in this study, where Rc = 1 during statistics.
Comparison test results between satellite-estimated rainfall and ground-measured rainfall during different periods.
The current scheme (matching up the satellite-estimated rainfall with the records of 10-min gauge rain rate synchronized with satellite scanning) is used, with the cc of 0.569 and rmsd of 1.4457 mm (10 min)−1. If the satellite image is matched up with ground-measured rainfall lagging 10 min behind the start time of satellite scanning, the cc will be slightly lower as 0.5184, and rmsd will be 1.5047 mm (10 min)−1. When other time comparison schemes are used (10 min prior or 20 min lagging behind), cc is very low in all cases, while rmsd all exceed 2 mm (10 min)−1.
Statistical results show that the satellite-estimated rainfall has the best matching relationship with the first 10-min rain gauge rainfall synchronized with satellite scanning. If accumulated rainfall within shorter periods could be achieved with rain gauges, then a further discussion could be carried out on the impact of time.
5. Conclusions
As the first part of “Rainfall Retrieval and Nowcasting Based on Multispectral Satellite Images,” this study focuses on the basic method of the rainfall retrieval with IR–VIS images, using 10-min measured rain rate as the true value. Based on reasonable division of rainfall areas, the multispectral segmented curve-fitting rainfall algorithm (MSCFRA) is used for estimating the 10-min rainfall with this method.
The rainfall assessment results show that when Rc = 3, the correlation coefficient between estimated rain rate and measured rain rate has already reached 0.8, and rmsd reaches 0.9097 mm (10 min)−1. In the investigations of rainfall grade estimation, regardless of whether for stratiform cloud rainfall or for convective cloud rainfall, a majority of estimated rainfall grades are consistent with rain gauge measurements. Especially, for convective cloud rainfall, the heavy rainfall center can be accurately estimated in most cases.
Favorable rainfall estimation results are achieved with the algorithm proposed in this study, and the characteristics are summarized as below:
This algorithm is conducted in two steps: first, rainfall and nonrainfall areas are determined, and then the rain rate at the pixel of rainfall area is retrieved, thus excluding the interference of nonrainfall pixels on rain rate retrieval process.
A reasonable RPIM is established with adequate samples, thus accurately distinguishing rainfall from nonrainfall areas.
MSCFRA algorithm is established with a full use of IR and VIS spectral characteristics. First, IR brightness temperature range of cloud top is used to characterize the different types of rainfall clouds and their development intensities. Then, a function relationship is established on each brightness temperature range by using VIS albedo and rain rate directly.
The modeling, retrieval, and test data are of high temporal–spatial resolutions (the original resolution information of 0.05° is maintained in satellite image, and the measured rainfall data are 10-min gauge rainfall) to ensure the rainfall retrieval with high accuracy with this method.
The normalization scheme of VIS images used in this study extends the hour of VIS images available for use in the summer half year from 0800 to 1600 local time. However, since this scheme is inapplicable to the areas with solar zenith angle or satellite zenith angle greater than 60°, a great error will occur with this algorithm in rainfall estimations of high-latitude areas. In the winter half year, VIS image during a limited number of hours can be applied to midlatitude areas (Zhuge et al. 2011, manuscript submitted to Atmos. Res.). Therefore, the rainfall algorithm in this study is mainly used to estimate the midlatitude rainfall in the summer half year. Whether this algorithm can be applied to tropical convective rainfall remains to be studied further.
To make the satellite rainfall retrieval method significant in synoptic and climatic analysis and able to be put into operational application, the continuous no-gap rainfall retrieval should be implemented. At present, the commonly used geostationary satellite observation data are basically obtained every half hour, and therefore, the half-hour rainfall retrieval should be achieved based on the assurance of rainfall retrieval accuracy. In part two of this study (Yu et al. 2011) the 10-min rainfall retrieval method proposed here will be used to further conduct half-hour satellite rainfall retrieval tests by combining the cross-correlation cloud tracking technique, and a continuous retrieval and nowcasting method for half-hour rainfall will be proposed.
The 10-min rainfall at nighttime, dawn, and dusk will be estimated in the subsequent study.
Acknowledgments
The authors would like to thank Dr. Yong Huang for providing FY-2C satellite data and rainfall data and Mr. Xun-shu Song for drawing some of the figures in this paper. The authors would also like to thank the anonymous reviewers for their helpful comments and valuable suggestions which improved the manuscript. This work was financially supported by the National Natural Science Foundation of China under Grant 40875012, the National Grand Fundamental Research 973 Program of China under Grant 2009CB421502, and by the Huaihe River Basin Meteorological Open Fund (HRM200704).
REFERENCES
Adler, R. F., and Mack R. A. , 1984: Thunderstorm cloud height–rainfall rate relations for use with satellite rainfall estimation techniques. J. Climate Appl. Meteor., 23, 280–296.
Adler, R. F., and Negri A. J. , 1988: A satellite infrared technique to estimate tropical convective and stratiform rainfall. J. Appl. Meteor., 27, 30–51.
Adler, R. F., Negri A. J. , Keehn P. R. , and Hakkarinen I. M. , 1993: Estimation of monthly rainfall over Japan and surrounding waters from a combination of low-orbit microwave and geosynchronous IR data. J. Appl. Meteor., 32, 335–356.
Arkin, P. A., and Meisner B. N. , 1987: The relationship between large-scale convective rainfall and cold cloud over the Western Hemisphere during 1982-84. Mon. Wea. Rev., 115, 51–74.
Ba, M. B., and Gruber A. , 2001: GOES Multispectral Rainfall Algorithm (GMSRA). J. Appl. Meteor., 40, 1500–1514.
Behrangi, A., Hsu K. , Imam B. , Sorooshian S. , and Kuligowski R. J. , 2009: Evaluating the utility of multispectral information in delineating the areal extent of precipitation. J. Hydrometeor., 10, 684–700.
Cheng, M., Brown R. , and Collier C. G. , 1993: Delineation of precipitation areas using Meteosat infrared and visible data in the region of the United Kingdom. J. Appl. Meteor., 32, 884–898.
Crosson, W. L., Duchon C. E. , Raghavan R. , and Goodman S. J. , 1996: Assessment of rainfall estimates using a standard Z–R relationship and the probability matching method applied to composite radar data in central Florida. J. Appl. Meteor., 35, 1203–1219.
Ferraro, R. R., 1997: Special sensor microwave imager derived global rainfall estimates for climatological applications. J. Geophys. Res., 102, 16 715–16 735.
Ferraro, R. R., and Li Q. , 2002: Detailed analysis of the error associated with the rainfall retrieved by the NOAA/NESDIS Special Sensor Microwave/Imager algorithm 2. Rainfall over land. J. Geophys. Res., 107, 4680, doi:10.1029/2001JD001172.
Griffith, C. G., Woodley W. L. , Gruber P. G. , Martin D. W. , Stout J. , and Sikdar D. N. , 1978: Rain estimation from geosynchronous satellite data—Visible and infrared studies. Mon. Wea. Rev., 106, 1153–1171.
Hong, Y., Hsu K. , Soorooshian S. , and Gao X. , 2004: Precipitation estimation from remotely sensed imagery using an artificial neural network cloud classification system. J. Appl. Meteor., 43, 1834–1852.
Hsu, K., Gupta H. V. , Gao X. , and Sorooshian S. , 1999: Estimation of physical variables from multichannel remotely sensed imagery using a neural network: Application to rainfall estimation. Water Resour. Res., 35, 1605–1618.
Huffman, G. J., Adler R. F. , Rudolf B. , Schneider U. , and Keehn P. R. , 1995: Global precipitation estimates based on a technique for combining satellite-based estimates, rain-gauge analysis, and NWP model precipitation information. J. Climate, 8, 1284–1295.
Huffman, G. J., Adler R. F. , Morrissey M. M. , Bolvin D. T. , Curtis S. , Joyce R. , McGavock B. , and Susskind J. , 2001: Global precipitation at one-degree daily resolution from multisatellite observations. J. Hydrometeor., 2, 36–50.
Kidd, C., Kniveton D. R. , Todd M. C. , and Bellerby T. J. , 2003: Satellite rainfall estimation using combined passive microwave and infrared algorithms. J. Hydrometeor., 4, 1088–1104.
Kuligowski, R. J., 2002: A self-calibrating real-time GOES rainfall algorithm for short-term rainfall estimates. J. Hydrometeor., 3, 112–130.
Kummerow, C., and Giglio L. , 1995: A method for combining passive microwave and infrared rainfall observations. J. Atmos. Oceanic Technol., 12, 33–45.
Kummerow, C., Olson W. S. , and Giglio L. , 1996: A simplified scheme for obtaining precipitation and vertical hydrometeor profiles from passive microwave sensors. IEEE Trans. Geosci. Remote Sens., 34, 1213–1232.
Kummerow, C., and Coauthors, 2001: The evolution of the Goddard profiling algorithm (GPROF) for rainfall estimation from passive microwave sensors. J. Appl. Meteor., 40, 1801–1820.
Lethbridge, M., 1967: Precipitation probability and satellite radiation data. Mon. Wea. Rev., 95, 487–490.
Lovejoy, S., and Austin G. L. , 1979: The delineation of rain areas from visible and infrared satellite data for GATE and mid-latitudes. Atmos.–Ocean, 17, 1048–1054.
Luque, A., Gómez I. , and Manso M. , 2006: Convective rainfall rate multi-channel algorithm for Meteosat-7 and radar derived calibration matrices. Atmósfera, 19, 145–168.
McCollum, J. R., and Ferraro R. R. , 2003: Next generation of NOAA/NESDIS TMI, SSM/I, and AMSR-E microwave land rainfall algorithms. J. Geophys. Res., 108, 8382, doi:10.1029/2001JD001512.
Miller, S. W., Arkin P. A. , and Joyce R. , 2001: A combined microwave/infrared rain rate algorithm. Int. J. Remote Sens., 22, 3285–3307.
Morrissey, M. L., and Janowiak J. E. , 1996: Sampling-induced conditional biases in satellite climate-scale rainfall estimates. J. Appl. Meteor., 35, 541–548.
Nauss, T., and Kokhanovsky A. A. , 2006: Discriminating raining from non-raining clouds at mid-latitudes using multispectral satellite data. Atmos. Chem. Phys. Discuss., 6, 1385–1398.
Olson, W. S., and Coauthors, 2006: Precipitation and latent heating distributions from satellite passive microwave radiometry. Part I: Improved method and uncertainties. J. Appl. Meteor. Climatol., 45, 702–720.
Rosenfeld, D., 2000: Suppression of rain and snow by urban and industrial air pollution. Science, 287, 1792–1796.
Rosenfeld, D., and Gutman G. , 1994: Retrieving microphysical properties near the tops of potential rain clouds by multispectral analysis of AVHRR data. J. Atmos. Res., 34, 259–283.
Rosenfeld, D., and Lensky I. M. , 1998: Satellite-based insights into precipitation formation processes in continental and maritime convective clouds. Bull. Amer. Meteor. Soc., 79, 2457–2476.
Scofield, R. A., 1987: The NESDIS operational convective precipitation-estimation technique. Mon. Wea. Rev., 115, 1773–1792.
Soman, V. V., Valdes J. B. , and North G. , 1995: Satellite sampling and the diurnal cycle statistics of Darwin rainfall data. J. Appl. Meteor., 34, 2481–2490.
Stout, J. E., Martin D. W. , and Sikdar D. H. , 1979: Estimating GATE rainfall with geosynchronous satellite images. Mon. Wea. Rev., 107, 585–598.
Tapiador, F. J., Kidd C. , Levizzani V. , and Marzano F. S. , 2004: A neural networks–based fusion technique to estimate half-hourly rainfall estimates at 0.1° resolution from satellite passive microwave and infrared data. J. Appl. Meteor., 43, 576–594.
Todd, M. C., Kidd C. , Kniveton D. , and Bellerby T. J. , 2001: A combined satellite infrared and passive microwave technique for estimation of small-scale rainfall. J. Atmos. Oceanic Technol., 18, 742–755.
Vicente, G. A., Scofield R. A. , and Menzel W. P. , 1998: The operational GOES infrared rainfall estimation technique. Bull. Amer. Meteor. Soc., 79, 1883–1898.
Wang, C., Yu F. , and Zhao Y. , 2008: Inversion study of rainfall intensity field at all time during Mei-Yu period by using MTSAT multi-spectral imagery. Multispectral, Hyperspectral, and Ultraspectral Remote Sensing Technology, Techniques, and Applications II, A. M. Larar, M. J. Lynch, and M. Suzuki, Eds., International Society for Optical Engineering (SPIE Proceedings, Vol. 7149), 1–12.
Wei, C., Hung W. , and Cheng K. , 2006: A multispectral spatial convolution approach of rainfall forecasting using weather satellite imagery. Adv. Space Res., 37, 747–753.
Woodley, W. L., Olsen A. R. , Herndon A. , and Wiggert V. , 1975: Comparison of gage and radar methods of convective rain measurement. J. Appl. Meteor., 14, 909–928.
Xie, P., and Arkin P. A. , 1996: Analyses of global monthly precipitation using rain-gauge observations, satellite estimates, and numerical model predictions. J. Climate, 9, 840–858.
Xu, L., Gao X. , Sorooshian S. , Arkin P. A. , and Iman B. , 1999: A microwave infrared threshold technique to improve the GOES precipitation index. J. Appl. Meteor., 38, 569–579.
Yan, H., and Yang S. , 2007: A MODIS dual spectral rain algorithm. J. Appl. Meteor. Climatol., 46, 1305–1323.
Yu, F., 1998: Extraction of meteorological characteristic elements from GMS images and their application to mesoscale NWP (in Chinese). Ph.D. dissertation, Nanjing University, 137 pp.
Yu, F., 2003: Analysis of rainfall intensity field by using multi-spectral GMS imagery (in Chinese). Acta. Meteor. Sin., 61, 334–345.
Yu, F., Zhuge X. , and Zhang C. , 2011: Rainfall retrieval and nowcasting based on multispectral satellite images. Part II: Retrieval study on daytime half-hour rain rate. J. Hydrometeor., 12, 1271–1285.
Zhang, P., Du B. , and Dai T. , 2001: Radar Meteorology. China Meteorological Press, 499 pp.
