Final Report on the Visiting Scientist activities:Refinement and operational implementation of a rain ratealgorithm based on AMSU/MHS and rain gauge data over

H-SAF area

Paolo Antonelli

1 May 2007

Acknowledgments

This work would have not been done without the contributions of several colleagues. A deep sense ofgratitude goes to Ralf Bennartz and Hal Woolf who kindly shared their software and more importanttheir valuable experience to guide me throughout the entire duration of the project. Same gratitudegoes to Alessandra Tassa for her contribution in improving the precipitation retrieval algorithm, for herfriendship and for her patience in dealing with me. It has been a privilege to work closely with AntonioGioia, Pietro Giordano, Luca Rossi, William Nardin, Davide Melfi, Daniele Biron, and Corrado deRosa, not only for their technical support, but also for their friendly and professional attitude. Itwould be impossible not to thank the H-Saf management, in the person of Gen. Sorani, and his staffin the person of Angela Corina, for the interest they constantly showed in the project. A special thankgoes to the H-SAF science manager, Luigi de Leonibus, for his kindness and more important forhis scientific enthusiasm, which characterizes his activities from the presentations to every personaldiscussions. Finally my full respect goes to the tutors: Silvia Puca and Francesco Zauli for theirdetermination and for the impressive volume of work they have done to support me in any possibleaspect. It has been a true pleasure to work in collaboration with the Italian Civil Protection and theItalian Meteorological Service of the Airforce, for the resources they provided and for their pricelesshospitality.

1

Summary

The Visiting Scientist (VS) activities focused on the generation of satellite-derived precipitation prod-ucts potentially useful for the hydrological and Numerical Weather Prediction (NWP) communities.Existing AMSU and SEVIRI products were merged to provide estimates of convective precipitationat SEVIRI spatial resolution. The derived SEVIRI, AMSU, and combined products were qualitativelyand quantitatively validated using in-situ radar and rain-gauge observations.

In this framework a new approach was developed within the VS project aimed at combining existingAMSU precipitation products with SEVIRI convective products in order to derive information on con-vective precipitation at spatial resolution useful for the hydrological models and Quality PrecipitationForecast (QPF). In spite of the pioneering stage of the approach, validation results obtained from radarand rain-gauge products indicated that considerable benefits could be achieved by an operational andrefined implementation of the proposed technique.

Specific activities for the time period September 2006 - April 2007 can be summarized as follows:

• Collection of AMSU Data from NOAA 15/16/17/18 over H-SAF area starting Feb 1st, 2006(updated daily);

• Selection of precipitation retrieval algorithm from AMSU data for Visiting Scientist (VS) projectpurposes;

• Acquisition and Installation of Precipitation Classification (PC) package for likelihoods of dif-ferent intensity classes of precipitation;

• Extension of PC coverage to H-SAF area;

• Selection of 12 Test Cases with intense convective and frontal systems;

• Application of PC algorithm to the Test Cases;

• Acquisition of RADAR and RAIN GAUGE data from the Italian Civil Protection (DPC) for theselected Test Cases;

• Acquisition of SEVIRI derived product (NEFODINA) from the Operational infrastructure ofItalian Meteorological Service of Airforce (CNMCA) for the selected Test Cases;

• Acquisition of RADAR and RAIN GAUGE data for the selected Test Cases on SEVIRI gridfrom DPC and CNMCA;

• Development of co-location software to re-map RADAR data on AMSU grid;

• Development of co-location software to re-map AMSU data and AMSU derived precipitationproducts on SEVIRI grid;

2

• Definition of a qualitative procedure to compare PC results with RADAR and RAIN GAUGESon native (AMSU) grid;

• Definition of a qualitative procedure to compare PC results with RADAR, RAIN GAUGES andNEFODINA on SEVIRI grid;

• Definition of a quantitative procedure to compare PC results with RADAR on AMSU grid;

• Publication of the results on a web page (http://lupin.ssec.wisc.edu/HSAF/VS.html);

• Set up of a CVS repository (archiving and versioning system) for software development;

• Presentation at the ITSC- 15 International Tovs Study Conference Maratea, Italy 4-10 October2006;

• First Interim Report;

• Operational implementation of PC algorithm at the CNMCA;

• Modification of Nefodina output format for merging of NEFODINA product with PC retrievals;

• Evaluation of parallax error for NEFODINA and for AMSU PC retrievals;

• Implementation of a new background Brightness Temperature (BT) calculation scheme for thePC product;

• Development of mapping software to calculate Effective Field Of View (EFOVs) for AMSUand NEFODINA;

• Implementation of a strategy for the merging of SEVIRI convective products (NEFODINA)and PC product to derive precipitation with spatial resolution higher than the native AMSU-Bresolution;

• Analysis of the results and recommendations for future activities;

• Final report on the Visiting Scientist activities.

This report aims to provide a detailed description of the summarized activities. Chapter 1 describesthe precipitation products derived from microwave data and covers the description of the satelliteobservations used, of the precipitation retrieval algorithm, of the derived products with the relatederrors and uncertainties. Chapter 2 introduces the SEVIRI data and the derived convective product,covering the satellite observations used, the automatic detection algorithm, the derived product alongwith the related errors and uncertainties. Chapter 3 is the core of the document and describes the pro-cedure adopted to merge precipitation products derived from polar orbiting microwave observationsand convection products from geostationary infrared observations. Chapter 4 presents the validationresults obtained comparing in-situ observation with microwave derived precipitation products, withconvection products derived from infrared data, and with the convective precipitation product ob-tained merging infrared and microwave derived products. Chapter 5 provides the final conclusionsand the recommendations.

3

Chapter 1

Precipitation retrieval from microwave data

1.1 AMSU Data

The Advanced Microwave Scanning Unit (AMSU) is a cross-track scanning microwave radiometerconsisting of two separate modules: AMSU-A and AMSU-B. The first one has sounding channelsin the water vapor (23.8 GHz) and oxygen absorption bands (50 GHz) complex plus some windowchannels (at 31.4 and 89GHz) and is dedicated mainly to the retrieval of temperature and water vaporprofiles. AMSU-B has two window channels (at 89 and 157 GHz) and three channels in the watervapor absorption band at 183 GHz, and is dedicated to the retrieval of ice cloud and precipitation[15]. Being a sounding instrument, AMSU-A has a much lower resolution than AMSU-B. Channelsand spatial resolutions are detailed in the Tab. 1.1. AMSU data used for this study were obtained from4 platforms (NOAA 15, 16, 17, 18 ) over the H-SAF domain (Europe), for a total of 8 overpasses perday, were archived starting from Feb 1st 2006 and will continue at least until June 1st 2007. AMSUdata from METOP were added to the archive after launch. Examples of AMSU-B data at 89 and157 GHz are showed in figures 1.1 and 1.2.

1.2 Precipitation Algorithm

1.2.1 Algorithm Selection

Among different physical and statistical candidates (i.e.[5, 2, 4, 14]), the Bennartz’s ([14]) operationalalgorithm was chosen as baseline. Bennartz’s algorithm, hereafter referred to as Precipitation Clas-sification (PC) algorithm, provides likelihoods for four different intensity classes of precipitation. Ithas been developed by Prof. R. Bennartz of the University of Wisconsin - Madison in collaborationwith the Swedish Meteorological and Hydrological Institute (SMHI) within the Now-Casting SAF.

Satellite characteristics AMSU-A AMSU-B

Spatial resolution 3.3◦ 1.1◦

Nadir effective FOV 50x50 km2 20x16 km2

Scan edge effective FOV 150x80 km2 64x52 km2

Channels23.8, 31.4, 50 GHz

O2complex, 89.0GHz89.0, 150.0, 183

(WV abs. Band) GHz

Table 1.1: AMSU-A and AMSU-B (MHS) instrument characteristics.

4

Figure 1.1: Example of AMSU-B at data 89 GHz.

5

Figure 1.2: Example of AMSU-B at data 157 GHz.

6

Currently the algorithm runs operationally at SMHI using AMSU data (see 1.1). The decision wasbased on the following considerations:

• Bennartz’s algorithm runs operationally over Europe at the Swedish Meteorological and Hy-drological Institute (SMHI). It was developed and implemented within the Now-Casting SAF;

• The system provides likelihoods for four different intensity classes of precipitation, and itsextension to Rain Rate estimation is feasible and desirable within both the Now-Casting andthe Hydrological SAFs;

• The algorithm is robust and, provided the availability of gauge calibrated radar data, can easilybe adapted to provide Rain Rate estimates;

• The algorithm has been already trained over Northern Europe and with the availability of radarand rain gauge data over Italy, the extension of the training to Southern Europe climate regimesis straightforward.

In particular, immediate availability of a validated algorithm, the idea of building upon the work donewithin a different SAF, and the opportunity of improving the algorithm with beneficial effects for twodifferent SAFs, without compromising on the accuracy of the final products, have been key factors inthe algorithm selection.

1.2.2 Algorithm Theoretical Basis

In the proposed algorithm, channels from both AMSU-A and AMSU-B modules are used, dependingon surface conditions. The main principle is that rainfall is derived from evaluating the variation ofthe measured Brightness Temperatures with respect to the background BTs obtained in absence ofprecipitation. Over land, the radiative cooling due to the scattering of the radiation by the ice parti-cles overshooting precipitating convective clouds is taken as a rain indicator, and AMSU-B windowchannels (89 and 157 GHz) are used for deriving a so-called Scattering Index (SI). Over ocean,AMSU-A channels at 23.8 and 31.4 GHz are used, taking advantage of the depolarizing effects thatrainfall has on the emitted radiation with respect to the oceanic background, which appears highlypolarized. Over mixed areas, AMSU-B is more useful, thanks to its higher spatial resolution.

Fig. 1 outlines the general data flow of the PC retrieval. In the first step, the type of surface isdetermined. This is done because the different types of surface are characterized by different radiativesignatures: their impact on rainfall estimation is therefore of extreme importance. Primarily, land andwater are separated into two different classes using a land/sea mask. Then, for every surface class,also snow cover and sea ice is distinguished. Since the spatial resolution of the AMSU is very low, itis further necessary to define a particular surface type for coastal observations, that are partly coveredby land and water.

Depending on the surface type, different linear combination of AMSU channels are used to derive theSI according to the rules defined in Tab. 1.2, where TXX is the Brightness Temperature at XX GHzand ∆Tbk is the background BT difference dynamically calculated. SI is successively used to calculatethe precipitation rates. For a case-by-case detailed description, please refer to [14]. Final products areobtained at AMSU-B resolution [11].

Distinguishing mixed surfaces, and especially coastal areas, is crucial in regions characterized bylarge spatial heterogeneity, as it is the case for Mediterranean regions. It must be highlighted, however,that the impact of the surface characteristics is inversely proportional to the precipitation intensity:clouds and precipitation may transmit, absorb/emit and/or scatter away the radiation emitted from

7

Figure 1.3: Precipitation Classes algorithm data-flow [Figure exerted from [11]]

Type of surface Scattering index computation

Ocean homogeneous only for AMSU-B (AMSU-A mixed) sl1 = (T89 − T150)−∆Tbk (89−150)

Land homogeneous for both AMSU-A and AMSU-B sl2 = (T23 − T150)−∆Tbk (23−150)

Land homogeneous only for AMSU-B sl1 = (T89 − T150)−∆Tbk (89−150)

Coast sc = (1− l)sl1 + lsl2

Table 1.2: Scattering Index computation according to different surface types, as from [14].

8

Class Type of precipitation Minimum rain rate (mm/h) Maximum rain rate (mm/h)

1 No precipitation 0.0 0.12 Risk of light precipitation 0.1 0.53 Light/moderate precipitation 0.5 5.04 Intensive precipitation 5.0

Table 1.3: Classes of different precipitation intensities used in this investigation (see [14]).

the surface, thus completely hiding its contribution to the overall radiation emerging from the top ofthe atmosphere and finally sensed by the radiometer. This “filtering”, of course, depends on severalfactors, such as the surface, the cloud/precipitation properties, and from the wavelength of interest.

The PC algorithm produces likelihoods of rainfall classes (i.e. the probability that the rainfall fitswithin predefined ranges of rain rates). Tab. 1.3 shows the different classes used. A given pixel willnot be assigned a certain value but rather a set of the probability according to which it belongs to eachof the four classes. The algorithm’s approach is empirical. The relationships between the precipitationclasses and the AMSU-derived scattering indexes is derived from the use of co-located, convolvedradar data. This approach allows to overcome the large systematic deviations that the current lack ofknowledge about the microwave response to cloud/precipitation microphysics would introduce in casea detailed inverse modeling approach was chosen. This is especially true where surface characteristicsare highly heterogeneous. The precip-radiation database, created by convolving rain-gauge adjustedradar estimates, was used to calibrate the algorithm. The database is composed by thousands ofcouples of AMSU observations and related precipitation classes. These data have been collectedduring eight months (April November 1999) at SHMI, and cover the Baltic Sea region. Due to thelimited representativeness of this data set, the algorithm sensitivity, in the current implementation,does not allow for the discrimination of more classes.

It is worth emphasizing that the derived thresholds and probabilities and the created database maynot be exported to other regions straightforwardly. In fact, an adjustment to other climate regions isdeemed necessary. It is also important to recognize that the database must be statistically significant,and all the meteorological and climatological situations of interest must be conveniently represented.In fact, the problem of the representativeness of the database is a well-known problem in rainfallretrievals, and may hamper the validity of the retrieval if not properly handled [9, 1, 16, 3].

Finally, when using the precip-radiation database, one must be conscious that several errors areintrinsically taken on-board like the spatial and temporal collocation errors between radar data andsatellite observations (due for example to temporal misalignments, to satellite navigation errors, tothe different geometry of the observations).

1.3 PC products

The PC products generated on the AMSU native grid are:

• likelihoods for four different intensity classes of precipitation;

• PC standard RGB composite (Fig. 1.1) The RGB is obtained mapping the probability of eachpixel belonging to class 1 to BLUE, class 2 to GREEN, and class 3 to RED. The RGB obtainedin this way shows high precipitation in RED and low precipitation in BLUE. Absence of pre-cipitation is shown in white. The RGB conveys the full information about the likelihood of eachpixel belonging to any of the four classes, however it does not provide a single value per pixeland therefore can be used only for qualitative comparisons with RADAR or RAIN GAUGES.

9

In order to obtain point values of precipitation, the PC product is further manipulated in two differentways:

• Maximum probability approach (Fig. 1.6). The class with the highest probability is chosen.It is a streightforward approach, however it can be misleading: for example, if for a givenpixel the probability across two or more different classes was uniform, the algorithm wouldstill assigning one class and part of the information regarding the actual probability distributionamong the different classes would be lost;

• Weighted average Rain Rate approach (Fig. 1.7). The weighted average approach provides anarbitrary estimate of the rain rate according to the following equation: RR =

∑pirri where

pi is the likelihood, and rri is the mean rain rate value for class i. This quantity conveysthe full information about the likelihood of each pixel belonging to any of the four classes,but the actual RR values are arbitrary as the mean value for each class (especially class four,intense precipitation) are set to a given value in an arbitrary way. In spite of this limitation,under specific circumstances, the representation is still useful for qualitative and quantitativecomparisons.

1.4 Improvements to the PC algorithm

As described in section Section 1.2.2 the performances of the PC algorithm largely rely on the com-putation of a realistic radiative background. In fact, the rainfall estimation is based on the deviationof the measured BTs with respect to the background BT (i.e. the BT that would be measured in clear-sky conditions). In its original implementation, the background BT is computed dynamically over awide region surrounding the pixel of interest (2 degrees by 2 degrees). This technique is simple andefficient but has several drawbacks, since the possibility that the resulting BT is contaminated by pre-cipitation and/or by heterogeneous backgrounds is extremely high. As a consequence, a major efforthas been devoted to developing a new technique for deriving background BTs: a grid of clear-skypixels is built by making use of a certain number of AMSU overpasses. For a given time intervalpreceding the one of interest, AMSU data are projected onto a fixed regular grid (the SEVIRI one hasbeen used). This projection is done by using the re-mapping process described in section 4.2. Thewarmest AMSU clear-sky BTs at 89 and 157 GHz, among all the overpasses within a certain timewindow (4 days in this study), are retained and mapped onto the the higher resolution SEVIRI fixedgrid.

This approach provides the best results but is not optimal. In fact, over the (radiatively) hot landsurfaces, scattering cooling from precipitation causes BTs to decrease. Thus, hot backgrounds easilymark clear sky pixels. However, the same reasoning cannot be applied over oceanic surfaces. In fact,since sea surface emissivity at 89 and 157 GHz vary between 0.6 and 0.7, ocean backgrounds arealways radiatively cold. Convective events still cause radiative cooling by scattering, however watervapor and light precipitation can warm the measured BTs with respect to the clear sky signatures.Therefore, hottest pixels might not necessarily correspond to clear sky ones. However, since thewarming varies with the wavelength (the 89GHz being more sensitive to water vapor), the differentialsignal should in theory be kept as an indicator of clear sky pixels. The SI should be maximum atclear sky (due to the differences in SSE) and minimum during convective precipitation (due to theenhanced scattering at 157 GHz than at 89 GHz). The differential discrimination was attempted butcaused significant errors by detecting light precipitation over clear sky ocean regions, therefore wasnot adopted for this study as deeper investigation is needed for its use.

10

Figure 1.4: Example of PC standard RGB composite. The RGB is obtained mapping the probabilityof each pixel belonging to class 1 to BLUE, class 2 to GREEN, and class 3 to RED. The RGB obtainedin this way shows high precipitation in RED and low precipitation in BLUE. No precipitation is shownin white. The RGB conveys the full information about the likelihood of each pixel belonging to anyof the four classes, however it does not provide a single value per pixel and therefore can be used onlyfor qualitative comparisons with RADAR or RAIN GAUGES.

11

Figure 1.5: Example of likelihood for each individual class of precipitation as described in Tab. 1.3.

12

Figure 1.6: Example of origial PC product: classification using the maximum probability approach.Classifying each pixel according to the maximum probability among the different classes provides aunique value (class), however this value can be misleading when, for example, the probability, fora given pixel, is spread across two or more different classes. In this case the algorithm would stillassigns the pixel to one class and of the information regarding the probability distribution is lost.

13

Figure 1.7: Example of original PC product: classification using the weighted average Rain Rate. Theweighted average conveys the full information about the likelihood of each pixel belonging to any ofthe four classes, but the values above 5 mm/hr are arbitrary as the mean value for class four (intenseprecipitation) has been set to 10 mm/hr in an arbitrary way. In spite of this, the representation is stilluseful for qualitative and quantitative comparisons.

14

Figure 1.8: Example of RADAR data convolved at AMSU-B spatial resolution and binned using thesame classification thresholds used by the PC algorithm.

15

A time window of 4 days used for the computation was proved to be large enough as to allow thepossibility that clear sky conditions are found for every pixel, and short enough to remain withinbackground conditions that are representative of the overpass under examination.

In this regard, it is worth noting that a considerable impact has been observed from night/day varia-tions (implying deviations as large as 30K on the background BTs), a fact which suggested to restrictthe computation of the background temperatures to overpasses within a few (6 in this study) hoursfrom the time of interest. In order to avoid problems linked to the instrument calibration, also separatecalculations were done for the different satellites.

The described approach results in a considerable improvement on the computed clear-sky BTs, andthis turned into a significant impact on the derived precipitation estimates as shown in figures 1.9and 1.10. The improvements are visible especially over coastal regions (North Africa) and over nonvegetated land (Sicily and Sardinia) where with original scheme, false precipitation was detected inclear sky regions (see fig. 2.2), whereas with the new scheme, the misclassification tends to disappear.Figures 1.11 and 2.3 show the improvements even over ocean with respect to the original version ofthe PC algorithm for a different NOAA 18 overpass (14 09 2006 at 01 : 56 UTC) with respect to theone shown in figure 1.1. For the sake of clarity, figures 1.4, 1.5, 1.6, 1.1 presented in this report showexamples obtained with the original version of the PC algorithm, and were retained for consistencywith the first interim report of the VS activities.

The algorithm also allows for an improvement in the estimation of precipitation over very hetero-geneous surfaces (namely over coastal areas): the fact that the background BT selection is done byconsidering several overpasses and on a fixed, high-resolution grid (the SEVIRI one is used), theheterogeneity of the background BTs is better retained, thus allowing a mitigation of the estimationerrors.

A possible improvement for this algorithm could consist in a smoothing of the computed resultingbackground BTs before ingesting them in the retrieval scheme. In fact, being computed over differentoverpasses (and therefore even from different observation angles) the radiative backgrounds maypresent some important discontinuities among adjacent pixels (e.g. differences as large as 10% couldderive from the different scan angle of the retained clear sky pixel) that could impact on the final rainestimation.

1.5 Errors and uncertainties

1.5.1 AMSU geolocation uncertainties

AMSU data suffer from considerable geolocation errors, depending on the platform. NOAA scientistshave evaluated these errors by analyzing the discontinuity in correspondence of coastline in clear-skyconditions.

Errors have been found in both the along and cross track direction. The displacement has been quan-tified as large as 4.2 and 5.5 km (cross-track and along-track respectively) for NOAA-17 and 3.2 kmand -16.4 km for NOAA-18. As for NOAA-18, an example of the considerable displacement (2pixels) that has been evidenced on the Red Sea with such a technique is clearly visible in figure 1.12.

1.5.2 Scan-dependent uncertainties

The fact that AMSU is a cross-track scanning instrument has a major impact on the relative quality ofthe resulting rainfall estimations, due to several reasons.

16

Figure 1.9: NOAA 18 16-08-2006 at 01:52 UTC: PC retrieval obtained with original estimationof background BTs. Light precipitation is erroneously detected in clear sky regions (over Sicily,Sardinia, and North Africa), see fig. 2.2.

17

Figure 1.10: NOAA 18 16-08-2006 at 01:52 UTC: PC retrieval obtained with estimation of back-ground BTs derived from previous time-coincident, same-platform overpasses. No precipitation isdetected in the clear sky regions over Sicily, Sardinia, and North Africa.

18

Figure 1.11: NOAA 18 14-09-2006 at 01:56 UTC: PC retrieval obtained with improved estimation ofbackground BTs derived from previous time-coincident, same-platform overpasses. No precipitationis detected in the clear sky regions over Sicily, and over ocean. In this figure PC values are mappedonto actual AMSU Effective FOVs (EFOVs) calculated according to Bennartz [11].

Figure 1.12: NOAA-18 AMSU-B image at 89GHz over the Red Sea (figure exerted fromhttp://www.orbit.nesdis.noaa.gov/smcd/spb/n18calval/calval/mhs.html#mhsgeo )

19

Figure 1.13: Schematic representation of the localization error for surface rainfall rate position. Atslant angle β, rainfall is located in correspondence of position P, whereas the current location shouldbe P’. The corresponding shift x is referred to as the parallax error.

First of all, different viewing geometries correspond to different sensed areas and different radiativeproperties. Atmospheric optical thickness, surface emissivity, hydrometeors particle emissivity de-pend to different extent from the viewing geometry. Even purely geometrical effects related to thechange in the equivalent area of the sensed portion of the cloud play an important role in this respect.

In addition, off-nadir slanted observing geometries may introduce important geometrical distortionsin the rain cell localization at surface level, as schematically sketched in figure 1.13. At the frequen-cies of interest, the source of the (scattering) signal are not the surface rainfall layers, but the icelayers aloft. At nadir, the surface rainfall can be assumed co-located with the BT scattering signal1. However, when observing through slanted views, significant displacements may exist between thelocation of the scattering source location and the effective location of the convective rain at surfacelevel. This displacement is often referred to as the parallax error, and it increases proportionally withthe altitude of the cloud and with the observation angle. Theoretically, a graupel cloud situated at10 km height, could be located almost 10 km away from the real surface rain rate location, if sensedat maximum scan edge.

A straightforward correction could be therefore attempted once the corresponding cloud top heightis supposed to be known. However, from using AMSU brightness temperatures, there is no meanto derive this piece of information2.Therefore an automatic correction for such a displacement is not

1This is true only in the hypothesis that the convective towers extend vertically under the sensor field of view. However,horizontal shears may introduce possible shifts and generate tilted events (see for instance [7] about a tilted convectiveevent observed from the TRMM Precipitation Radar)

2Or in any case without introducing significant errors. As an example, the possibility to use SEVIRI-derived cloud topheights has been discarded due to the fact that the cloud top height sensed by the IR channels might in general be much

20

deemed possible: estimating the height of the main source of the scattering signal (i.e. the centerof gravity of the generalized weighting function) seems not reasonable, due to the complexity andthe uncertainties related to the characterization of the micro and macro-physical properties of theprecipitating cloud.

Another scan-dependent source of error is due to the calibration of the instrument itself: NOAAscientists have observed an asymmetric behavior of the sensed radiances across the scan lines whichamounts to more than 5K (http://www.orbit.nesdis.noaa.gov/smcd/spb/n18calval/calval/mhs.html#mhsgeo, http://www.orbit.nesdis.noaa.gov/smcd/spb/mirs/validation/characterization.htrr).

The main impact, however, is due to the fact that the dimensions of the sensed area vary by more than200% from nadir to scan edges. According to [11], the relationship between EFOV and scan angle,both in cross-track and along-track direction, can be modeled through the following interpolatingformula :

EFOV (across− track) = 79.08 + 2.84m− 14.78m0.66

EFOV (along − track) = 28.72− 0.9m + 0.094m1.5

Where m is the scanning position.

The cross-track EFOV passes from 15km at nadir to more than 50km at scan edge.

Within such a variation, the heterogeneity of the sensed scene (and the possibility that the observedprecipitation event ”fills” the EFOV homogeneously - the so called ”beam filling” problem, see forinstance [8]) augments considerably. Everything that falls inside the sensed area is integrated andfiltered in the sensor view, but the significance of the resulting information decreases with the het-erogeneity of the observed scene. At scan edges, the probability to filter out and even loose narrowisolated sub-pixel convective cells augments considerably.

From all the above discussion, it should be clear that an increasing degree of uncertainty should be as-sociated to AMSU measurements, depending on the scan angle: as a consequence, rainfall estimationsat nadir are more reliable than the ones located along scan edges.

Instead of attempting risky (and subjective) corrections to the estimates, we propose that a scan-dependent confidence index should associated to AMSU measurements, being maximum at nadir,and decreasing with the scanning angle. The confidence index should be taken into account whenevaluating the performances of AMSU-derived rainfall rates.

higher than the source of scattering signature characterizing the microwave frequencies, and also to the time-shift that inany case would exist between the two observations.

21

Chapter 2

Convection Detection from SEVIRI data

2.1 Data description

SEVIRI is a scanning radiometer which operates on Meteosat Second Generation (MSG). It providesdata in Visible, Near Infrared and Infrared channels. Full spatial resolution in 12 spectral channels.Nominal Coverage includes Europe, Africa and locations with satellite elevation greater than or equalto 10◦. The IR channels are designed with 3 narrow band detector elements per channel, to scan theEarth every 3 km at the sub-satellite point. The High Resolution Visible (HRV) channel providesmeasurements with resolution of 1 km.

The full Earth disc image is obtained after 1250 scan line steps (south - north direction) of 9 km SSPper line step. The satellite spins at 100 rpm allowing to complete (east - west direction) a full imagein about 12.5 min. The Earth observation repeat cycle is of 15 minutes. These real-time data areprocessed to Level 1.5, i.e. are corrected for radiometric and geometric non-linearity, before onwarddistribution to the user. All SEVIRI data are available through EUMETCast.

2.2 NEFODINA

2.2.1 Algorithm description

NEFODINA is an automated tool developed at the Italian Meteorological Service an automatic tool,named NEFODINA, has been developed to detect convective cloud systems, to monitor their life cycleand to forecast their development. The product, using a varying threshold method in infrared window10.8 µm and absorption channels 6.2 and7.3 µm allows to individuate inside the cloudy cluster all theConvective Object (CO) with a top BT lower than a temperature threshold fixed at 236 K.

The detection method relies on the following basic assumptions:

• the temporal and spatial satellite data sampling is compatible with the corresponding scales ofthe phenomena;

• the evolution of the cloud top temperature joined with the water vapor amount in the mediumand high troposphere represent a good tracer of convective cell;

• it is possible to represent the life cycle of the convective cell with a linear combination of cloudtop temperature and the water vapor amount in the high troposphere.

22

Figure 2.1: NEFODINA Flow Diagram

23

The NEFODINA flow chart is reported in figure 2.1.

Validation on the detection efficiency of NEFODINA was performed, following the idea that a COduring its life has an electric activity. Often this happens during the mature stage. The validation hasbeen based indeed on the observation of lightning measured by the LN of the IAFMS during the lifeof the CO and it was conducted on a set of 12000 data uniformly selected along one year of MSG data.Probability Of Detection (POD) and False Alarm Rate (FAR) were obtained respectively equal 0.84and 0.17. 60 % of the CO were detected and classified as convective by NEFODINA 30-45 minutes(2-3 MSG slot) before any electric activity was measure by LN

2.2.2 Algorithm Improvement

The parental relationship applied to identify the daughter and the mother objects in two consecutiveslots, is based on a cross correlation between the CO detected at time t and the COs detected at timet − 1. The parental function is evaluated minimizing a weighted distance function d on the positionof the center of gravity d1, minimum temperature d2, modal temperature d3:

The CO with a distance d from a CO´, detected at the previous slot, less than a radius Rmax is saidto be a parent of CO´. A statistical study on the distances of the position of the center of gravity,minimum temperature and modal temperature in two consecutive slots was performed on the casestudies selected. The new Rmax evaluted on the bases of these statiscs allowed to increse the capacityto trak the same convective object in different slots.

2.2.3 Product description

The product is a Portable Network Graphic (PNG) image of the last available SEVIRI infrared(10.8 µm) image where the detected cells, their development and their tracking are color coded togive a quick overview to the forecaster (Fig. 2.2). This output image is associated to an ASCII filewhere the minimum, medium and modal BT of the 10.8, 6.2 and 7.3 µm channel is reported withshape, slope area and other information.

2.3 Errors and uncertainties

2.3.1 Parallax Error

The concept of parallax error (i.e. the mislocation, on the Earth surface, of rain cells due to slantedsatellite observations) has been already introduced in 1.5.2 for AMSU. However, even SEVIRI rainproducts may be affected by such a problem, as already noticed by Vicente et al. [6]. In fact, theSEVIRI camera is observing the Earth with a spherical geometry centered in the Gulf of Guinea(approximately at 0 degrees latitude and longitude). The sensor viewing angle increases with theradial distance from the nadir point. The viewing angle increases as function of the geographicallocation, introducing potential distortions and pixel enlargement (the effective pixel size increasesradially from the nadir point)(see for complete details about SEVIRI [17].

For the purposes of our study, the SEVIRI slanted observing geometry may have an impact in thatthe location of the observed signal source (cold pixels associated to convective events and possiblyto the largest cloud top heights) may be shifted by several kilometers with respect to the effectivelocation of the precipitation at the surface. Therefore, the convective pixels are relocated so that arein correspondence with the effective location of the surface rain. From [17], we computed such a

24

Figure 2.2: Example of NEFODINA product at 02 : 00 UTC on 16 08 2006. The colorbar shows thecloud top temperature. The red pixels show increasing CB instead the pink pixels show the decreasingCB.

25

Figure 2.3: NEFODINA 14-09-2006 at 02:00 UTC.

26

Figure 2.4: Parallax Correction: the colored circles represent radar data interpolated and re-projectedon the fixed SEVIRI grid, the red crosses are SEVIRI FOV characterized by convection before paral-lax correction, and the green circles are the same SEVIRI FOVs after parallax correction.

displacement as function of the latitude and longitude of the observation (radial distance from nadirsub-satellite pixel). As an example, for a 10 km height cloud top in the Mediterranean, we obtain adisplacement as large as 7 km in the radial direction. It must be highlighted, though, that this calcu-lation relies on the assumption that the observed event is perfectly vertical and that horizontal shearsare minimal. Correcting SEVIRI precipitation products for the parallax did not account for the wholemisplacement of the convective cell location with respect to the region of high radar precipitation.However the correction did lead to improved geolocation with respect to radar data (figure 2.4)

27

Chapter 3

Convective Precipitation Retrieval fromcombined AMSU and SEVIRI data

3.1 Algorithm description

In the proposed approach AMSU PC data are co-located with NEFODINA COs. The basic idea is toidentify, for each CO detected in the closest SEVIRI time slot, the AMSU FOVs affected by convec-tion and use the corresponding AMSU precipitation rates to calculate the Mean Precipitation associateto the CO. The estimated MEan CO PREcipitation (ME-CO-PRE), Φ, is computed as follows:

Φ =

∑Ni=1 RRi ∗ Ai∑M

j=1 aj

(3.1)

where:

• RRi is the AMSU derived Rain Rate for the j-th;

• Ai is the area of the i-th AMSU FOV;

• i goes from 1 to N with N being the number of AMSU FOVs affected by the convective object;

• aj is the area of the j-th SEVIRI FOV;

• j goes from 1 to M , with M being the number of SEVIRI FOVs within the convective object.

In summary, Φ represents the precipitation that would characterize the CO if all the precipitation ob-served in the corresponding AMSU FOVs was confined to the convective region only. This assump-tion represents an approximation, and it is intended to provide an upper bound to the precipitationassociated to the CO.

In its current implementation the algorithm calculates the AMSU IFOVS according to [11] and itaccounts for the parallax error on the SEVIRI FOVs (section 2.3.1), however it does not correct forAMSU geolocation errors (section 1.5.1). The algorithm also generates RR values for each SEVIRIpixel identified as convective by NEFODINA according to the following expression:

φi,j =RRi,j ∗ Ai,j

∑Lk=1 ak

(3.2)

where:

28

• RRi,j is the Rain Rate for the AMSU FOV in the i-th row and j-th column of the satellite swath;

• Ai,j is the Area of the AMSU FOV in the i-th row and j-th column of the satellite swath;

• ak is the area of the k-th SEVIRI convective FOV within the AMSU Ai,j FOV, with k goingfrom 1 to the total number of convective pixels, L, within the AMSU FOV.

The single pixel estimate φ is in general prone to geolocation and co-location errors. If fact if, dueto a small geolocation error, a single SEVIRI pixel belonging to a broad CO which densely populatean AMSU FOV, is erroneously located within a different AMSU FOV, its rain rate will be muchhigher with respect to the average of the CO. Since even small errors in the AMSU geolocation candetermine whether a SEVIRI FOVs belongs to one (less densely populated) or another (more denselypopulated) AMSU FOVs, this estimate might vary significantly within the same CO, and it is to beused in a careful way.

3.2 Product description

The co-location product consists of a CO MEan PREcipitaiton value which define the mean RR fora NEFODINA CO. It also consists of estimates of RR for individual SEVIRI FOVs belonging to aNEFODINA CO. The first product (ME-CO-PRE) is more robust to geolocation and parallax errors,while the second product is generally prone to these errors and provides highly variable estimateseven within the same CO.

29

Chapter 4

Product Validation

This chapter describes the in-situ observations (radar and rain gauges) used in the validation processand introduces the approaches used for qualitative and quantitative validation of the original PC andthe ME-CO-PRE products. Due to the limited duration of the VS project no extensive quantitativevalidation has been performed on the improved version of the PC algorithm neither on the ME-CO-PRE product.

4.1 In-situ Data

4.1.1 RADAR

The Italian Department of Civil Protection (DPC), through the network of the Regional FunctionalCenters, collects in real time/near real time precipitation products and data as radar data for the hydro-meteorological monitoring and management. The Italian system for early warning has been developedin order to issue warning , with a minimum of false alarm.

The Meteorological Service of the Air Force (CNMCA) produces national mosaic of rainfall intensityon a grid of 1400x1400 km2 by with a spatial resolution of 2.5 km2 and a time sampling of 30 min.This product (see Fig. 4.1) covers mainly North and Center of Italy. It is worth emphasizing that thenational SRI (Surface Rainfall Intensity), at the moment, is not validated with rain gauges.

The Italian SRI mosaic, is composed by several radars managed directly by Regional Government,the Italian Air Force and some foreign countries such as France, Swiss and Slovenia (see Fig. 4.2),however at the moment Italian radar network data are used. Characteristics of the most part of theseradar systems have been designed, in order to form a modern network able to operate with Doppler &polarimetric quantities. The most important technical features are:

• C band

– Frequency: 3.900− 5.750 GHz

– Wavelength: 7.69− 5.20 cm

• Doppler capacity up to 125 km;

• Operation with polarimetric quantities (e.g. dual polarization);

• Antenna with characteristics compatible with polarimetric observations;

30

Figure 4.1: RADAR coverage

• Beam width = 1o;

• Radar system remotely controlled with 24-hour operability.

4.1.2 RAIN GAUGES

Rain gauge data acquisition and processing are performed at different temporal intervals rangingfrom 5 to 30 min. The observation are collected by the National Centers and Cumulate Maps ofprecipitation are made available within 45min from the acquisition. Figure 4.3 shows the distributionof the rain gauges over Italy and provides example of the available products. Technical details aboutthe rain gauges can be found (insert reference). At the DPC an automatic quality check is applied tothe rain gauges data.

4.2 Grids remapping

In order to perform the inter-comparisons among the various products, several procedures have beenimplemented for remapping data:

• from AMSU to SEVIRI grid,

• from RADAR to AMSU grid,

• from RADAR to SEVIRI grid,

31

Figure 4.2: RADAR National Mosaic

Figure 4.3: Rain Gauge distribution

32

• from GAUGES to SEVIRI.

In particular the SEVIRI grid has been chosen as reference, since it provides a fixed grid at a conve-nient resolution.

4.2.1 The AMSU-to-SEVIRI remapping process

This process was designed to allow the comparison of current PC products with SEVIRI-derivedrainfall products, and with RADAR and RAIN GAUGE data on a fixed grid. However it was alsoused to compute the background BTs for the PC algorithm (as described in Section 1.2.2).

The basic concepts of the implementation are hereafter described.

Given the location of a SEVIRI pixel, the bounds of the corresponding AMSU pixel are found. routinebegins by determining if it lies within the bounds of the AMSU grid, and if so, then searches for theclosest AMSU observation. (Incidentally, this is done in "brute-force" fashion by cycling through rowand column do-loops, and is not particularly efficient; it would be possible to speed up this part of thealgorithm). Having found the nearest AMSU, the routine then looks to the left and right (on the sameAMSU row) to find which of these two is closer to the SEVIRI, and then the closest and its neighborto the left or right are chosen as two of the four AMSUs surrounding the SEVIRI pixel. A check isperformed to make sure the selected pixel is not on the lateral edge of the AMSU array, so that itmakes sense to look left or right. If the pixel is on the edge, then the algorithm simply uses the twopoints on the row nearest the edge. A weight is assigned to these two based on their distance from theSEVIRI pixel. The distances are determined by computing the angle between the SEVIRI positionvector, and either of the two chosen AMSU position vectors. This part of the algorithm, using vectoralgebra, is quite efficient.

Once the distances are calculated there are two different strategies to derive the BT on the SEVIRIgrid. The first one makes simply use of the nearest neighbors, while the second interpolates (weightedaverage) among the four closest AMSU FOVs according to the following procedure: returning to thepreviously-found nearest AMSU, the routine looks at the two AMSU points above and below, inadjacent rows, to find which of these two is closer to the SEVIRI point. The chosen row (above orbelow the row in which the closest AMSU lies) is then used, and the four points used for interpolationare the two found in the first row, and the two in the next-best row.

An interpolated estimated is made in each row (Wi) between the two chosen points on that row, usingweights

Wi1 = 1.− s1/(s1 + s2) (4.1)

Wi2 = 1.−Wi1

where s1 and s2 are the distances (in latitude degrees) between SEVIRI and the first and secondclosest AMSU points on that row. In the above case, if s1 = 0 (i.e. an AMSU coincides withSEVIRI), then that AMSU has weight 1 and its neighbor has weight 0. A similar weighting schemeis then used vertically (Wj), along the AMSU column containing the closest AMSU, to arrive at theweights which are used to interpolate as many of the AMSU channels as are needed to the SEVIRIlocation.

Wj1 = 1.− d1/(d1 + d2) (4.2)

Wj2 = 1.−Wj1

33

Figure 4.4: Example of PC remapping onto SEVIRI grid.

where d1 and d2 are distance of two AMSUs along the minimal column from SEVIRI.

The horizontal weights, and the vertical weights, are then used to average the AMSU BTs (or anyother quantity defined on the AMSU grid) on the two rows (e1 and e2), :

e1 = Wi1 ∗BT (1, 1) + Wi2 ∗BT (2, 1) (4.3)

e2 = Wi1 ∗BT (1, 2) + Wi2 ∗BT (2, 2)

e = Wj1 ∗ e1 + Wj2 ∗ e2

to get the final estimate e of the BT.

The co-location scheme based on the nearest neighbors has been used for the derivation of backgroundBTs (as described in section 1.4), while the second one (based on weighted averages) has been usedfor validation purposes to keep consistence with the results presented in the first VS report.

4.2.2 The RADAR-to-AMSU remapping process

Radar data was convolved to the AMSU footprint using the methods described in [10, 13, 14]. Themethods used in this study were initially derived for Baltex Radar Data Center (BRDC) compositesbut were adjusted to account for the radar composites provided by CNMCA. The convolution takesinto account the actual spatial sensitivity of AMSU-A and AMSU-B as outlined in [11]. A fixed Z-R

34

Figure 4.5: RADAR-to-SEVIRI remapping scheme.

relation of Z = 200R1.6 was used in this study. Due to missing information about the actual positionof the radar in the composite imagery, a parallax correction could not be performed. Also, the radardata used in this study were not gauge adjusted.

4.2.3 The RADAR-to-SEVIRI remapping process

The observations available for Surface Rain Intensity (SRI) are on a different scale (spatial resolutionand projection) compared to satellite grids and since the variability of precipitation fields stronglydepends on the scale at which the fields are considered a meaningful comparison is not trivial. The up-scaling technique (fine to course resolution) used to remap radar data onto MSG grid, here described,is very simple but numerically effective. National mosaic of SRI generated by Radars and MSGproduct are composed of two static grids, each radar cell has been linked to the SEVIRI pixel whichcontain the center of radar pixel. Therefore radar data are remapped onto geostationary grid throughthe mean value of SRI calculated on radar cells linked to each satellite grid (fig. 4.5)

4.2.4 The RAIN GAUGE-to-SEVIRI remapping process

The re-mapping of rain gauges data on the SEVIRI grid was done using the Thiess Polygons andconsidering the precipitation constant inside each polygon.

4.3 Validation

4.3.1 Qualitative Comparison: PC-GAUGES

Fig. 4.10 shows that the comparison between PC weighted average RR derived from AMSU and therain gauge observations, at this stage, does not provide conclusive evidence of the PC accuracy. Thisis due to the following considerations:

• too many gauge observations are suspiciously reported to be zero even where AMSU (as RADAR)sees precipitation;

35

Figure 4.6: Example of RADAR derived Rain Rate on SEVIRI grid.

36

Figure 4.7: Example of RADAR derived precipitation classes on SEVIRI grid.

37

Figure 4.8: Example of RAIN GAUGES 30 min. cumulated rain.

38

Figure 4.9: Example of RAIN GAUGES 30 min. cumulated rain binned according to PC classes.

39

Figure 4.10: PC weighted average on AMSU native grid at 01 : 52 UTC over the Gulf of Genova.AMSU derived precipitation is described by the values next to the colorbar. The colored circles repre-sent rain gauges observations. RR intensity estimated from 10 min cumulate precipitation measuredby the gauges at 02 : 00 UTC, is described by the values in parenthesis next to the colorbar.

• for a fair comparison, the rain gauges observations should be convolved to the AMSU FOVusing the same convolving scheme used for RADAR data;

• the impact AMSU geolocation uncertainty should be carefully estimated.

Those issues will be further investigated by inter-comparisons of AMSU, RADAR, and GAUGESestimation in the following section. The 10 min (cumulated rain) temporal resolution of the raingauge observations, and their dense spatial distribution, make them potentially useful for qualitativeand quantitative estimate of the PC accuracy.

4.3.2 Qualitative Comparison: RADAR-GAUGES

Since RAIN GAUGES and RADARS represent the main validation instruments for satellite de-rived precipitation estimates, it has been considered worth to evaluating how they compare to eachother. This section describes the comparison of RADAR data on SEVIRI grid and individual RAINGAUGES. Fig. 4.12, and 4.13 show some of the inherent difficulties in using these kind of observa-tions, especially for quantitative estimations. Some of the issues summarized in Sec. 4.3.1 are showedto be present even when comparing those two ground based instruments. In particular:

40

Figure 4.11: PC weighted average on AMSU native grid at 02 : 35 UTC over the Gulf of Genova.AMSU derived precipitation is described by the values next to the colorbar. The colored circles repre-sent rain gauges observations. RR intensity estimated from 10 min cumulate precipitation measuredby the gauges at 02 : 30 UTC, is described by the values in parenthesis next to the colorbar.

41

Figure 4.12: RADAR estimated RR on AMSU native grid between 02 : 00 and 02 : 30 UTC overthe Gulf of Genova. RR is derived precipitation is described by the values next to the colorbar. Thecolored circles represent rain gauges observations. RR intensity estimated from 10 min cumulateprecipitation measured by the gauges at 02 : 20 UTC, is described by the values in parenthesis nextto the colorbar.

• too many gauge observations are suspiciously reported to be zero even where RADAR (asAMSU) sees precipitation;

• for a fair comparison, the rain gauges observations should be convolved to the AMSU FOVusing the same convolving scheme used for RADAR data;

• often the RADAR temporal resolution (30 min) does not allow for a fair comparison with10 min cumulated rain observed by the gauges;

• in the comparison the distance of the RADAR observation from the actual RADAR coordinatesshould be taken into account as the error on the RADAR derived RR depends on this distance.

4.3.3 Quantitative Comparison PC-RADAR on AMSU grid

The only quantitative validation of the original PC product was done for about 43000 observationstaken in 7 different days by the NOAA 16 and 18 satellites. It was performed using RADAR dataconvolved on the AMSU grid. The results were compared in the following two ways:

42

Figure 4.13: RADAR estimated RR on AMSU native grid between 02 : 30 and 03 : 00 UTC overthe Gulf of Genova. RR is derived precipitation is described by the values next to the colorbar. Thecolored circles represent rain gauges observations. RR intensity estimated from 10 min cumulateprecipitation measured by the gauges at 02 : 50 UTC, is described by the values in parenthesis nextto the colorbar.

43

1. AMSU precipitation classes (max probability approach) were compared to RADAR derivedclasses of precipitation, pixel by pixel on AMSU native grid;

2. AMSU recalculated classes (weighted average approach) were compared to RADAR derivedclasses of precipitation, pixel by pixel on AMSU native grid;

(a) AMSU derived RR is calculated using the weighted average of the likelihoods. A meanRR value was assigned to each class (as described third point of section 1.3)

(b) weighted averages of RR were classified again according to Tab. 1.3. This process wasrequired because the weighted average is an arbitrary estimate (in particular the meanvalue of the intense precipitation class is completely arbitrary), and the one based on themaximum likelihood does not take into account the probability distribution among thedifferent classes but only the maximum probability;

(c) RADAR data were convolved on AMSU grid and classified according to Tab. 1.3;

(d) The contingency tables were calculated for all the selected overpasses.

These procedures do not take in account several of the issues discussed in the previous section butstill provide some useful information on the performances of the PC algorithm. It looked evident, infact, that:

• In both cases (Max Probability and Weighed average) PC misclassified several pixel assigningthem to class 2 and 3 (.1 < RR < 5 [mm/h] ) while according to the radar the pixels were be-longing to class 1 (no precipitation) as showed in tbl 4.3. This misclassification issue appearedto be more evident for daytime passes (tbl.s 4.1) than for nighttime ones (tbl.s 4.2). The problemis very likely to due the combination of two effects: 1) the background BT calculation are notcalculated correctly especially over coastal warm regions); 2) the algorithm was calibrated in aNorthern European climate regimes, therefore it should be expected that it would not performoptimally in Southern Europe, especially daytime, summer, over the warmest part of Italy.

• Several pixels were classified class 4 (RR > 5 [mm/h]) by PC while according to the radarthey were belonging to class 3 (.5 < RR < 5 [mm/h]). In addition to the fact the radar, withthe distance, tends in general to underestimate rain rate intensity, this problem is also clearlyrelated to the arbitrary mean value associated to class 4 in the intermediate step for the PCproduct when the weighted average RR was calculated.

Both the issues are well understood and were expected from the application of the PC algorithm toItalian areas.

For the described validation exercise the results obtained were, on average, 18% worse, with a peakof 58% for class 4, with respect to those obtained by Bennartz in [12]. The discrepancies can beexplained considering that:

• the radar date used in the validation were not gauge adjusted, underestimation of precipitationfrom radar data was therefore expected;

• the algorithm was trained only with North European radars and for high latitude climate regimes;

• in his validation Bennartz discriminated between land and water pixels and excluded coastalregions, which were less relevant over Northern Europe, than for the Italian case;

44

PC/RADAR R1 R2 R3 R4

C1 93.69 (95.02) 4.74 (3.86) 1.52 (1.10) 0.05 (0.01)C2 66.05 (93.66) 21.21 (4.03) 12.15 (2.11) 0.73 (0.20)C3 23.66 (63.33) 36.62 (19.79) 37.22 (16.03) 1.85 (0.84)C4 3.40 (0) 16.54 (5.81) 68.98 (71.32) 5.82 (22.86)

Table 4.1: Contingency table for daytime only for the max probability approach and for the weightedaverage approach (in parenthesis) . PC classification are on the rows while RADAR estimationsare on the columns.

PC/RADAR R1 R2 R3 R4

C1 92.47 (93.52) 5.34 (4.70) 2.17 (1.76) 0.01 (0.01)C2 53.50 (82.75) 28.78 (13.55) 16.99 (3.70) 0.73 (0.00)C3 21.69 (50.87) 37.03 (26.24) 39.41 (21.70) 1.85 (1.18)C4 13.60 (12.10) 17.67 (12.81) 62.91 (69.39) 5.82 (5.69)

Table 4.2: Contingency table for night-time only for the max probability approach and for theweighted average approach (in parenthesis) . PC classification are on the rows while RADAR es-timationsare on the columns.

• over coastal (and in general heterogeneous) and non vegetated pixels the current scheme forbackground temperature estimation does not provide optimal results, this issue appears to bemore evident over Italy;

• class 4 pixels over Northern Europe are much less frequent than over Southern Europe, thereforethe data set used to calibrate the algorithm might not, at this stage, enough representative forthe Italian cases.

It is worth mentioning that overall the radars detected 37044 FOVs with no precipitation (C1), 3427of C2, 2464 of C3, and 171 of C4; PC detected 36740 FOVs with no precipitation (C1), 3955 of C2,1219 of C3, and 1192 of C4.

4.3.4 Qualitative validation of NEFODINA/AMSU combined products

The merging of PC and NEFODINA has a key role in the VS activity because:

• for convective systems, it allows to determine if the intense precipitation, as observed byAMSU, actually occurs in regions were convection is active;

PC/RADAR R1 R2 R3 R4

C1 93.11 (94.28) 5.01 (4.28) 1.82 (1.42) 0.03 (0.01)C2 62.14 (88.89) 23.56 (8.20) 13.65 (2.8) 0.63(0.10)C3 23.05 (59.79) 36.75 (21.62) 37.90 (17.63) 2.29 (0.94)C4 7.80 (6.03) 17.03(9.46) 66.35 (70.31) 8.81 (13.91)

Table 4.3: Contingency table for day and night time for the max probability approach and for theweighted average approach (in parenthesis) . PC classification are on the rows while RADAR estima-tions are on the columns. Overall the radars detected 37044 FOVs with no precipitation (C1), 3427of C2, 2464 of C3, and 171 of C4; PC detected 36740 FOVs with no precipitation (C1), 3955 of C2,1219 of C3, and 1192 of C4.

45

Figure 4.14: PC weighted average on AMSU native grid at 01 : 52 UTC and on SEVIRI FOVs.

• it opens the way to a merging of geostationary and polar orbiting products which could allowfor the identification of highly precipitating regions at SEVIRI spatial resolution (3 to 8 timesfiner than the native AMSU resolution), and under well defined approximations it could provideupper bound estimates of the Rain Rate, within the convective regions.

The comparison hereafter presented is based on the co-location of CO, as detected by NEFODINA ,on the AMSU PC product (Fig. 1.7). The selected case study for this comparison is 16 Aug. 2006.Fig. 4.14 shows the PC weighted average for the AMSU overpass NOAA-18 at 01 : 52 UTC. Dif-ferent convective systems are active in the scene, but for sake of clarity only the system active in theproximity of the Gulf of Genova was selected to demonstrate the concept. The figure shows the PCweighted average RR for the AMSU overpasses, NOAA-18 at 01 : 52 UTC (large FOVs) along withthe smaller SEVIRI FOVs belonging to a CO. For every overpass the NEFODINA detected convectiveobjects co-locate nicely with the highly probable (P ' 1) class 4 (RR > 5 mm/hr) FOVs charac-terized by an weighted average rain rate of about 10 mm/hr. Under the assumption that most of theprecipitation occurs in the highly convective region, NEFODINA provides relevant information aboutthe distribution of precipitation within the AMSU FOVs. For this particular case the NEFODINA de-tected convection cell has a mean precipitation (according to eq. 3.1) of 47mm/hr while uncalibratedradar data provide a precipitation value of about 42 mm/hr. In this study NEFODINA was proved tobe an ideal source of information to couple with the PC product because, for convective systems: itprovides evidence of probable precipitation, and, even more important, it provides information withregards to the developmental stage of the precipitating convective system.

46

Figure 4.15: RADAR RR interpolated on SEVIRI grid 02 : 00 UTC.

47

4.4 Selected Case Studies: data distribution and visualization

Detailed description of the synoptic situation for the selected case studies is presented on the H-SAFVS web page at: http://lupin.ssec.wisc.edu/~paoloa/HSAF/VS.html . Quicklooks of AMSU observa-tions, derived products, different re-mapping of in-situ observations are also published on the sameweb site. Access to the site is granted to H-SAF members upon e-mail request to paoloa@ssec.wisc.edu.

48

Chapter 5

Conclusions and recommendations

5.1 Conclusions

This VS activity has shown an important relation between the precipitation estimates derived by PCand COs detected by NEFODINA. This relation opens the way to a potentially fruitful merging ofgeostationary and polar orbiting products which could improve the accuracy and the usefulness ofthe AMSU derived precipitation products to hydrological and QPF activities. The case study ofthe 16 August 2006 here reported showed the good accordance between heavy rainfall (Class 4:RR > 5mm/hr) deduced by PC and convection detected by NEFODINA and showed also the basicconcept of the proposed merging strategy. For this study, the PC algorithm [14] developed by RalfBennartz within NWC-SAF was selected as the most suitable AMSU-based precipitation retrievalalgorithm for H-saf VS project. The selection was based on pre-operational software availability,knowledge of performance characteristics over a large part of H-saf area (Northern Europe), use ofAMSU channels more sensitive to heavy rainfall. The PC algorithm outputs was extended to thewhole H-saf area allowing for testing and validation in the Mediterranean regions (Italy). Improve-ments in the performances of the algorithm were obtained through a refinement of the estimation ofthe radiative background field. AMSU data from NOAA 15,16, 17 and 18 (and METOP after launch)were archived starting Feb. 1st 2006. The precipitation observations from radar and rain gauges werecollected for the selected Italian convective cases and up scaling and down scaling algorithms weredeveloped to compare precipitation ground-based observations with PC algorithm retrievals and SE-VIRI products. The data collected with the synoptic analysis are available to those who are involvedin H-saf activities at (http://lupin.ssec.wisc.edu/paoloa/HSAF/VS.html). The validation, on more than43000 observations taken in 7 different days by the NOAA 16 and 18 satellites, was performed usingRADAR data convolved on the AMSU grid. The results obtained for the convective events stud-ied appeared to be different from those over North Europe. Initial comparison results indicate thatthe background brightness temperature tuning that was originally developed for northern Europe fre-quently produces spurious rainfall signatures, especially in arid regions during daytime. The resultsobtained applying the PC algorithm are on average 18%, with a peak of 58% for class 4, worse thanthose obtained by Bennartz in [12]. This was probably due to a combination of effects such as differ-ent climate regimes (North and South Europe), lack of severe events in the current calibration data set,and limitations in the calculation of the background temperature especially in coastal regions. Dueto the limited duration of the VS project, no quantitative validation was performed on the improvedversion of the PC algorithm neither on the ME-CO-PRE product.

49

5.2 Recommendations

The use of Italian radar and rain gauge network is encouraging, so, for the future activities, it is recom-mended to re-calibrate the PC algorithm using both radar and rain gauge data over Southern Europe.In the same way the merging of NEFODINA with the PC product was demonstrated to be potentiallyuseful in providing information about the precipitation spatial distribution within the AMSU pixel,and the temporal development of the convective systems. It is therefore recommended to further testthis strategy, with extensive validation, to fully exploit the benefits of geo and leo observation cou-pling. Finally it is recommended to further develop or adopt of a true rain rate retrieval, taking thefeasible steps towards the goal of a future fully operational implementation.

50

Contents

1 Precipitation retrieval from microwave data 4

1.1 AMSU Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Precipitation Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.1 Algorithm Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.2 Algorithm Theoretical Basis . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3 PC products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.4 Improvements to the PC algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.5 Errors and uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.5.1 AMSU geolocation uncertainties . . . . . . . . . . . . . . . . . . . . . . . . 16

1.5.2 Scan-dependent uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Convection Detection from SEVIRI data 22

2.1 Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.2 NEFODINA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.2.1 Algorithm description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.2.2 Algorithm Improvement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.2.3 Product description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3 Errors and uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3.1 Parallax Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Convective Precipitation Retrieval from combined AMSU and SEVIRI data 28

3.1 Algorithm description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2 Product description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4 Product Validation 30

4.1 In-situ Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.1.1 RADAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.1.2 RAIN GAUGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.2 Grids remapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.2.1 The AMSU-to-SEVIRI remapping process . . . . . . . . . . . . . . . . . . 33

51

4.2.2 The RADAR-to-AMSU remapping process . . . . . . . . . . . . . . . . . . 34

4.2.3 The RADAR-to-SEVIRI remapping process . . . . . . . . . . . . . . . . . . 35

4.2.4 The RAIN GAUGE-to-SEVIRI remapping process . . . . . . . . . . . . . . 35

4.3 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.3.1 Qualitative Comparison: PC-GAUGES . . . . . . . . . . . . . . . . . . . . 35

4.3.2 Qualitative Comparison: RADAR-GAUGES . . . . . . . . . . . . . . . . . 40

4.3.3 Quantitative Comparison PC-RADAR on AMSU grid . . . . . . . . . . . . 42

4.3.4 Qualitative validation of NEFODINA/AMSU combined products . . . . . . 45

4.4 Selected Case Studies: data distribution and visualization . . . . . . . . . . . . . . . 48

5 Conclusions and recommendations 49

5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

52

Bibliography

[1] P. Bauer. Over-ocean rainfall retrieval from multisensor data of the tropical rainfall measur-ing mission. part i: design and evaluation of inversion databases. J. Atmos. Oceanic Technol.,18(8):1315–1330, 2001.

[2] R. Bennartz and G.W. Petty. The sensitivity of microwave remote sensing observations of pre-cipitation to ice particle size distributions. J. Appl. Meteorol., 40:345–364, 2001.

[3] Kummerow C. and L. Giglio. A passive microwave technique for estimating rainfall and verticalstructure information from space, part i: Algorithm description. . J. Appl. Meteor., 33:3,18, 1994.

[4] F.W. Chen and D.H. Staelin. Airs/amsu/hsb precipitation estimates. IEEE Trans. On Geosci.and Remote Sensing, 41:410,417, 2003.

[5] Ferraro et al. The current status of passive microwave precipitation retrievals at noaa/nesdis.Proceedings of SPIE Vol. 5654, 2004.

[6] J.C.Davenport G.A.Vicente and R.A. Scofield. The role of orographic and parallax correctionson real time high resolution rainfall rate distribution. International Journal of Remote Sensing,pages 221–230, 2007.

[7] J. Hafermann W.S. Olson Hong., Y. and C.Kummerow. Microwave brightness temperaturesfrom tilted convective systems. J. Appl. Meteor., 39:983–998, 2000.

[8] C. Kummerow. Beam-filling errors in passive microwave rainfall retrievals. J. Appl. Meteor.,pages 356–370, 1998.

[9] G. Panegrossi, S. Dietrich, F.S. Marzano, A. Mugnai, E.A. Smith, X. Xiang, G. J. Tripoli, P.K.Wang, and J.P.V. Poiares Baptista. Use of cloud model microphysics for passive microwave-based precipitation retrieval: significance of consistancy between model and measurement man-ifolds. J. Atmos. Sci., 54:1644,1673, 1998.

[10] Bennartz R. On the use of ssm/i measurements in coastal regions. J.Atmos.Ocean.Technol.,16:417–431, 1999.

[11] Bennartz R. Optimal convolution of amsu-b to amsu-a. J.Atmos.Ocean.Technol., 17:1215–1225,2000.

[12] Bennartz R. Nwc-saf vs: Final report. 2005.

[13] Bennartz R. and D. B. Michelson. Correlation of precipitation estimates from spacebornepassive microwave sensors and weather radar imagery for baltex pidcap. Int.J.Remote Sens.,24:723–739, 2003.

53

[14] Bennartz R., A. Thoss, A. Dybbroe, and D. B. Michelson. Precipitation analysis using theadvanced microwave sounder unit in support of nowcasting applications. Meteorological Appli-cations, 9(2), 2002.

[15] N.C. Grody L. Zhao H. Meng C. Kongoli P. Pellegrino S. Qiu R.R. Ferraro, F. Weng andC. Dean. Noaa operational hydrolocial products derived from the advanced microwave soundingunit (amsu). IEEE Trans. Geosci. Rem. Sens, 2004.

[16] S. Di Michele A. Mugnai F. S. Marzano J. P. V. Poiares Baptista Tassa, A. Cloud model-basedbayesian technique for precipitation profile retrieval from the tropical rainfall measuring missionmicrowave imager. Radio Sci., 38:8074, 2003.

[17] Wolf. Lrit/hrit global specification coordination group for meteorological satellites. TechnicalReport 2.6, EUMETSAT, August 1999.

54