Characteristics of different precipitation measurements in a tropical mountain valley in southern Ecuador are compared in this study to determine potential errors. The instruments are used for different ecological purposes like erosion studies, through fall measurements, investigation of atmospheric chemistry, and modeling of area rainfall distribution. Five recording devices (two precipitation radars, an electro-optical present weather sensor, and two tipping buckets) and three totaling gauges were operated in parallel at a designated site. Data were taken between 1998 and 2003 with different temporal resolution and different operational periods. The general agreement between the instruments is rather good; deviations are in the expected range of 10%–20% of the annual total of about 2200 mm. The remote sensing devices are superior in registering the frequent occurrence of light rain but are not capable of detecting the full range of rain intensities observed. The tipping buckets and the totaling gauges are reliable instruments, but a certain fraction of light drizzle and wind-driven rain is not detected. The present weather sensor has the widest range of sensitivity and supplies additional information on drop spectra. All datasets are affected by operational problems (interruptions, synchronization errors); hence, the redundancy given here seems reasonable for an ecosystem study.
The current study is part of an interdisciplinary ecological research program carried out in the tropical high mountains of southern Ecuador (Beck and Müller-Hohenstein 2001; Bendix et al. 2004a, b). There is a great demand for precipitation data in different temporal and spatial resolutions by ecologists, hydrologists, and soil scientists involved in the program. Several groups are operating different types of standard rain gauges as well as more sophisticated instrumentation. To enable a realistic estimate of rainwater input into the forest ecosystem, a comparison of the performance of all systems (point measurements and rain radars) is executed at a special point (supersite) in order to estimate the scatter from different technologies. The resulting correction factors will be used to convert measured rainfall to an area-specific reference value.
Rainfall is one of the dominant factors that controls ecosystem dynamics like water and energy fluxes, landslide activity, and vegetation zonation in tropical montane forests. However, accurate measurement of rainfall in time and space is a demanding task, even in lowlands. In remote areas of tropical montane forests with their complex terrain and precipitation gradients on the small and medium scale (e.g., Sklenář and Lægaard 2003) the situation is much more complicated. To account for the high spatial variability of rainfall and to obtain parallel sets of data from inside and outside the forest as required for ecological research projects, many gauging stations have to be installed, even in small study areas. This demand is frequently accompanied by budget limitations and a lack of fixed power supplies, which normally does not allow the use of advanced sensor technology.
The importance of scavenging of wind-driven rain and cloud water in these ecosystems has been expressed in many studies (e.g., Barry 1992; Bruijnzeel 2001). A deeper analysis of wind-driven rain is part of our work, but due to the complexity of the problem it is beyond the scope of the current paper. All sensors compared in this paper, including the remote sensing devices, only detect falling rain.
Tipping-bucket or totalizing rain gauges are standard instruments for point measurements at single sites. Such observations are characterized by several problems, such as evaporation losses during light rain, underestimation of high intensities, or errors due to high wind speed. Most of the well-known drawbacks can be corrected to a certain degree if additional data (e.g., wind speed) are available [for problems and solutions refer to Sevruk (1981), Groisman and Legates (1994), and Serra et al. (2001)]. More sophisticated direct and indirect methods, such as automated weighing bucket gauges, disdrometers (acoustic and optical types), and present weather sensors have been developed to overcome problems of the standard rain collecting systems, partly using assumptions [like the type of drop size distribution (DSD), etc.] to derive rain rates. Normally all instruments reveal similar results within a certain range of scatter (∼±10%) but have sensor-specific advantages or disadvantages, mostly dependent on prevailing rainfall conditions (refer, e.g., to Nystuen 1999).
However, point measurements with any of these instruments may give a poor representation of the spatial distribution of rainfall, especially in regions with predominately convective regimes and complex terrain like tropical montane forests. One solution, which is only applicable to limited-area research, is simply to increase the density of the rain gauge network, to register and minimize the effect of spatial heterogeneity (e.g., Yoo and Ha 2002; Changnon 2002; Dalitz et al. 2004). For larger regions as well as high mountain areas, sophisticated interpolation techniques can help to a certain degree, especially for climatological purposes (Morissey et al. 1995; Bendix and Bendix 1998; Frei and Schär 1998; McCullom and Krajewski 1998). However, errors tend to increase with increasing complexity of terrain and rainfall conditions, and most approaches need a larger station network to give satisfying results (e.g., Daly et al. 1994).
Rain radars can generally help to overcome the problem of spatial representation, even in high mountains under convective precipitation regimes (e.g., Hagen et al. 2000). However, these indirect measurements reveal problems resulting from the conversion of radar reflectivity to rain rate. Various other sources of errors exist, like ground clutter contamination of the signal in complex terrain and uncertainties in the calibration (refer to Austin and Wickham 1995; Bringi and Chandrasekar 2001; Andrieu et al. 1997). Also, rain radars do not measure the rain rate at the ground level but at certain altitudes, depending on the elevation angle and the distance from the radar site. Moreover, the achievable spatial resolution of X-band radars lies at the upper limit appropriate for limited-area ecological studies.
Over the last decades, several techniques of rainfall retrieval have been developed based on optical sensors and/or active and passive microwave instruments [a good overview is given by Levizzani (2003)]. However, some techniques still show large deviations to ground-based measurements (Masunaga et al. 2002), while others lack a very poor temporal and spatial resolution (especially microwave sensors on polar-orbiting platforms). Although some techniques could successfully be used for tropical high mountains (Bendix 2000; Bendix et al. 2003a, b), their spatial resolution is still inadequate for limited-area ecological research.
2. Material and methods
a. Study area
The study area is located in southern Ecuador (3°58′S, 79°05′W) ranging between 1800 and 3200 m above sea level (ASL) and covering about 1100 ha. The study site of the joint research group is the Reserva Bíologica de San Francisco (Fig. 1; refer also to Beck and Müller-Hohenstein 2001). Prototypes of all point measurements considered here (and partly used elsewhere in the study area along the altitudinal gradient between 1800 and 3200 m ASL) were installed at one supersite close to the Estación Científica San Francisco (ECSF) research station, where they are operating in parallel, along with the main automatic climate station (see Fig. 2). Besides commercially available samplers, two types of self-constructed samplers are used as well, due to a limited budget. The supersite is situated in a clearing on a ridge just above the valley bottom. Forest margins are some 100 m away, but wind speeds are generally low (daily mean about 1.4 m s−1 with a standard deviation of 0.6 m s−1; Richter 2003); hence, no considerable effect of wind is expected. Average annual rainfall is about 2200 mm, as determined by the longest available dataset (1998–2005). Rain intensities are generally low with 73% of all rain rates between 0.1 and 1 mm h−1, because the main type of precipitation is from advective clouds, which are orographically enhanced. The interannual variability of precipitation is low. In the 5 yr observed so far, the maximum annual total was 2473 mm and the minimum was 2171 mm. The seasonality is relatively constant too. The main rainy season is from April to July but all months of the year receive more than 100 mm. The greatest variability occurs in the weak dry season in November, but no month exceeds a standard deviation of 50% of the 5-yr average (1998–2003). The periods considered for this study did not depart significantly from these average values. Not even the final phase of the El Niño event of 1998 caused notable changes in the annual cycle (Rollenbeck et al. 2005).
The basic system established in 1998 was an automatic climate station (refer to Richter 2003). This station is equipped with a standard tipping-bucket gauge of 200 cm2 collecting surface, called “THIES-Ombrometer” by the manufacturer and widely used by the German weather service (referred to as THIES below). The attached datalogger records hourly totals of the registered rainfall with a nominal resolution of 0.1 mm.
Other measurements are provided by the soil scientific and hydrological research group (Wilcke et al. 2001) using a self-constructed totaling gauge of 104 cm2 collecting surface and a plastic ball mechanism to reduce evaporation losses from the reservoir (Fig. 3, left; referred to as “104” below). The funnel holds a 0.5-mm-mesh-wide filter sieve to prevent contamination by insects and debris. This type of gauge was used to determine subcanopy rainfall and chemical characteristics of rainwater at six sites in different catchments. Five of these instruments were installed at the supersite for comparison purposes and were sampled weekly. The average of the five gauges is used to determine the weekly precipitation totals for this type of gauge.
In addition to these collectors, five totaling gauges were installed for comparison purposes by the plant ecology group (Dalitz et al. 2004). They were used to study spatial heterogeneity of throughfall in relation to canopy structure (149 collectors in 12 forest plots; Oesker et al. 2005). The design had to follow the need for chemical analysis of collected water; hence, they were entirely made of polyethylene (PE) and polyvinylchloride (PVC) and offer a collecting surface of 198 cm2 (Fig. 3, right; referred to as “198” below). Measurements were maintained throughout September 2001–November 2002 with a weekly collection interval.
The last additions to the station network are more sophisticated sensors, which were introduced in March 2002 by one of the climatological groups [precipitation dynamics and chemical properties subproject (PREDICT); Rollenbeck et al. 2006]. At the reference climate station a scatterometer, an electro-optical present weather sensor of the nephelometer type, and a precipitation monitor were installed. A further totaling gauge, with a 314-cm2 collecting surface (Fig. 3, center; referred to as “314” below) provides weekly amounts of rainfall but is mainly used for chemical analyses of the rainwater.
The precipitation monitor Niederschlagsmonitor (NMO) was installed as an automatic sampling and real-time monitoring device for wet deposition. It is equipped with a tipping bucket of 0.1-mm resolution and registers 5-min totals.
The scatterometer offers not only high-resolution rain rates (down to 0.036 mm h−1), but also determines drop size distribution and a velocity matrix of the observed precipitation (BIRAL 2000).
On top of the highest summit in the area (Cerro de Consuelo, 3180 m ASL), an X-band weather radar [Local Area Weather Radar (LAWR)] that supplies data for the extended study area (120 km × 120 km) was installed. The LAWR is a cost-effective system based on ship radar technology (Furuno ship radar). After applying a specific calibration to the raw data, the system provides images of rain distribution every 5 min [for the system, refer to Jensen (2001, 2002); for calibration to Rollenbeck and Bendix (2006)]. The resolution of each radar pixel is 500 m × 500 m; hence, the whole study area is covered by about 53 pixels. The vertical beamwidth of the radar is rather large with 20° and the elevation angle is 0°. The radar rainfall of the supersite pixel is used for the comparison in this study and represents the volume from 500 to 1900 m above that point.
From 2001 to 2003, a vertical Doppler K-band rain profiler [METEK Micro Rain Radar (MRR-2)] was operated at the research station some 400 m off the reference climate station (Bendix et al. 2006). This device offers very high resolution data down to 0.009 mm h−1 sampling every 10 s (Klugmann et al. 1996; Löffler-Mang et al. 1999; Peters et al. 2002). The MRR-2 offers 30 altitudinal range bands of 269-m altitude, reaching a top altitude of about 8000 m. Fourteen months of useable data were obtained from this device, and the near-surface level was used for comparison.
Table 1 summarizes the technical characteristics of the different precipitation measurements used in this study.
b. Evaluation scheme
The comparison of the different sensors uses the individual overlapping observation periods with the THIES rain gauge. The recording devices delivered one complete year of usable data (April 2002–April 2003) whereas the totaling gauges were evaluated on the basis of their different operation periods (see Fig. 4).
In a first step of the comparison, all datasets had to be transformed to the appropriate scale (mm h−1 for the recording devices and mm month−1 for the totaling gauges) and different temporal resampling procedures had to be applied.
The 10-s and 5-min samples were aggregated to hourly totals, which is the basis for the comparison of the recording devices (THIES, scatterometer, NMO, LAWR, and MRR).
The weekly samples of the totaling gauges (104, 198, and 314) were not collected synchronously; sometimes weekly intervals were shifted for operational reasons. Thus, a heterogeneous dataset of different accumulation times resulted. To enable correlation statistics, the hourly data of the THIES rain gauge were grouped and integrated into the variable-length accumulation periods of each of the totaling gauges individually. The weekly data were resampled to monthly totals by splitting the total of transitional weeks (weeks that include the end of the old and the beginning of the new month), according to the number of days in that week belonging to the respective month. Because the datasets are covering different periods, the overall average of all instruments was taken from the period April 2002–April 2003 (vertical lines in Fig. 4). No correction was applied for potential wind errors, because wind speeds are rather low at this site.
After applying this preprocessing, straightforward descriptive statistics and linear regression at different temporal scales (hourly, daily, and monthly accumulations as well as averaged rain rates) were used to examine the systematic bias in the relation between the different instruments. Slope and intercept were determined to assess sensitivity for different rain intensities.
Since frequencies of rain rates show a skewed distribution, correlations tend to be unstable. To gain further insight into the causes of the deviations, the cumulative frequency distribution of the observed rain rates was examined. Distribution functions show the sensitivity for certain rain rates. The cumulative form allows a better distinction of the different datasets.
Each rain-rate class contributes a certain fraction to the rainfall total. Registered totals are more affected by the less frequent higher rain rates, although their frequency is lower. This can be expressed by multiplying the frequency distribution by the value of each rain-rate class. These normalized curves then show how much of the total is registered at a certain rain rate and reflect the capability of each device to cover the actual range of rainfall dynamics.
Finally, a corrected dataset is presented, which serves as a best estimate of the actual monthly rain totals for the period considered here (May 2000–May 2003) and enables us to derive correction factors for other sites of the research program.
3. Results and discussion
For the comparison analysis, the standard tipping bucket (THIES) was chosen as the independent variable because it was calibrated every 2 months, by determination of the precise volume of the tipping bucket. All instruments were maintained and cleaned weekly, as far as required. In general, all measurements show good agreement as should be expected for data obtained within a range of a few meters of each other. All correlations are highly significant (p < 0.01), but the required normal distribution applies to the monthly totals only.
An overview of descriptive statistics for the recording devices is given in Table 2. With the exception of the MRR, all datasets cover the same period. The MRR was used as a mobile measuring unit; hence several interruptions occurred during the considered time period. The rain-rate range (mean, minimum, and maximum) is almost equal for the THIES precipitation sensor, the NMO, and the scatterometer. The LAWR and the MRR show limited maximum values, pointing to a systematic error (see below). The average rain rate should be the same for all instruments, and variations are possibly due to varying sensitivity. Because of intermittent interruptions in the operation of the devices, only 67% of the complete period could be evaluated [n(100%) = 9857]. The total and n given in Table 2 hence represents only a fraction of the actual rain total of the observed period.
Figure 5a presents the comparison of the indirect measurements and NMO data with the THIES standard gauge. The largest differences result from measurement with a different spatial focus. While the MRR rain profiler measures altitudinal bins with a vertical extension of 269 m each, the LAWR produces averages of 0.25 km2 within an altitudinal range of approximately 500–1900 m above the climate station. This results in a blur effect of peak rain rates for the radar datasets, because the volume average is always lower than the peak values within a heterogeneous rain field. The re-evaporation of falling rain in the near-surface layer may introduce further scatter in the relation. The MRR dataset is also affected by the strong spatial gradient of the research site, which gives significantly lower values for the valley bottom some 400 m away from the supersite, where this instrument was located.
The scatterometer shows a higher sensitivity than the tipping bucket because it is able to detect rain rates as small as 0.036 mm. Compared to the nominal sensitivity of the tipping bucket of 0.1 mm, an intercept of 0.04 results. The scatterometer detects about 13% more rain, which might be considered plausible, as substantial amounts of rain falls as light drizzle with rather low intensities. Nevertheless, some noise may be included, because insects sometimes interfere with the measurement field and erroneously are detected as raindrops (see Leroy et al. 1998). With 0.89, the coefficient of determination between the Thies and the scatterometer shows the best fit of all datasets.
A lid on the NMO tipping bucket is controlled by a rain detector and needs a certain amount of rain before it is opened. Low intensities may be lost but higher rain rates are well detected, because the large 500 cm2 funnel is less affected by adhesion losses of raindrops. Furthermore, the material of the funnel [polytetrafluoroethylene (PTFE)] exerts less adhesion force to individual raindrops than conventional metal collector funnels. Re-evaporation from the funnel is inhibited by the automatic lid, enabling faster reaction to a following rain event. Nevertheless, the actual rain rate is slightly underestimated (0.87) while the correlation is still good.
The remote sensing devices are volume-average measurements; hence, they cannot reproduce point measurements precisely. The MRR underestimates rain rate by about 32%, but as mentioned above, a spatial gradient may be involved as well. A great advantage is the extremely high sensitivity: The MRR is capable of detecting rain rates of 0.009 mm h−1. This is illustrated by the very low intercept of 0.10, showing that light rain is well detected. A problem common to MRR and LAWR is an upper limit to the detectable rain rate. According to detailed data analysis of original minute samples, the MRR has an upper limit of about 4.8 mm of detectable rain rate. This results in a greater scatter for data containing rain intensities above that value. Hourly data above 4.8 mm cannot be detected if the rain rate is equally distributed in time. If rain rates are varying heavily, even lower hourly totals are affected. If all rain rates above 3 mm are discarded from the analysis, the slope is 1.0 and the coefficient of determination increases to 0.80. The possible reason for this effect is a saturation of the receiver unit, which cannot handle higher levels of reflectivity.
For the LAWR the saturation threshold is in the range of 7 mm of rain rate. The hourly data show increasing scatter at this value. The intercept of the derived relation reveals a much higher sensitivity, which can mainly be attributed to the extremely large sample area of one radar pixel. Compared to the tipping-bucket surface area, the radar registers 12.5 × 106 times the area; therefore, the observed frequency of rainfall is much higher. Omitting above saturation rain rates does not improve the coefficient of determination significantly. The slope of the relation cannot be interpreted because the LAWR calibration is derived from several tipping-bucket measurements in the radar range (Rollenbeck and Bendix 2006).
If the data are integrated to daily totals (Fig. 5b), the fit between the different devices improves further, which might be due to slight synchronization errors of the hourly datasets that are eliminated by averaging. Furthermore, the influence of unsystematic errors is canceled out by averaging procedures, so the better fit is not surprising. The MRR dataset, however, does not benefit from the averaging, because the dataset has too many interruptions.
The problematic coverage of higher rain rates by the MRR is also depicted in the comparison of the two rain radars (Fig. 6). Hourly averaged rain rates only show a loose fit between the data; they are cut off at about 5 mm. The average of the daily rain rate has a higher coefficient of determination but the slope value is even lower. Because of its high sensitivity, the MRR registers rainfall more frequently, which increases the average value. The maximum rain rates, however, are not well detected and reach only about 50% of the LAWR values. The situation is almost the same for the daily totals: Low intensities are well detected; days with higher totals and probably higher intensities are not.
Further insight into the characteristics of the different devices can be obtained from the cumulative distribution function (CDF) of rain-rate frequency and the normalized distribution curves as shown in Fig. 7a. Both tipping-bucket instruments are not capable of detecting rain rates below 0.1 mm, but about 7% of all rain events are registered in this lowest class. The distribution rises asymptotically to the maximum value, which is considered a probable representation of the real maximum intensities occurring at this site. The THIES and the NMO devices register the rain rate in a similar quality over the observed range, showing an intermediate sensitivity.
The LAWR measures 14% of all rain events between 0.004 (minimum value) and 0.1 mm and has the highest frequency in the range between 0.3 and 3.8 mm; it shows a positive deviation from the other curves at intermediate intensities. This may be attributed to the higher sensitivity and the integrating measurement principle, but can also be caused by the linear calibration scheme applied to the original data. Raw data of the LAW radar are not clearly described as radar reflectivities or rain rates (Jensen 2002) but are some kind of intermediate value. Possibly a nonlinear relation between rain rate and radar data has to be used at this point in contradiction to what is proposed by the system manufacturer (Pedersen 2004). Further investigation of this problem will be necessary.
The scatterometer and the MRR both show that almost half of the rain events are registered in the extremely low range between 0.01 and 0.1 mm. The MRR, however, although very sensitive to light rain, departs significantly from the other curves and ends at 4.9 mm. This means that higher rain rates are never detected; obviously, the MRR is too sensitive for low rain rates and heavily affected by the saturation effect described above.
Regarding the normalized distribution (Fig. 7b), the point measurements (THIES, NMO, and scatterometer) show a continuous curve over the whole range, meaning that all occurring rain rates are detected with a similar sensitivity. The scatterometer has the best sensitivity to all rain rates, because it shows a uniform curve up to the highest rain rates. According to the manufacturer, this instrument can detect rain rates up to 250 mm. Again, the radars depart in the lower range of intensities, a fact that points to a potential overestimation of this range. While the MRR has the highest sensitivity for rain rates up to 1.5 mm, the LAWR departs from the point measurements at 1.0 mm. If the saturation problem can be overcome, the curve will resemble more the form of the scatterometer curve.
The dataset of the totaling gauges is evaluated on the basis of the weekly samples. Linear correlation (Fig. 8) shows the expected results: The more similar the collector, the more stable the observed values (Sevruk 1981). The comparison of the 104 cm2 gauge versus the THIES tipping-bucket instrument shows 10% higher values than measured with the tipping bucket. The scatter of this totaling gauge relative to the THIES data may be caused by the fact that values were not exactly synchronized with the THIES instrument, because data on the exact hour of reading for the 104 were not available. The intercept of 4.06 mm only accounts for about 2% of average monthly totals, and therefore it is negligible. The coefficient of determination is 0.78.The 198-cm2 gauge shows a perfect slope value for the correlation, and the intercept of 0.61 is very low. This value could be confirmed by a manual calibration of the evaporative loss. This collector resembles most closely the measures of the tipping bucket and hence shows the best agreement. The largest totaling gauge, the 314-cm2 collector, produces a slight underestimation of the tipping-bucket rain rates; the negative intercept value, however, points to a higher sensitivity for light rainfall. The fit is very good too, reaching 0.95.
Table 3 details descriptive statistics for the used datasets. All instruments are in good agreement, although they are not covering the same data collective. F-test statistics show that datasets are not differing significantly (α = 0.05).
The CDF of relative frequency of observed rain totals for this dataset supports the findings of the correlations. The 314 and 198 gauges are in perfect fit with all observed weekly totals, whereas the 104 collector shows a slight deviation from the distribution of the THIES gauge in the range between 50 and 125 mm (Fig. 9).
To achieve a homogeneous dataset, all observations were used to assess the possible error of the largest dataset stemming from the THIES precipitation sensor. The presented correlations show that most of the instruments show unsystematic errors, which can generally be attributed to handling errors or operational problems (interruptions, synchronization errors). Assuming that these errors are unsystematic, all common monthly values were averaged and compared against the monthly totals of the THIES instrument.
The derived monthly average of all datasets is in good agreement with the tipping bucket. The slight underestimation of the THIES (Fig. 10a) probably is a consequence of the measurements, which are capable of detecting light rain (scatterometer, LAWR, and MRR). The hourly dataset shows the same slope value and a slightly higher scatter. It is very likely that light rain and spatial heterogeneity are not as well detectable in the high temporal resolution of one hour by the tipping bucket; actual area rain input may be higher (Fig. 10b).
More important, however, are the ecological consequences of the observed deviations. If the dataset of the THIES instrument is up-scaled by 2% to meet the average dataset and the average of all available measurements is taken, a 12% higher annual total for the supersite results (Fig. 11a) and the secondary dry season from August to November is less pronounced (see Fig. 11b). The rainfall frequency obviously is much higher than can be deduced from the tipping-bucket gauges: while the tipping bucket registers 73% of rain-free periods, the scatterometer and the remote sensing devices show that only 35% of all hours are free of rain. Although this does not greatly increase the registered totals, it has a significant effect on the ecosystem.
The comparison of rain measurements obtained with different methods reveals some important results for the ecological work of this group and for precipitation measurements in tropical montane environments in general. Because the vast majority of rain events have low intensities, the potential errors from neglecting light rainfall is larger than in lowland applications. Rain frequency is drastically higher if the detection threshold is lowered to values that are in the range of modern precipitation sensors.
Totaling gauges, although most commonly used at least in developing countries, are not always capable of registering the full range of all occurring rain rates. Obviously only sophisticated instruments can detect light and heavy rainfall at the same time. This is an important fact for ecological investigations in tropical mountain forests, which are dominated by light advective rainfall (drizzle), but may also experience heavy convective events. An improved detection limit and the superior measurement principle of the radars and the scatterometer enable a better estimation of the frequent light rain. Although small rain intensities do not contribute much to the overall total, the length of rain-free periods is much shorter as can be deduced from standard measurements. The ecological importance of this fact may be explained by the example of epiphytic vegetation: These plants form an important part of the tropical montane ecosystems, but they can only survive short periods of dryness and cannot rely on soil water storage. Hence, the rain frequency plays an important role in the development of the high biodiversity.
According to our results, the scatterometer is the most recommendable instruments for automatic stations. Regarding the fact that all precipitation measurements tend to underestimate real rainfall, the best results for point values in this study are probably obtained from this instrument, which measures precipitation without physically interfering with the raindrops. A deeper analysis of the drop spectra registered by this instrument will enable an assessment of the potential noise, which increases the amount of light rain detected. Nevertheless, totaling gauges are indispensable because of their reliability and simplicity of setup and operation, especially under budget limitations. A special recommendation for one of the three instruments cannot be given; however, the 198 collector reproduced the data of the recording devices most closely.
The tipping bucket is a rather reliable and precise instrument but compared to the remote sensing methods and the scatterometer a slight underestimation of very light rain seems probable. The remote sensing methods used here have shown their usefulness for determining area rainfall as atmospheric contribution to the hydrological cycle. The spatial scale and averaging capabilities of radar data make them most useful for modeling purposes, since arbitrary sampling errors of conventional point measurements are avoided. Their high sensitivity for light rain and the accordingly higher frequency of rainfall disguises the errors of these instruments occurring with higher rain intensities. The totals detected are the same (MRR) or even higher (LAWR) than those recorded by the point measurements, but actually a part of the rainfall is not registered because it is above the upper threshold of rain rates. Only a thorough investigation of the data on different time scales can reveal this problem. It will be addressed by changing hardware configuration and data processing of the radars.
With regard to operational considerations, the importance of redundancy must be stressed: although automatic measurements offer improved frequency and resolution, many problems remain unsolved. The reliability of modern equipment is not yet satisfying. The most frequent reason for interruptions of the measurements were failures of the power source, be it the very unstable supply in this part of Ecuador with frequent power spikes and blackouts or the premature deterioration of the batteries used in dataloggers. This often was caused by the high humidity and corrosion of electric connectors. Even the introduction of uninterruptible power sources could not improve this significantly, because they cannot handle long interruptions and the extremely bad quality of power supply. Software glitches in the control hardware of the radars and the scatterometer were another source of errors and failures. Regarding the fact that all of these devices or their software used in this study have prototype character, this problem will be overcome in the future.
The data of the totaling gauges mainly suffered from wrong handling during the fieldwork, deterioration of the material and errors introduced, while manually copying the field data to the database.
Project PREDICT (FOR 402/1, BE 1780/5-2, FA 62/17-2) and several other contributing subprojects of the researcher unit FOR 402 are funded by the German Research Council (DFG). We greatly acknowledge the support of NCI/FCSF and INAMHI (Instituto Nacional de Meteorologia e Hidrologia) and the help of Michael Richter and many field assistants of the research program for supplying field data of precipitation measurements. We furthermore thank the Ecuadorian Ministry of Environment for the permission to conduct this work (research permit 001-2005-DBAPVS-RLZCH/MAE).
Corresponding author address: Rütger Rollenbeck, Dept. of Geography, University of Marburg, Deutschhausstr. 10, 35032 Marburg, Germany. Email: email@example.com