AVI 080 Project "ACROSS"
| |
AVI 080 Project "ACROSS" |
|
"Analyzed Climatology of Rainfall Obtained from Satellite and Surface data
for the Mediterranean Basin"
by M.C. Todd, E.C. Barrett and M.J. Beaumont
Existing algorithms for estimation of rainfall from passive microwave satellite data can be grouped into three categories:
Empirical algorithms are based on a statistical relationship derived between a single channel brightness temperature or particular combination of multi-channel Tbs, and rainfall measured by some other means, usually weather radar or rain gauges. The statistical relationship is then used to derive rain rates and accumulated rainfall from measured Tbs. The derived rainfall/Tb relationship is therefore simply an 'average' of all individual rainfall/Tb relationships experienced over the calibration area and period. Such algorithms have the advantage of computational simplicity, are based on 'real' data, and have shown useful results at least in areas experiencing similar rainfall characteristics to the calibration region. They do however lack sensitivity to space/time variations in rainfall processes. All the Bristol methods fall into this category.
Physical/empirical algorithms also utilise a statistical relationship between Tbs and rainfall, but one which is derived not from actual measured Tbs and rainfall but from the output of radioactive transfer modelling using a simulated modelled atmospheric profile. Thus they are not dependent on accurate collateral measurements of rainfall, itself a problematic issue, but are limited by the physical accuracy of the atmospheric models used.
Physically based algorithms attempt to use complex models of cloud and rainfall processes in order to more accurately represent the vertical profiles of ice and liquid hydrometers in the atmosphere and the resulting multi-channel Tbs. Such methods are based strongly on models of cloud physics and have so far been largely directed at estimation of rainfall from individual storm events. Currently these methods are computationally intensive but may be expected to underpin a future generation of improved rainfall algorithms.
SMMR algorithms for rainfall estimation over oceans
Barrett et al. (1990) provides an overview of some of the rainfall algorithms developed for use with SMMR data in the context of rainfall estimation over the North Sea. Listed below are a selection of the principle rainfall algorithms developed to date:
Barrett et al. (1991) presents results of the application of the H37 algorithm to SMMR data for rainfall estimation over the North Sea. The algorithm is applied to all footprints with a 37Tbh exceeding 160K. To account for the poor temporal sampling of the SMMR instrument and the saturation of the 37Tbh at rain rates beyond about 5 mm hr-1, a statistical connection was additionally applied. All footprints with a 37Tbh > 229 K were ascribed a rain rate of 10.1 mm hr', representing a weighted average of rain rates between 5 mm hr-1 and an estimated absolute footprint maximum of 20 mm hr-1. This algorithm was successfully used to derive a rainfall climatology of the North Sea region.
A similar method that has been described by Prabhakara et al. (1992)
The depolarising effect of rain liquid emission is utilised in the polarisation difference algorithms (Kidd and Barrett, 1990):
PD = Tbv - Tbh
The 37 GHz frequency has proved most popular to date due largely to
the higher resolution at this frequency. Use of both vertical and horizontal
polarisation reduces the temperature dependence which affects a single
channel approach. The 37 GHz PD algorithm is still limited by saturation
effects at moderate/high rain rates. In the WetNet PIP-l Project the PD37
algorithm was used to retrieve rainfall over oceans globally as part of
the ensemble of Bristol algorithms.
SSM/I algorithm for estimation of rainfall over oceans
Of the algorithms described above, algorithms 1 and 2 can be used with data from the SSM/I sensor too. The addition of the 85 GHz channel opens up a further opportunity to utilise the scattering signal. The failure of the 85 GHz channels on the SSM/I DMSP onboard F8 in December 1987 means that this high frequency data is unavailable from then until the launch of DMSP F10 in December 1991. A full description of the available SSM/I data set is provided by Todd and Barrett (1996). For the purposes of this study therefore only algorithms utilising the low frequency channels (19, 22 and 37 GHz) can be considered, e.g. the Fiore et al. NESDIS Algorithm now used operationally by NOAA.
Calibration and validation data sets
Establishing suitable data sets for calibrating and validating satellite estimates of rainfall over oceans has long been a major problem. In this study the following data are used.
Meteo-France currently operate a network of 13 weather radar. Only the radar at Marignane (operating between 1987 and April 1990) and at Nimes (operating since 1990) provide coverage over the Mediterranean Sea. Reflectivities are converted to rain rates which are then binned into 14 rain rate categories. Raw data is not archived. Radar imagery in this form was obtained for 19 cases in 1992 selected on the basis of known rain events in the region as defined from quick-look SSM/I. The radar image immediately following the SSM/I overpass time was chosen. Thus Meteo-France radar data is available only for comparisons with SSM/I data, not SMMR data. For quantitative comparisons each pixel was assigned a rain rate equivalent to the mid point of the rain rate category.
Because of this limitation and the restricted number of cases Meteo-France data is utilised only for validation of SSM/I rainfall estimates.
The Centre for Remote Sensing has over the years acquired an extensive set of data from the UK FRONTlERS weather radar network. This network has grown over the years and now includes 13 radar covering all of the UK, Ireland and surrounding seas. The data consists of FRONTIERS products produced from raw radar estimates through an interactive correction process involving:
For present purposes, calibration of SMMR rainfall algorithms is based on data from a selection of dates in 1982 to 1987, chosen to provide samples from a range of seasonal rainfall conditions. These data are restricted to the Cambome radar, the only radar providing significant coverage of sea areas at this time, and are not deemed extensive enough to permit independent validation of satellite estimates.
SSM/I algorithm calibration and validation is based on the period 1/5/92 - 30/4/93 and involves all radar covering water surfaces.
Calibration of the SMMR rainfall algorithm
Of the algorithms available for rainfall retrieval over oceanic surfaces, attention was directed principally at the H37 algorithm for a number of reasons. First, the algorithm provides estimates at the highest resolution possible from SMMR data. Second, it is computationally very simple thereby minimising processing time, an important consideration given the size of the data set. Third it has been shown to provide acceptable estimates over the North Sea. Fourth it includes an empirical correction to account for the sampling characteristics of the SMMR sensor. Single channel algorithms have been developed elsewhere and are suitable for climatological purposes where the effects of parameter crosstalk from other features such as sea surface roughness, cloud and water vapour can be statistically averaged out.
The H37 was chosen in preference to the similar Prabhakara et al. (l992) algorithm which was developed for estimation of tropical rainfall. The methods of Hinton et al. 1992 and Prabhakara et al. (l986) were deemed inappropriate due to the poorer spatial resolution of the low frequency channels used and the irregular footprint pattern of the SMMR sensor. In addition these algorithms are bad on the relationship of cloud liquid and rainfall for tropical ocean regions and are therefore likely to underestimate rainfall in regions where the atmosphere is typically cooler and drier.
Algorithm calibration
In its original formulation the relationship relating 37Tbh to rain rate was derived empirically from averaging all observed Tbs coincident with radar pixels in particular rain rate categories at 0.5 mm hr-1 intervals (Barrett et al. 1991). The resulting relationship shows a non linear increase 37Tbh with rain rate until 5.0 mm hr-1 - the maximum rain rate for which a significant number of observations occurred. To account for both the saturation of the 37Tbh channel at rain rates above 5.0 mm hr-1 and the poor temporal sampling of short duration peak in events an empirical correction is applied. This algorithm is hereafter referred to as H37A.
Todd (1993) found that direct pixel by pixel comparisons of satellite and radar data may not be the most effective method to establish an empirical relationship relating satellite and radar quantities. Such relationships are typically characterised by extensive scatter contributed by external sources such as temporal and spatial displacement of imagery and errors within the calibration data. This can seriously bias a regression type empirical relationship. While both the satellite and radar sensors may produce errors in the spatial distribution of rainfall, overall rainfall histograms are generally very stable. Thus for ACROSS a new method of algorithm empirical calibration was developed based on the strength of histogram stability. In this approach histograms of radar rain rates and SMMR 37Tbh are proportionally equalised. For radar rain rates at 0.1 mm hr-1 intervals above 0.0 mm hr-1, the proportion of the total number of observations are calculated from the histogram. The 37Tbh values which are associated with the boundaries of these same class proportions are extracted from the Tb histogram. Thus a look-up table is created with rain rate category boundaries and associated 37Tbh values. In this histogram equalisation approach it is simply assumed that increasing rain rates are associated with increasing 37Tbh values, and the spatial correspondence of radar rain and 37Tbh observations is ignored. The 37Tbh/rain rate look-up table was derived using the same calibration data set of SMMR and FRONTIERS observations as used in the calibration of the H37A algorithm. The 37Tbh algorithm using the optimised Tb/rain rate look-up table is referred to as H37B.
A similar exercise was conducted using the SSM/I calibration and validation data set where alternate days were chosen for calibration and validation purposes. The H37B relationship was derived from the calibration cases and then applied to the validation period.
All these arguments which indicate that the optimised approach may be more accurate are made from the comparison of 37Tbh values and rainfall over the FRONTIERS area. It is possible that the Mediterranean basin experiences a different rainfall regime. However, analysis of the rainfall histograms derived from the GPCP TOGA COARE data set, obtained by radar over the tropical Pacific (climatologically even more remote from the UK), shows a remarkable similarity with that derived over the UK (Kidd 1995 p.c.). It may be that over water where there is less thermal and orographic convective forcing than over land rainfall intensity histograms are far more uniform and low rain rates proportionally dominate. This increases our confidence in extrapolating relationships derived from the UK FRONTIERS region to the Mediterranean region.
Calibration of the SSM/I rainfall algorithm
The SSM/I sensor represents an improvement on the SMMR sensor for rainfall monitoring for a number of reasons, including the fact that the SSM/I records at 85GHz providing an opportunity to quantify rainfall at relatively high resolution from the scattering signal as well as from the emission signal at lower frequencies. However until the launch of the DMSP F10 satellite there was a substantial gap in the 85 GHz data record. This discontinuity in data restricted the choice of algorithm for the ACROSS study, to those using the low frequency channels only.
This Section therefore focuses on derivation of a suitable Tb/rain rate relationship for the H37 algorithm and comparison of resulting rainfall estimates with those from other SSM/I rainfall algorithms.
Derivation of the 37Tbh rain rate relationship
Following from the findings of the work on calibrating the H37 algorithm for SMMR data a 37Tbh/rain rate look-up table was created from large SSM/I/FRONTIERS data set using histogram equalisation method. The resulting algorithm, referred to as H37C, was compared against the other algorithms available.
The H37C algorithm produces the best results in terms of mean rain rate and the Normalised Root Mean Squared Error (NRMSE), with only slight overestimation of rainfall amount. All other algorithm overestimate ranging from 20% to 282%.
For climatological purposes, particularly over ocean regions where low rain rates appear to dominate, accurate estimation of the mean rain rate is of primary importance, suggesting that the H37C algorithm represents the optimum choice for the SSM/I period within the ACROSS Project.
A key element of the ACROSS Project objectives is the investigation of the surface effect of significant rain anomalies using the SMMR and SSM/I satellite data. Thus the identification, quantification and mapping of rainfall anomalies from surface and satellite observations is to be linked with surface processes in the eastern Mediterranean region. Historically SMMR and SSM/I data have been used to monitor changes in soil moisture, vegetation characteristics and surface inundation and these will be the primary parameters of interest in this work. The theoretical bases for surface monitoring are outlined in the report, followed by a description of algorithms developed to retrieve data and a proposed analysis strategy for the ACROSS Project.
