A comparison of Meteosat rainfall estimation techniques in Kenya

Meteorol. Appl. 8, 107 117 (2001) A comparison of Meteosat rainfall estimation techniques in Kenya M R Tucker and C B Sear, NRI, Medway University Campus, University of Greenwich, Central Avenue, Chatham Maritime, Kent ME4 4TB, UK Two methods for estimating ten-day rainfall totals from Meteosat infra-red imagery were compared for the April June 1996 long rains of Kenya in an area covering the eastern highlands and the Tana and Athi river basins. One of these (the Bristol B4 method) was then used for rainfall estimation for the whole of Kenya, for November 1996 and the other, the TAMSAT Cold Cloud Duration (CCD) method was used to estimate rainfall for the whole of Kenya for November 1997 to April 1998. April June 1996 was an unusual season with very few large rainstorms. For this comparison period the B4 method gave better estimates of actual rainfall than the TAMSAT method because it used a variable cold cloud threshold temperature and ongoing calibration against rain gauge data. Comparison of ten-day CCD totals with rainfall for the 1997 1998 period indicated that using the TAMSAT method gave best rainfall estimates for the arid and semi-arid areas of eastern and northern Kenya and for months other than the main rainy season months of November and April. Both methods could be used successfully to identify periods with well below or well above average rainfall even over highland areas, and they are therefore useful for providing food security early warnings. 1. Introduction Satellite data have been used for rainfall monitoring for many years (Barrett & Martin, 1981) and in particular, the use of geostationary Meteosat infra-red imagery to estimate convective rainfall by monitoring the presence of cold cloud tops is well established (Milford & Dugdale, 1989). Tropical highland areas, especially those with complex topography such as eastern Africa, have presented a considerable challenge to users of satellite rainfall estimation methods because of complex local rainfall variations and the occurrence of non-convective rainfall. The aim of the present work is to compare the ability of two satellite rainfall estimation methods to give useful results for the Kenya highlands. A brief description of some of the methods currently in use is followed by a study using two of the methods for April-June 1996, November 1996, and November 1997 to April 1998. The Cold Cloud Duration (CCD) method of rainfall estimation, used by the TAMSAT (Tropical Applications in Meteorology using Satellite data) Group at the University of Reading (Milford & Dugdale, 1989), assumes both that the significant rainfall in the monitored area is convective and that there is a linear relationship between the length of time that convective clouds are present and the amount of rain that falls. This generally restricts the method to the tropics and subtropics where convective rainfall predominates over frontal or orographic rainfall. The method relies on infra-red imagery from geostationary satellites to provide estimates of cloud-top temperature. The validation of satellite rainfall estimation also presents problems because the difference between satellitederived rainfall estimates for pixels with areas of 25 km 2 or greater and rain-gauge point values means that dense networks of rain-gauge measurements are required for accurate calibration (Flitcroft et al., 1989). In the Sahel of Niger, Flitcroft et al. found that the best pixel estimate is obtained as a weighted average of the gauge value and the climatological mean rain per rain day. This tends to reduce high values of estimated rainfall but increase low ones. Burt et al. (1995) found that in Mali, in the Sahel, the maximum rain gauge separation for reliable area estimates was 8 km. This study also found that the probability of rain from cold cloud increased with decreasing cold cloud temperature from 0.71 for 223 K to 0.93 for 203 K. As high densities of surface rain-gauge data are very rarely available for operational rainfall forecasting, time or spatial averaging of satellite data are often used together with readily available synoptic weather information. For many applications, CCD-based rainfall estimates are made for ten-day periods for each pixel, although for some river catchment studies daily area averages are made for the whole river basin. The TAMSAT method first establishes the optimum cold cloud threshold temperature, which distinguishes rain days from non-rain days for an area (typically about 25 degrees square but depending on topography) for each month. Typical 107

M R Tucker and C B Sear threshold temperatures are 233 K for much of Africa but 223 K or even 213 K for the Sahel of West Africa (based on intervals of 10 K). Then, using this threshold, a simple linear regression between CCD and rainfall amount is calculated for each month for the trial period. The threshold temperature and regression parameters are then taken to be typical for a broad zone for a given month so that they can be used for future rainfall forecasting. This method for estimating rainfall has been used in conjunction with the Meteosat Primary Data User Systems installed widely in Africa and elsewhere under the LARST (Local Application of Remote Sensing Techniques) initiative managed by the Natural Resources Institute (NRI) (Sear et al., 1993). It has also been used with other methods, such as that developed by Bristol University (Barrett, 1993). In eastern Africa the TAMSAT method has been used successfully, together with NOAA vegetation index imagery, in Ethiopia to aid food security programmes (Tadesse et al., 1995) and in Kenya and Tanzania for insect-pest forecasting (Tucker, 1997). However, it has not been considered suitable for rainfall estimation in the Kenya highlands because of the occurrence of orographic rainfall and the complexity of the topography. The University of Bristol Centre for Remote Sensing developed a more complex satellite rainfall estimation method which uses both variable rain/no rain threshold temperatures and variable rain-rates (Barrett et al., 1986). The method uses thresholds that vary from pixel to pixel depending on an ongoing calibration against synoptically reported rainfall, then deploys a weighting for mean rain per rain day at nearby rain gauges, rather than assuming that the association between cold cloud and rainfall is constant (Barrett, 1993). This B4 method has been used for rainfall estimation in the upper Nile River basin, including the Ethiopia highlands (Todd et al., 1995) and has been applied operationally to several other areas of the tropics for several years (Barrett et al., 1996). The Todd et al. study modelled spatial variations in optimum cold cloud rain/no rain thresholds in relation to latitude, longitude and elevation. In July 1991, optimum thresholds varied from 252 K in the eastern Ethiopia highlands to 205 K in the Sudan border area west of the highlands, with thresholds generally decreasing from south-east to north-west. In August, however, elevation appeared to be the dominant factor with warmer thresholds over the highlands of 240 to 250 K. Herman et al. (1997) described a similar method for estimating ten-day African rainfall totals that is used operationally by the USAID/FEWS programme. They started with the GOES Precipitation Index (GPI) developed by Arkin & Meisner (1987) which assumes 3 mm of precipitation for each hour of cloud top temperature below 235 K and then corrected for biases by computing differences from accumulated precipitation 108 determined from rain gauge reports. They also developed a technique for estimating orographic precipitation from warm cloud (cloud top temperatures between 275 and 235 K inclusive). When the low-level wind direction is favourable for orographic lifting, the rainfall rate is estimated using a procedure combining relative humidity, wind direction and the terrain slope. Menz (1997) used two techniques to estimate rainfall in Kenya: the Cold Cloud Threshold method (the result of which he called the Meteosat Precipitation Index MPI) and the Convective Stratiform Technique developed by Adler & Negri (1988). Monthly maps of estimated precipitation for a dry season month (September 1993) and a wet season month (April 1994) were produced. Menz concluded that there was little variation in linear correlation coefficients for precipitation threshold temperatures between 223 K and 285 K with maximum values at 252 K for September and 280 K for April. This contradicts the conclusion of Todd et al. (1995) that variations in threshold temperature are of crucial importance. However, 280 K does not represent cold cloud, though it may distinguish warm stratus from ground surface temperatures, and since the MPI (Cold Cloud) method was developed for convective rainfall, we consider the results presented by Menz to be of doubtful validity. Menz also used the convective stratiform technique, with a threshold temperature of 285 K for convective rainfall and 230 K for stratiform rainfall (from high level stratus). This, he found, improved monthly rainfall estimates for the Kenya highlands and coast but gave a zero correlation for the semi-arid lowlands. This is not surprising given the high cloud top temperatures. Menz made no attempt to estimate precipitation for periods of less than a month. Research has been carried out on a potentially useful method of estimating rainfall from warm clouds in Eritrea by correlating the negative deviation of the pixel temperature from a rainfall threshold temperature to rainfall amount, rather than just using the CCD. However, this method has not yet been proved operationally (Van Buskirk, pers. comm., 1998). The aim of the present work is to compare the Bristol B4 method and the TAMSAT method in estimating rainfall in the Tana and Athi river basins of east-central Kenya. These river basins extend from the Aberdare mountains of the Central Highlands to the Indian Ocean coast (Figure 1). The work was then extended to the rest of Kenya (Figure 2) to investigate further the spatial and temporal variation in the ability to these methods to estimate rainfall from Meteosat imagery. 2. Data and methods Rainfall data for the long-rains period of April-June 1996 were obtained from the Kenya Meteorological

Meteosat rainfall estimation in Kenya Figure 1. Location of the Tana and Athi river catchments and the Kenya rainfall stations used in the April June 1996 study. Figure 2. Location of the Kenya meteorological stations used in the November 1997 to April 1998 study (the area covered by Figure 1 is delimited). Department for 28 stations in east-central Kenya (Figure 1). These included both synoptic meteorological stations and other climatological stations. Rainfall data for the short rains month of November 1996 and also November 1997 to April 1998 (which includes both rains and the January to March dry season) were then obtained for 27 Kenya synoptic meteorological stations from the NOAA web site. The location of stations for which data were complete (or nearly so) for the period November 1997 to April 1998 is shown in Figure 2. Climatological mean data (1951 1980) were obtained from data published by the Kenya Meteorological Department in 1984. Bristol B4 Meteosat-based estimates of rainfall were provided by the University of Bristol Centre for Remote Sensing (Barrett, pers. comm., 1996, 1997). No prior training of the B4 method was undertaken before the generation of the test results, which were provided in the form of hard copy maps for April and May 1996 and for November 1996. For April June 1996, tabulated daily and dekadal (ten-daily) rainfall estimates for each rainfall station and monthly analyses were also provided. TAMSAT Meteosat-based dekadal rainfall estimates and analyses were provided by the University of Reading TAMSAT Group for April June 1996, excluding the first dekad (ten-day period) of April (Grimes, pers. comm., 1997). The parameters used for this estimation were based on calibration for Kenya and surrounding countries for the periods 1991 1995 (for April) and 1985 1995 (for May and June). Meteosat imagery was collected for Kenya for the period November 1997 to April 1998 and daily CCD imagery was generated using a Bradford University Remote Sensing (BURS) Ltd. Primary Data User System and NRI Meteosat Operation Manager software. CCD images were archived for three cold cloud thresholds (243 K, 233 K and 223 K). Both the TAM- SAT analyses for April June 1996 and trials with the 1997 1998 data indicated that the 243 K threshold generally gave better differentiation between rain and no rain days for Kenya than did other, colder thresholds. This threshold was therefore used in the present analyses except for February 1998 for which month 233 K gave better rain/no rain discrimination. Meteosat images were converted for analysis in an image analysis package. The locations of Kenya synoptic meteorological stations were entered into the package. Meteosat is above the Greenwich Meridian and a correction for parallax was needed to examine high cloud tops above Kenya at 35 to 40 E, which would otherwise appear to be located at the Earth s surface further east than their actual locations. On the assumption that cold cloud tops were 10 km above ground, 0.8 degrees was added to station longitudes. No correction was made for latitude because Kenya is on the Equator and Meteosat is above the equator. The pixel value of CCD was extracted automatically for each day of a month for each meteorological station. Output text values were converted into spreadsheets for analysis by comparison with rainfall station data. Some CCD data were missing for January and March so there was a total of 14 complete dekads from November 1997 to April 1998. From the 1996 analyses (see below) an hypothesis was formulated that, in order to use CCD to estimate rainfall in a topographically variable area, variation between locations is more important than variation between months. However, while B4 uses a variable cold cloud threshold to account for spatial variation, we decided to use a constant threshold and investigate variations in the slope and intercept of the regression equation. The 11 dekadal totals of CCD and gauge 109

M R Tucker and C B Sear rainfall for the period November 1997 to March 1998 were combined in a single spreadsheet and compared for each rainfall station separately, for the whole period. We made the assumption that each station was representative of the topographic and geographical area in which it was located. A linear best fit was used to estimate April rainfall from CCD and these rainfall estimates compared with rain gauge data. 3. Results 3.1. April-June 1996 The 1996 Kenya long rains were well below average. For the gauges used in the analysis, the April rainfall was only 34% of the long-term mean and the May rainfall was 74% of the mean. On average, June is a dry month except in western Kenya, but in 1996 it was wetter than usual (87% above average). This did not compensate for the earlier poor rains (Figure 3). In the TAMSAT analyses (23 rainfall stations) a total of 12 out of 69 station-dekads had no rain in April, 10 in May and 22 in June. The Bristol B4 method analysis correctly allocated 84% of the dekads to rain or no rain categories while the TAMSAT method over-estimated the number of dry dekads and only allocated 53% of dekads correctly (Table 1). This indicates that the 243 K cold cloud temperature threshold used in the TAMSAT analyses was too cold to discriminate between raining and non-raining cloud, while the individual pixel calibration method used by Bristol was able to give reasonable rain/no rain differentiation. The TAMSAT Group does not archive CCD images for temperatures above 243 K because tropical convective rainstorms usually have cloud-tops below this temperature (Grimes, pers. comm., 1998). Regressions were calculated for monthly and dekadal totals for B4 rainfall estimates against reported rain gauge data (Figure 4). The proportions of variance explained (R 2 ) were between 0.36 and 0.58 for monthly totals and between 0.24 and 0.59 for dekads. The slopes of the regression lines were between 0.34 and 0.77. This indicates a tendency for higher rainfall totals to be underestimated, a characteristic of most satellite rainfall estimation techniques, both because of the difference between pixel and rain gauge estimates discussed earlier and because empirical calibration methods often miss the infrequent, high intensity rainfall events. Because of the large number of zero estimates, a regression of TAMSAT estimates against gauge data was possible only for all three months combined and only for dekads with non-zero cold cloud. This gave the equation: Rain = 2.61 CCD + 14.3 where Rain is in mm and CCD is in hours. The overall proportion of variance explained was 0.53, indicating that some prediction of non-zero rainfall totals would be possible. In summary, for the period April-June 1996 both Bristol B4 and TAMSAT CCD methods under-estimated reported rainfall but B4 performed significantly better than the TAMSAT method. 3.2. November 1996 B4 rainfall estimates were also tested against rain gauge data for November 1996. For this period rainfall data were only available for 17 stations within the central Kenya area covered by the satellite-derived rainfall estimates. For most of these stations, November 1996 rainfall was close to the long-term climatic mean (Figure 3). However, at Makindu, 336 mm was reported on one day, which massively exceeds the climatological maximum 24 hour fall for November of 112.5 mm. We assumed, therefore, that it was erroneously reported and replaced it with 33.6 mm (the most likely correct value). Table 1. Percentage of dry/wet dekads and correlation coefficients for dekadal rainfall estimates by Bristol B4 and TAMSAT CCD methods for 1996 Technique Statistic April May June November Bristol Percentage of correct dry/wet dekads 84 92 Correlation coefficient of estimate 0.74 0.49 0.77 0.57* (0.36) Number of stations 28 28 28 15 TAMSAT Percentage of correct dry/wet dekads 53 Correlation coefficient of estimate 0.73 # Number of stations (excluding cost) 23 23 23 * Excluding exceptional Makindu rainfall ( ) Including exceptional Makindu rainfall # Excluding zero estimates 110

Meteosat rainfall estimation in Kenya Figure 3. Monthly Kenya rainfall compared with climatic mean rainfall for (a) April, (b) May, (c) June and (d) November 1996. The B4 method allocated 92% of dekads correctly as rain or no rain. The linear regressions of estimates versus gauge data for monthly total rainfall and for all dekadal rainfall estimates for this month (Figures 4(d) and 4(h)) had slopes of 0.40 and 0.28 respectively with low rainfall over-estimated and high rainfall under-estimated. R 2 were 0.47 and 0.32 for monthly totals and dekads respectively. Monthly November rainfall gauge totals were well correlated with the climatic mean (R 2 = 0.77). Somewhat surprisingly, the B4 monthly estimates were more highly correlated with mean values (R 2 = 0.67) than with the rain gauge totals (R 2 = 0.48). This suggests that, at least in this month, the B4 method was better at distinguishing between different stations than measuring differences between actual and mean rainfall. 3.3. November 1997 to April 1998 For the period November 1997 to March 1998, monthly and dekadal CCD totals were compared with rainfall station data and the linear best-fit lines used to attempt to forecast April rainfall. Monthly rainfall from November 1997 to February 1998 was well above the long-term climatic mean. March 1998 was below average and April 1998 was generally below average (Figure 5 shows the comparison for those months with complete CCD data). Monthly total CCDs were highly correlated with monthly rainfall for December (R 2 = 0.61), moderately so for February (R 2 = 0.39) but not for either November or April, which are usually the wettest months (Figure 6). Taking dekadal totals for each month (Figure 7), the correlation between CCD and station rainfall was low except for February (R 2 = 0.50). This suggests that the CCD estimation method is less useful in the middle of the short and long rains. However, December 1997 rainfall totals were higher than those in April 1998 for many stations, so the lower correlation in April was more likely to be related to the type of rainfall system rather than the amount of rainfall. Taking each rainfall station separately to investigate the spatial variation in the relationship between dekadal rainfall and CCD, we find R 2 for the whole November to March period taken together were variable (Table 2). High correlations (R 2 > 0.49) were found for Lodwar and Moyale in northern Kenya, Eldoret and Kericho in western Kenya, Narok in the Rift Valley and Makindu and Voi in south-east Kenya. Very low correlations were also found, especially for Kakamega in western Kenya, Nairobi in east-central Kenya, Garissa in eastern Kenya and Mombasa on the coast. Stations with 111

M R Tucker and C B Sear Figure 4. Bristol B4 monthly rainfall estimates versus rain gauge totals for (a) April, (b) May, (c) June and (d) November 1996. (e), (f), (g), (h) as (a), (b), (c), (d) but for B4 dekadal rainfall estimates. 112

Meteosat rainfall estimation in Kenya Figure 5. Monthly Kenya rainfall compared with climatic mean rainfall for (a) November 1997, (b) December 1997, (c) February 1998 and (d) April 1998. Table 2. Correlation coefficient of dekadal rainfall versus CCD for November 1997 to March 1998 for individual Kenya meteorological stations Station Correlation coefficient Lodwar 0.985 Moyale 0.948 Kitale 0.518 Eldoret 0.910 Kakamega 0.214 Meru 0.562 Kisumu 0.383 Kericho 0.816 Nakuru 0.362 Embu 0.358 Garissa 0.128 Narok 0.852 Nairobi JKA 0.175 Nairobi Dag 0.372 Nairobi WILS 0.009 Makindu 0.788 Voi 0.706 Mombasa 0.086 moderate or high correlations were grouped by region to provide a combined regression model (Figure 7). The resulting equations are: Northern Kenya Rain = 1.63 CCD + 1.3 (R 2 = 0.89) South-east Kenya Rain = 2.39 CCD + 17.0 (R 2 = 0.53) Western Kenya Rain = 0.74 CCD + 14.8 (R 2 = 0.49) Kenya Rift Valley Rain = 1.18 CCD + 21.9 (R 2 = 0.47) We did not consider it possible to derive estimation equations either for the eastern highlands of Kenya (including Nairobi, Meru and most of the upper catchment of the Tana and Athi Rivers) or for the Kenya coast. The best-fit regression lines suggest that when cold cloud was present in northern and south-east Kenya, rainfall increased rapidly with CCD, while in western Kenya, rainfall increased more slowly with CCD. This reflects the fact that the November March period includes the short rains which are often the dominant rains in eastern Kenya but only the beginning of the long rains, which are more important in western areas. We applied the November March regression results to 113

M R Tucker and C B Sear Figure 6. Monthly Kenya rainfall totals versus monthly total CCD for (a) November 1997 (b) December 1997 (c) February 1998 and (d) April 1998. (e), (f), (g), (h) as (a), (b), (c), (d) but for dekads. 114

Meteosat rainfall estimation in Kenya Figure 7. November 1997 to March 1998 dekadal rainfall versus CCD for (a) Northern Kenya (b) Southeast Kenya (c) Western Kenya and (d) Kenyan Rift Valley. April 1998 dekadal CCD to give estimated rainfall for stations in the four areas (see Figure 8). The correlation between estimated and gauge rainfall was low (R 2 = 0.05) indicating that the November March derived regressions could not be used to estimate April 1998 rainfall. For comparison, the equation derived by TAMSAT for non-zero dekads for April June 1996 was also used to predict April 1998 rainfall but the correlation was no better (R 2 = 0.04). As noted above (Figure 6) the correlation between dekadal CCD and rainfall for April 1998 was also low (R 2 = 0.13) which may explain these poor predictions. 4. Discussion Our results show that there is considerable variation in both space and time in the relationship between cold cloud duration and rainfall over Kenya. The poor rain/no rain discrimination shown by the 243 K threshold in April June 1996 indicates that significant amounts of rain fell from warmer clouds during this period. The Bristol B4 variable threshold method was able to take this into account (although untrained from previous data) and gave good dekadal rainfall estimates for April and June but less good estimates for May and November 1996. For the period November 1997 to April 1998 there was no correlation between monthly total CCD and rainfall for the main rainy season months of November and April but good correlations at the end of the short rains in December and for the usually dry month of February. Considering the different regions of Kenya, correlations were high for the arid or semi-arid regions of northern Kenya, south-eastern Kenya and the Rift Valley but low for the more humid western Kenya and very low for the eastern highlands from Nairobi northwards and for the coast. This indicates that the coldcloud threshold method works best in climatologically drier areas in spite of the fact that western Kenya (where the main rains are from April to June) was drier than the other areas during the November March period. The TAMSAT CCD method has previously been shown to work well in semi-arid areas such as the Sahel and is used in Ethiopia (Tadesse et al., 1995). However, it would need further modification to be of regular operational use for the Kenya highlands. The complexity of the relationship between rainfall, relief and location in eastern Africa suggests that multiple linear regressions, as used by Herman et al. (1997) may not give much improvement. Milford & Dugdale (1989) stated the fundamental limits of the methods using 115

M R Tucker and C B Sear (a) (b) Figure 8. Dekadal Meteosat rainfall estimates versus rain gauge totals for April 1998. (a) Estimate based on November 1997 to March 1998 data. (b) Estimate based on April May 1996 data. cold cloud statistics should be recognised before effort is spent on fruitless attempts at refinement. The strengths of the method include its ability to give real time estimates over wide areas, particularly for discriminating between days with rain above and below fairly small rainfall thresholds. Thus, the basic CCD method is perhaps best used for identifying periods of wellabove or well-below average rainfall rather than trying to estimate actual rainfall amounts, provided that a suitable CCD threshold is used. Its operational use, together with NOAA-based vegetation monitoring to identify possible crop failure and subsequent food security problems, has proved valuable (Tadesse et al., 1995). Daily rainfall estimation for small catchment hydrological forecasting in topographically complex areas is probably beyond the method at present. This said, the TAMSAT Group is currently refining and improving their CCD methodology for non-operational use by enhancing, via Kriging, the incorporation of rain gauge data into satellite-derived rainfall estimates (Grimes et al., 1999). The Bristol B4 method, using flexible thresholds and rain-rate with continuous re-calibration against available rain gauge data, gave better estimations for this area over our test period. Further development of this technique is being undertaken to incorporate rain-rate evidence from passive microwave imagery from the US DMSP satellite family (Todd, pers. comm. 1999), which will help improve performance further, especially in areas and/or periods with no rain gauge reports. 5. Conclusions 116 This comparison between two methods of rainfall estimation for the Kenya highlands and the Tana and Athi river basins indicates that the Bristol B4 method gives better results than a simple Cold Cloud Duration method for the three months for which direct comparisons were possible. B4 estimates were, however, only available to this study for four months in 1996 and further tests are needed for other years and months to draw firm conclusions about the ability of the B4 method to give consistently reliable estimates of rainfall in the Kenya highlands. The B4 method requires the regular availability of rain gauge data for continuous calibration and these can be difficult to obtain quickly enough in Africa. The TAMSAT method does not have this disadvantage. The analysis of the period November 1997 to April 1998 indicates that correlations between CCD and rainfall (the basis of the TAMSAT method) are highest for semi-arid areas and for months in which rainfall is likely to fall as isolated convective showers rather than more continuously. The three months for which direct comparison between the B4 method and the TAMSAT method was possible in this study were unusually dry and atypical in that rain appeared to fall from warm clouds rather than vigorous convective storms. Therefore, further comparisons between the two methods should be carried out for other months and years. Acknowledgements We would like to thank Eric Barrett and Paul Brown of the University of Bristol and David Grimes and Virginia Thorne of the University of Reading for their inputs to this paper. We would also like to thank colleagues at NRI, University of Greenwich, especially Ken Campbell and Richard Pole. References Adler, R. F. & Negri, A. J. (1988). A satellite infrared technique to estimate tropical convective and stratiform rainfall. J. Appl. Meteorol., 27: 30 51. Arkin, P. A. & Meisner, B. N. (1987). The relationship

Meteosat rainfall estimation in Kenya between large-scale convective rainfall and cold cloud over the Western Hemisphere during 1982 84. Mon. Wea. Rev., 115: 51 74. Barrett, E.C. (1993). Satellite rainfall monitoring for agrometeorology: operational problems, practices and prospects. EARSeL Advances in Remote Sensing, 2: 66 72. Barrett, E. C., D Souza, G. & Power, C. H. (1986). Bristol techniques for the use of satellite data in raincloud and rainfall monitoring. J. British Interplanetary Soc., 39: 517 526. Barrett, E. C. & Martin, D. W. (1981). The Use of Satellite Data in Rainfall Monitoring. Academic Press, London, 340 pp. Burt, P. J. A., Colvin, J. & Smith, S. M. (1995). Remote sensing of rainfall by satellite as an aid to Oedaleus senegalensis (Orthoptera: Acrididae) control in the Sahel. Bull. Entomol. Res., 85: 455 462. Flitcroft, I. D., Milford, J. R. & Dugdale, G. (1989). Relating point to area average rainfall in semi-arid West Africa and the implications for rainfall estimates derived from satellite data. J. Appl. Meteorol., 28: 252 266. Grimes, D. I. F., Pardo-Iguzquiza, E. & Bonifacio, R. (1999). Optimal areal rainfall estimation using raingauges and satellite data. J. Hydrology. In press. Herman, A., Kumar, B., Arkin, P. A. & Kousky, J. V. (1997). Objectively determined 10 day African rainfall estimates created for famine early warning systems. Int. J. Remote Sensing, 18: 2147 2159. Menz, G. (1997). Regionalization of precipitation models in East Africa using Meteosat data. Int. J. Climatol., 17: 1011 1027. Milford, J. R. & Dugdale, G. (1989). Estimation of rainfall using geostationary satellite data. In Applications of Remote Sensing in Agriculture. Proceedings of 48th Easter School in Agricultural Science. University of Nottingham. July 1989. Butterworth, London. Sear, C. B., Williams, J. B., Trigg, S., Wooster, M. & Navarro, P. (1993). Operating satellite receiving stations for direct environmental monitoring in developing countries. In Proceedings of the International Symposium: Operationalisation of Remote Sensing, April 1993, ITC, Enschede, The Netherlands, 8: 293 302. Tadesse, T., Sear, C. B., Dinku, T. & Flasse, S. P. (1995). The impact of direct reception of satellite data on an African meteorological service: operational use of NOAA AVHRR and Meteosat products in Ethiopia. In Proceedings of the Eumetsat Meteorological Satellite Data Users Conference, Polar Orbiting Systems, Winchester UK, September 1995, 485 489. Todd, M. C., Barrett, E. C. & Beaumont, M. J. (1995). Satellite identification of rain days over the upper Nile River basin using an optimum infrared rain/no-rain threshold temperature model. J. Appl. Meteorol., 34: 2600 2611. Tucker, M. R. (1997). Satellite-derived rainstorm distribution as an aid to forecasting African armyworm outbreaks. Weather, 52: 204 212. 117