Interactive comment on “ Comparison and evaluation of satellite derived precipitation products for hydrological modeling of the Zambezi River Basin ” by T

The overall presentation of this manuscript is well structured and clear. However, some details of the description are not very clear and concise. The statistics used needs to be justified. The language, in general, is pretty good, but I did notice a few grammar errors and typos. In summary, this interesting research is conducted reasonably well, and results are helpful to partially mitigate the lack of adequate evaluation of satellite products over Africa. Therefore, I recommend that this manuscript be accepted for publication in Hydrology and Earth System Sciences, after the following comments are carefully addressed.


Introduction
Water resources management in tropical and semi-arid areas of Africa is particularly important due to the high temporal and spatial climatic variability that affects availability of water resources within and between countries and river basins. The overarching goal of the African Dams Project: Adapt planning and operation of large dams to meet social needs and environmental constraints (ADAPT) is to strengthen this interdisciplinary science. A consistent information platform for a large scale river catchment, the Zambezi River basin, is currently under development. Modeling the hydrology of this basin is a challenging task due to its size and heterogeneity, but mostly, due to the lack of reliable input and calibration data. In the past, several studies addressed 10 the problem by using or assessing novel satellite derived data sources in addition to rainfall, such as evaporation (Winsemius et al., 2008), terrestrial water storage change (Winsemius et al., 2006b) and soil moisture (Meier et al., 2011). However, the satellite derived rainfall data were rarely evaluated even if, concerning model performance, the selection of the type of input precipitation has been considered as equally or even more 15 important than the choice of the hydrological model.
In view of the sparse available gauging network for rainfall monitoring on the African continent, the observations from spaceborne instrumentation currently produce the only measured data for a large part of the territory. Two types of sensors are commonly used in the satellite rainfall estimation algorithms: Passive Microwave (PM) and 20 Visible and Infrared Radiance (VIS/IR). The PM sensors identify the precipitation particles by the scattering due to large ice particles present in the clouds. These sensors are installed on Earth-orbiting satellites which offer only intermittent coverage of a given region of interest (currently about ten observations per day). Therefore, the estimation of precipitation from proxy parameters such as cloud top temperature that can be in- 25 ferred from geo-stationary observations has been developed. The algorithms based on IR data relate rainfall to cloud top temperature and cloud optical properties through a precipitation index. The indexing method assigns a fixed rain rate to each identified  (Kidd, 2001). This assumption is most effective for convective conditions but can yield crude estimates because of the weak link between cloud properties and precipitation. Current approaches use rain rates estimated from coincident microwave observations to derive regional calibrations of Global-IR techniques (Anagnostou, 2004). However, both kinds of sensors have difficulties in capturing non-convective rainfall and 5 shallow "warm" rain events (Ebert et al., 2007). With the multiple products currently available, it is important to evaluate their precision and uncertainty, as well as their advantages and drawbacks, before opting for a specific application. Several studies have been conducted with the aim to intercompare, against locally observed data, rainfall estimates derived from satellite obser-10 vations. As particular reference is the work achieved by the International Precipitation Working Group (IPWG) (information available online at http://www.isac.cnr.it/ ∼ ipwg/). The project started in 2002 over Australia and United States and an additional verification was done over Europe in 2004. The results showed that PM-IR merged estimates perform about as well as radar in terms of daily precipitation bias and frequency over 15 the United States (Ebert et al., 2007). Such elaborated evaluation has not been undertaken over the African continent, as high quality networks of rain gauges and radars are needed in order to assess the quality of the estimates.
Nevertheless, the Tropical Rainfall Measuring Mission (TRMM) monthly estimates have been validated over major climatic regions in Africa (Adeyewa and Nakamura, 20 2003) showing the sensitivity of random and systematic error components to the seasonal and regional differences. Over West Africa, TRMM-merged product seems to be in excellent agreement with gauge data at monthly time step (Nicholson et al., 2003): the root mean square error is on the order of 1 mm day −1 and the bias is null. from 45 to 60 %, increasing while the time step decreases. FEWS RFE 2.0 performs worse than TRMM 3B42 because of the fixed temperature threshold and fixed rain rate used to compute IR estimates. CMORPH shows superior performance compared to TRMM 3B42. The performance of seven operational global products, including TRMM 3B42, CMORPH, and FEWS RFE 2.0 was also evaluated during West African mon-5 soon at 10-daily time step (Jobard et al., 2011). CMORPH exhibited the worst skills (strong positive bias), TRMM 3B42 displayed a moderate aptitude and FEWS RFE 2.0 the best performance in terms of distribution and bias. The Microwave Infra-Red Algorithm (MIRA) has been compared at daily time scale to ground station data over Southern Africa (Layberry et al., 2006) showing better agreement in the wet months than in 10 the drier ones, but overall quite poor skills for rainfall detection. Over the Okavango basin, a monthly dataset at 0.5 • based on the TRMM and Special Sensor Microwave Imager (SSM/I) datasets was found to overestimate the rainfall by 20 % (Wilk et al., 2006). The comparison of MIRA and FEWS estimates to in situ stations records over the Zambezi Basin at monthly time scale indicated that MIRA often overestimates (up 15 to 50 %) and produces rainfall during dry months whereas FEWS has less bias (Winsemius et al., 2006a). TRMM 3B42RT and CMORPH were evaluated over Ethiopian river basins (Romilly and Gebremichael, 2011;Bitew and Gebremichael, 2011) and CMORPH was found to underestimate rainfall by 11 % whereas TRMM 3B42RT overestimated by 5 %. However, the results varied depending on the geographical region 20 considered.
Regarding the divergent results obtained from the previous studies and the lack of validation at the daily time step, the objective of this paper is to provide a comparison and an evaluation of the different sources of input data that can be used for hydrological modeling of the Zambezi Basin at daily time step. The aim of the analysis is 25 to determine the appropriate size of sub-basins in terms of rainfall pattern and the reliable time step for modeling. Three operational and acknowledged high resolution satellite derived estimates (TRMM 3B42, FEWS RFE 2.0 and CMORPH) are analyzed and compared to ground data for the period from January 2003  The Zambezi river basin ( Fig. 1), located in the South of the African continent, is shared by eight countries, making it a particularly interesting system to further investigate the implementation of IWRM's (Integrated Water Resources Management) principles. From its headwaters in Angola to the delta in Mozambique, the Zambezi River runs over 2600 km and connects eight African nations that share different portions of its

15
The ITCZ is a convective front oscillating along the equator. It moves from 6 • N to 15 • S from July to January and back North from February to June. Associated with it, the peak rainy season occurs during the Southern Hemisphere summer (from October to April) and the winter months are dry. The diurnal cycle of precipitation depends also on the ITCZ. Usually, clouds form in the late morning and early afternoon hours and then 20 by the end of the afternoon, convectional short thunderstorms form and precipitation begins. In this study, the data from 32 Mozambican national meteorological stations collected by the the Regional Administration of Zambezi Water (ARA-Zambeze) and 48 meteorological stations from the Global Summary Of the Day (GSOD) international database were collected, resulting in an unequally distributed dataset over the basin The use of the potential of Zambezi River is currently mainly limited to hydropower production through a series of large impoundments: Kariba Dam between Zambia and Zimbabwe, Kafue hydropower scheme in Zambia and Cahora Bassa Dam in Mozambique.
CMORPH is constructed from similar inputs as those used in TRMM 3B42 with the difference that it does not merge PM and IR rain estimates. At times and locations when PM data are unavailable, it uses the motion vector derived from half-hourly geo-15 stationary satellite IR data to interpolate precipitation (Joyce et al., 2004). Therefore, the analysis does not rely on IR data for direct rainfall estimation. The original product, starting in December 2002, has a very high spatial resolution: 8 km grid and half-hourly time step. However, historical data are available only at a spatial resolution of 0.25 • and at 3-hourly temporal resolution (00:00, 03:00,. . . , 21:00 UTC) in a global belt extending 20 from 60 FEWS RFE is computed by the NOAA/CPC (Herman et al., 1997). Since January 2001, the version 2.0 of the algorithm is used, integrating PM estimates. The data consist of a combination of PM and IR precipitation estimates merged with daily rainfall data from Global Telecommunication System (GTS) records. The spatial resolution 25 corresponds to a 0. high spikes in the precipitation estimates. Thus, the data have to be screened for intensities higher than a certain threshold. The Global Precipitation Climatology Centre (GPCC) full data reanalysis product is based on synoptic weather observation data (SYNOP) and monthly CLIMAT report received near real-time via the World Meteorological Organization (WMO) 5 Global Telecommunication System (GTS) (7000-8000 stations). Additional data from dense national observation networks and global and regional collections complete the database which is the most comprehensive global compilation of monthly precipitation data from in situ observation (Schneider et al., 2008). The processing steps include quality-control, inter-comparison of the data from different sources and interpolation to 10 a regular mesh (0.5 • grid). The Version 4 of the product covers the period 1901 to 2007 at a monthly time step with varying data coverage.
The daily ground station rainfall observations are extracted from the Global Surface Summary of the Day (GSOD) product archived by the National Climatic Data Centre (NCDC) of the NOAA. Historical data are generally available from 1929 to the present.

15
In deriving the Summary of the Day data, a minimum of 4 observations per day must be present. The data are reported and summarized based on coordinated universal time (00:00-23:59 UTC). An extensive automated quality control is applied to correctly "decode" as much of the synoptic data as possible, and to eliminate the random errors.
The ground rainfall data registered at a daily basis on the Mozambican part of the 20 Zambezi basin are collected by the Regional Administration of Zambezi Water (ARA-Zambeze). Table 1 summarizes the different characteristics of the rainfall estimates used.

Methodology
The first part of the analysis is the comparison of the different satellite estimates in 25 order to bring out the similarities and discordances. The spatial distribution of rainfall for the dry season (from May to September) and the wet season (from October to April) is mapped on a grid of 0.25 • (Fig. 2). In addition, the zones of agreement and divergence between the different estimates are illustrated by correlation maps (Fig. 3). The Pearson correlation coefficient (R 2 ) between two time series at the same pixel is used for computation. In view of an application for hydrological modeling, the average size of a rainfall event 5 is assessed for each of the products by calculating temporal correlations (Pearson) between the pixels at different radius, assuming isotropy of the rainfall for the period 2003 to 2009. For each pixel p(i ,j ), the Pearson correlation coefficients are calculated between itself and all the other pixel of the matrix: where k and l are the index of the matrix, varying respectively between 18 • and 26 • longitude and −8 • and −20 • latitude.
The value of the coefficients obtained on a radius around the pixel is averaged, varying the radius from 0.01 • (FEWS FE2.0) or 0.25 • (TRMM and CMORPH) up to 15 6 • latitude/longitude and considering a rectangular mesh.
where i and j define the position of the pixel inside the matrix and r is the number of pixel corresponding to the radius of correlation.
The global mean correlation at each radius (COR r ) is then computed as an average 20 over the whole basin to underline the differences between the estimates and maps are produced for some of the key radius (Fig. 4). The analysis is done at daily, 10-daily and monthly time steps. During the second part of the analysis, the error property of the satellite derived data with reference to point ground station measurements is investigated for the wet  (Table 2). Since there is nearly no rain during the dry season, this period was not taken into account for the performance assessment. The goal of this analysis is to assess the quality of the satellite products and to select the most reliable for the hydrological modeling. The two ground data set, GSOD and ARA are separated for the analysis as they 5 come from different sources and don't cover the same area of the basin. Both sets of data (satellite product versus ground station) are plotted at daily, 10-daily and monthly time steps for the pixels on which at least one ground station is available (Fig. 5). The original grid size is used for each product. As the ground data contain large gaps, only time series with at least 20 continuous daily values have been integrated in the analysis at daily time step. For the 10-daily rainfall accumulation, one day of missing data is accepted in the calculation and for the monthly accumulation, up to 5 days of missing data are accepted. Statistics are calculated for each of the ground stations, weighted by the number of available data per season and a global value of the coefficients is determined by the 15 weighted mean of all stations based on the total number of data per station.
The ability for each of the products to detect rainfall is evaluated by the Probability Of Detection (POD) and the False Alarm Ratio (FAR) indices (Layberry et al., 2006;Stanski et al., 1989;Ebert et al., 2007). For each rainfall threshold associated with the time step, each point is estimated to rain or not. This leads to three outcomes: 20 estimated rain/observed rain (hit h), estimated rain/observed no rain (false alarm f ) and estimated no rain/observed rain (miss m). The indicators are derived from these outcomes:  (7) where "Sat" is the satellite data, "Obs" the ground observed data and "Obs" the mean of the ground observed data. Finally, the pixel to pixel approach is applied to the satellite products in comparison with the GPCC ground data grid, taking into account only the pixels with at least one 10 ground station. The data are compared by means of scatter plots (Fig. 8) and maps of volume ratio (Fig. 9) in order to evaluate the spatial distribution of the satellite precision.

Temporal and spatial repartition of the precipitation
The spatial variation analysis shows a general North-South gradient in the intensity 15 of precipitation (Fig. 2). The TRMM data set registers slightly lower rainfall intensities than the FEWS data set. The region of Lake Malawi, located in North-East side, is characterized by lower rainfall in comparison with the North-West area. The grid pixels above the ocean (South-East corner) reveal lower rainfall than those of the coastal 8183 Introduction areas. CMORPH has the highest spatial variability of rainfall, varying from 300 to 2000 mm year −1 and seems to overestimate the precipitated amount in the North-West region. During the dry season, it displays quite high rainfall intensity over the Kariba Lake area, probably due to some shortcomings in the computing procedure. FEWS reports the lowest rainfall volume and shows the lowest variability of precipitation. TRMM 5 spatial variability is moderate. Although the main characteristics are preserved in all estimates, the rainfall spatial patterns produced by the three algorithms show considerable differences. The global correlation coefficient is 0.54 between TRMM and FEWS data sets, 0.76 between TRMM and CMORPH and 0.60 between FEWS and CMORPH. In terms of 10 spatial repartition (Fig. 3), the area at the North-West corner, the region over Lake Malawi (North-East limit of the basin) and the coast line (South-East corner) show the lowest agreement between data sets. The overall low correlation (R 2 ) between TRMM and FEWS and FEWS and CMORPH is probably due to the difference in the IR-based estimates used in the algorithm. TRMM and CMORPH have the highest global correlation, reflecting that their algorithms are based on the same PM data and can indicate that the IR influence is not very important. The homogeneity of the rainfall was evaluated by the correlation of the time series at each pixel with the surrounding pixels. FEWS exhibits the highest internal correlation (COR r ) different from TRMM and CMORPH which show similar patterns (Fig. 4). At 20 daily time step, FEWS has a mean correlation of 0.5 computed on a radius of 2.25 • and the mean correlation of TRMM and CMORPH decreases rapidly with a correlation of 0.5 on a radius of only 0.75 • . The spatial repartition of the correlation coefficient (COR p(i ,j )r ) is different from one estimate to the other, however, regardless of the product, the central part of the basin seems to be homogeneous and the region over Malawi 25 Lake rather heterogeneous. At 10-daily time step, the 0.7 correlation pattern is similar for all the products: the area over the ocean has the highest heterogeneity along with the regions over Lake Malawi and the upper West corner of the basin. The zones of homogeneity over the 8184 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | delta present at daily time step for TRMM and CMORPH do not appear at 10-daily time step. At monthly time step, the difference between the products for the global correlation is close to zero. In terms of spatial pattern, the area over the ocean is still a heterogeneous zone for all the products. TRMM exhibits a high correlation over the Western 5 part of the basin whereas Kariba Lake is an area of high heterogeneity for CMORPH.

Point to pixel
Based on the scatter plots presented in Fig. 5, it is clear that the time step has an important influence on the quality of the satellite estimates. At daily time step, no direct 10 correlation exists between the satellite estimates and the ground data whereas monthly accumulation comparisons display already a marked trend.
Especially at 10-daily and monthly time steps, the satellite estimates are less correlated with the ARA-Zambeze data than with the GSOD data. FEWS has the lower dispersion as the algorithm uses GSOD data to rescale the satellite estimates. The

15
TRMM product does not appear to be biased but has an important dispersion of the cloud. A strong overestimation is visible on the CMORPH cloud.
The statistics for the ARA-Zambeze data are presented on Fig. 6. All the satellite products reach similar values, except for the volume ratio, for which CMORPH is overestimating the rainfall by about 40 %, TRMM is overestimating the rainfall by about 20 20 % and FEWS is close to 1. CMORPH strong positive bias has already been documented for West Africa (Jobard et al., 2011). However, it seems to be more reliable over Ethiopia where it performs better than TRMM 3B42 and FEWS RFE2.0 at 10-daily time step (Dinku et al., 2007) and underestimates the rainfall by 11 % at daily time step (Romilly and Gebremichael, 2011). As the time step increases, the performance of the 25 estimates also increases (higher POD, IA and R 2 and lower FAR and RRMSE). This is consistent with the results already published in terms of time step effect. For the GSOD data (Fig. 7), the differences between the satellite estimates are more marked. Surprisingly, although TRMM algorithm uses GSOD data to rescale the satellite estimates, it has not the highest performance: at 10-daily and monthly time steps, the POD, FAR, R 2 and IA of FEWS are better. In terms of volume ratio, CMORPH is still showing an overestimation of about 40 % but FEWS and TRMM have similar values 5 close to 1.

Pixel to pixel
The pixel to pixel comparison, carried out for a monthly time step on GPCC's 0.5 • grid (Fig. 8), shows the same trend than the point to pixel analysis. CMORPH is clearly overestimating the rainfall as the cloud is on the left side of the plot and FEWS has the 10 lowest dispersion of the cloud.
Regarding the spatial distribution of the satellite performance ( Fig. 9), the precipitations are overestimated in the South-West corner, especially with CMORPH (volume ratio of about 2). On the contrary, an underestimation (below 0.75) occurs on some pixels over Malawi Lake. For FEWS and TRMM, the major part of the basin has a 15 volume ratio between 0.75 and 1.25.

Conclusions
First, the three satellite estimates were compared. In terms of yearly rainfall, although main characteristics are preserved, the rainfall spatial patterns produced by the three algorithms show considerable differences. CMORPH seems to be highly influenced 20 by Kariba Lake. Regarding the spatial heterogeneity, FEWS pixels are much more inter-correlated than TRMM pixels. For a rainfall homogeneity threshold criterion of 0.5 global mean correlation coefficient, the area of each subbasin should not exceed a circle of 2.5 • latitude/longitude radius for FEWS and a circle of 0.75 • latitude/longitude radius for TRMM and CMORPH considering rectangular mesh. Secondly, the performance of the satellite estimates was assessed by comparisons with ground station. At a daily time scale, the probability of rainfall being detected by the satellite appears nearly equivalent to a random simulation (POD of about 0.6 and FAR of about 0.5). However, they cannot be expected to provide results identical to the gauge measurements at small time steps as both the temporal and the spatial sam-5 plings are different. The gauging stations provide point measurements observed over continuous periods of time, while satellites deliver spatial averages based on intermittent rain rate estimates. Furthermore, satellites produce estimates over a broad area, thus having a tendency to smooth localized phenomena which can substantially affect gauging stations. The reliability of gauging station data is also controversial as the 10 series are often not continuous. At monthly time scale, all estimates have a good correspondence, CMORPH being the less precise in terms of volume ratio, overestimating the rainfall by about 40 %. TRMM 3B42 and FEWS RFE2.0 show a very similar performance compared to ground data even if they are very different in the spatial repartition of the rainfall.

15
Due to the fact that TRMM produces data since 1998, which will increase the number of years available for calibration and validation of the model, it is chosen as the input data for hydrological modeling.
The results presented in this paper underline the fact that rainfall input data have to be studied before modeling the hydrological behavior of a basin in order to know the 20 size of rainfall events and their distribution through space and time. Moreover, they illustrate the very strong dependency of the satellite product quality with the region of interest. An interesting addition to the study would be to calibrate the model with the different possible input data and evaluate the performance in terms of runoff simulation. However, in a basin like the Zambezi one where only about 7 % of the rainfall is 25 contributing to runoff, the influence of other parameters like the wetland capacity, the evaporation and soil equations will be more important.