satellite radiometric remote sensing of rainfall ﬁelds ...€¦ · 268 f. s. marzano et al.:...

Advances in Geosciences, 2, 267–272, 2005SRef-ID: 1680-7359/adgeo/2005-2-267European Geosciences Union© 2005 Author(s). This work is licensedunder a Creative Commons License.

Advances inGeosciences

Satellite radiometric remote sensing of rainfall fields: multi-sensorretrieval techniques at geostationary scale

F. S. Marzano1,2, D. Cimini1, E. Coppola1, M. Verdecchia1, V. Levizzani3, F. Tapiador4, and J. F. Turk5

1Centro di Eccellenza CETEMPS, Universita dell’Aquila, Via Vetoio – 67010, L’Aquila, Italy2Dipartimento di Ingegneria Elettronica, Universita “La Sapienza” di Roma, Rome, Italy3Istituo di Scienze dell’Atmosfera e Clima (ISAC), Consiglio Nazionale delle Ricerche, Bologna, Italy4Instituto de Ciencias Ambientales, Universidad de Castilla-La Mancha (UCLM), Toledo, Spain5Marine Meteorology Division, Naval Research Laboratory (NRL), Monterey, California, USA

Received: 11 November 2004 – Revised: 11 July 2005 – Accepted: 13 July 2005 – Published: 22 July 2005

Abstract. The Microwave Infrared Combined Rainfall Al-gorithm (MICRA) consists in a statistical integration methodusing the satellite microwave-based rain-rate estimates, as-sumed to be accurate enough, to calibrate spaceborne in-frared measurements on limited sub-regions and time win-dows. Rainfall retrieval is pursued at the space-time scale oftypical geostationary observations, that is at a spatial resolu-tion of few kilometers and a repetition period of few tensof minutes. The actual implementation is explained, al-though the basic concepts of MICRA are very general andthe method is easy to be extended for considering innova-tive statistical techniques or measurements from additionalspace-borne platforms. In order to demonstrate the poten-tiality of MICRA, case studies over central Italy are also dis-cussed. Finally, preliminary results of MICRA validation byground based remote and in situ measurements are shownand a comparison with a Neural Network (NN) based tech-nique is briefly illustrated.

1 Introduction

The problem of using satellite remote sensing data to retrieverainfall is fairly complicated since presently there is not asingle spaceborne platforms which can carry all the suitableinstruments to ensure all the properties to the rainfall prod-uct (Levizzani et al., 1995; Vicente et al., 1998; Turk et al.,1999). From a meteorological point of view, visible (VIS)and infrared (IR) radiometers can give information on cloudtop layers. On the other hand, microwave (MW) radiometerscan detect cloud structure and rainrate since MW brightnesstemperatures (TB’s) are fairly sensitive to liquid and ice hy-drometeors (Ferraro, 1997). Regarding platforms, Geosyn-chronous Earth Orbit (GEO) satellites can ensure a cover-

Correspondence to:F. S. Marzano([email protected])

age with a high temporal sampling, while Low Earth Orbit(LEO) satellites have the advantage to enable the use of mi-crowave sensors, but with a major drawback of low tempo-ral sampling. Therefore, LEO-MW and GEO-IR radiometryare clearly complementary for monitoring the Earth’s atmo-sphere and a highly variable phenomenon such as precipita-tion.

Statistical integration of satellite infrared and microwavedata can be accomplished in several ways (Kummerow andGiglio, 1995; Bellerby et al., 2000; Miller et al., 2001;Marzano et al., 2001). On one hand, there is a choice ofwhat variables (i.e., predictors) to match in order to providethe final product. A possibility is represented by the directcombination of MW TB’s and thermal IR (TIR) radiances,having the advantage to exploit the observable informationwithout any post-processing and the disadvantage to requestIR and MW measurements matched in space and time. Thelatter condition is only satisfied a limited number of times ina given area if few LEO platforms are considered thus flar-ing the statistical significance of any empirically-trained re-trieval algorithm. The feasible approach would be that basedon physically-based retrieval algorithms which, on the otherhand, would need a climatological and microphysical tuning.

In order to avoid these difficulties, one can resort to ap-proaches whose aim is to combine IR measurements andMW-based estimates on a cumulative statistical distribu-tion basis. Indeed, artificial neural network (NN) canbe conveniently applied to the same problem dealing withempirically-trained algorithms showing comparable perfor-mances (Hsu et al., 1997; Grimes et al., 2003; Tapiador etal., 2004a, 2004b).

In this work, the Microwave Infrared Combined RainfallAlgorithm (MICRA) is presented, based on the statistical in-tegration of collocated GEO-IR and LEO-MW data accom-plished on a local scale (Marzano et al., 2003, 2004) and it iscompared with a NN-based algorithm: NEREMIS (NEuralRainfall Estimation algorithm using Microwave and Infrared

268 F. S. Marzano et al.: Satellite radiometric remote sensing of rainfall fields

Marzano, Cimini, Coppola, Verdecchia, Levizzani, Tapiador and Turk, “Satellite radiometric remote sensing ….”

July 11, 2005 2

2. Microwave infrared combined rainfall algorithm The general idea behind the considered integration techniques is to combine the appealing spatial and temporal sampling of IR sensors, mounted on geo-stationary platforms (GEO-IR), with the higher accuracy of passive Low-Earth-Orbit Microwave (LEO-MW) methods for rain-rate retrieval. The statistical integration techniques are applied within a procedure which is supposed to run continuously. This procedure is based on a background process and a foreground process. A schematic sketch of the block diagram and the temporal flow chart of MICRA is shown in Figure 1, while detailed equations are described in (Marzano et al., 2003; Marzano et al., 2004).

MW MW Inversion Inversion of of Rain Rain (INV) (INV)

Space Space - - Time Time Collocation Collocation (COL) (COL) R R

Background process ( BP )

T T B B

T T IR IR

Pre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- Thresholding

IR IR - - R R Statistical Integration Statistical Integration (INT) (INT)

IR IR Retrieval Retrieval of of Rain Rain (RET) (RET)

Foreground process ( FP )

T T IR IR




T T B B

T T IR IR




T T B B

T T IR IR

Pre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- ThresholdingPre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- ThresholdingPre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- Thresholding




T T IR IR




R

T T IR IR

Fig. 1. Schematic sketch of the MICRA block diagram and flow chart. i) The first step of the background process consists: in estimating the surface rain-rate from available LEO-MW measurements by means of either empirical retrieval algorithms or inversion schemes based on parametric cloud radiative models (inversion step). This enables the inversion of a set of brightness temperatures (TB) at different frequency and polarization to provide a rain-rate product spatially integrated within the nominal area A. ii) The second step of the background process pursues the combination of LEO-MW sensor data with data coming from GEO-IR sensor in space and time on a global scale (collocation step). This step consists in temporally locating the GEO-IR data within the past few tens of minutes of the LEO-MW data time and to re-map into the geographic coordinates both GEO-IR and LEO-MW measurements available observations. Note that, since spatial resolution of MW data is generally worse than IR ones, a MW field-of-view of nominal area A generally includes more than one IR pixel. As a result of the background process, a data set is generated, containing the per-pixel rain-rate retrieved from LEO-MW data, the co-located GEO-IR brightness temperature and the pixel geo-location. This process is continuously ongoing, since new LEO-MW and GEO-IR data are continuously ingested depending on available satellite platforms. iii) The third step is a foreground process, started to derive the R-TIR inverse relationship once the data set has been updated. A pre-processing stage is accomplished after each background process. The IR retrieval relationships are

updated every time a new set of combined data have been added to the data set relative to that sub-region, and are derived using data archived in a time window of several hours (integration step). To assure that only the most recent rain history is captured and to guarantee a statistical significance of the training set, the R-TIR inverse relationship for a given sub-region is derived using only the most recent combined data. iv) The last step is represented by the prediction of the surface rain-rate from IR measurements in a given sub-region by applying the derived R-TIR algorithm (retrieval step). Many attempts have been carried out so far to derive this R-TIR relationship, using different techniques such as probability matching formulations, regression methods and artificial neural networks. Details on the background and foreground processes are given section 4. 3. Neural rainfall estimation algorithm using

microwave and infrared sensors The NEREMIS techniques shares the same methodology of MICRA, but uses a Multiply Layer Feed Forward NN architecture as shown in Fig. 2 (Grimes et al., 2003; Tapiador et al., 2004b). It is a four layers network using as input the IR TB from GEO-IR sensor and latitude and longitude information.

Fig 2 Neural Network architecture for rainfall estimation

The NN is calibrated using input data derived from the GEO-IR sensor and output rain-rate retrieved from available LEO-MW measurements. This requires, as for MICRA, measurements from both IR and microwave sensors collocated in space and in time. The calibration period has been fixed for this study to 24 hours and the validation has been carried on for the following day. In particular, the calibration periods are the 24 hours of the Jan. 23, 2003, while Jan. 24, 2003 has been retained for the validation. There is not any constraint in choosing a calibration period. The one day time step has been selected due to the length of the precipitation event that has been observed for at least tree days. In case of a short rain event the time step could be reduced to few hours.

Fig. 1. Schematic sketch of the MICRA block diagram and flow chart.

Sensors). As an application, some case studies on the areaof central Italy are discussed, comparing satellite based es-timates with ground-based remote (radar) and in situ (raingauges) observations, in order to demonstrate the potentialityof monitoring precipitation using the proposed techniques.

2 Microwave infrared combined rainfall algorithm

The general idea behind the considered integration tech-niques is to combine the appealing spatial and temporal sam-pling of IR sensors, mounted on geo-stationary platforms(GEO-IR), with the higher accuracy of passive Low-Earth-Orbit Microwave (LEO-MW) methods for rain-rate retrieval.The statistical integration techniques are applied within aprocedure which is supposed to run continuously.

This procedure is based on a background process and aforeground process. A schematic sketch of the block dia-gram and the temporal flow chart of MICRA is shown inFig. 1, while detailed equations are described in (Marzanoet al., 2003; Marzano et al., 2004).

i) The first step of the background process consists in esti-mating the surface rain-rate from available LEO-MWmeasurements by means of either empirical retrievalalgorithms or inversion schemes based on parametriccloud radiative models (inversion step). This enablesthe inversion of a set of brightness temperatures (TB) atdifferent frequency and polarization to provide a rain-rate product spatially integrated within the nominal areaA.

ii) The second step of the background process pursues thecombination of LEO-MW sensor data with data com-ing from GEO-IR sensor in space and time on a global

scale (collocation step). This step consists in tempo-rally locating the GEO-IR data within the past few tensof minutes of the LEO-MW data time and to re-map intothe geographic coordinates both GEO-IR and LEO-MWmeasurements available observations. Note that, sincespatial resolution of MW data is generally worse thanIR ones, a MW field-of-view of nominal area A gen-erally includes more than one IR pixel. As a result ofthe background process, a data set is generated, con-taining the per-pixel rain-rate retrieved from LEO-MWdata, the co-located GEO-IR brightness temperature andthe pixel geo-location. This process is continuously on-going, since new LEO-MW and GEO-IR data are con-tinuously ingested depending on available satellite plat-forms.

iii) The third step is a foreground process, started to de-rive the R-TIR inverse relationship once the data set hasbeen updated. A pre-processing stage is accomplishedafter each background process. The IR retrieval rela-tionships are updated every time a new set of combineddata have been added to the data set relative to that sub-region, and are derived using data archived in a timewindow of several hours (integration step). To assurethat only the most recent rain history is captured andto guarantee a statistical significance of the training set,the R-TIR inverse relationship for a given sub-region isderived using only the most recent combined data.

iv) The last step is represented by the prediction of the sur-face rain-rate from IR measurements in a given sub-region by applying the derived R-TIR algorithm (re-trieval step). Many attempts have been carried out sofar to derive this R-TIR relationship, using differenttechniques such as probability matching formulations,

F. S. Marzano et al.: Satellite radiometric remote sensing of rainfall fields 269


July 11, 2005 2

2. Microwave infrared combined rainfall algorithm The general idea behind the considered integration techniques is to combine the appealing spatial and temporal sampling of IR sensors, mounted on geo-stationary platforms (GEO-IR), with the higher accuracy of passive Low-Earth-Orbit Microwave (LEO-MW) methods for rain-rate retrieval. The statistical integration techniques are applied within a procedure which is supposed to run continuously. This procedure is based on a background process and a foreground process. A schematic sketch of the block diagram and the temporal flow chart of MICRA is shown in Figure 1, while detailed equations are described in (Marzano et al., 2003; Marzano et al., 2004).




T T B B

T T IR IR

Pre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- Thresholding




T T IR IR




T T B B

T T IR IR




T T B B

T T IR IR

Pre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- ThresholdingPre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- ThresholdingPre Pre - - processing (PRE) processing (PRE) - Classification- Temporal memory- Thresholding




T T IR IR




R

T T IR IR

Fig. 1. Schematic sketch of the MICRA block diagram and flow chart. i) The first step of the background process consists: in estimating the surface rain-rate from available LEO-MW measurements by means of either empirical retrieval algorithms or inversion schemes based on parametric cloud radiative models (inversion step). This enables the inversion of a set of brightness temperatures (TB) at different frequency and polarization to provide a rain-rate product spatially integrated within the nominal area A. ii) The second step of the background process pursues the combination of LEO-MW sensor data with data coming from GEO-IR sensor in space and time on a global scale (collocation step). This step consists in temporally locating the GEO-IR data within the past few tens of minutes of the LEO-MW data time and to re-map into the geographic coordinates both GEO-IR and LEO-MW measurements available observations. Note that, since spatial resolution of MW data is generally worse than IR ones, a MW field-of-view of nominal area A generally includes more than one IR pixel. As a result of the background process, a data set is generated, containing the per-pixel rain-rate retrieved from LEO-MW data, the co-located GEO-IR brightness temperature and the pixel geo-location. This process is continuously ongoing, since new LEO-MW and GEO-IR data are continuously ingested depending on available satellite platforms. iii) The third step is a foreground process, started to derive the R-TIR inverse relationship once the data set has been updated. A pre-processing stage is accomplished after each background process. The IR retrieval relationships are

updated every time a new set of combined data have been added to the data set relative to that sub-region, and are derived using data archived in a time window of several hours (integration step). To assure that only the most recent rain history is captured and to guarantee a statistical significance of the training set, the R-TIR inverse relationship for a given sub-region is derived using only the most recent combined data. iv) The last step is represented by the prediction of the surface rain-rate from IR measurements in a given sub-region by applying the derived R-TIR algorithm (retrieval step). Many attempts have been carried out so far to derive this R-TIR relationship, using different techniques such as probability matching formulations, regression methods and artificial neural networks. Details on the background and foreground processes are given section 4. 3. Neural rainfall estimation algorithm using

microwave and infrared sensors The NEREMIS techniques shares the same methodology of MICRA, but uses a Multiply Layer Feed Forward NN architecture as shown in Fig. 2 (Grimes et al., 2003; Tapiador et al., 2004b). It is a four layers network using as input the IR TB from GEO-IR sensor and latitude and longitude information.

Fig 2 Neural Network architecture for rainfall estimation

The NN is calibrated using input data derived from the GEO-IR sensor and output rain-rate retrieved from available LEO-MW measurements. This requires, as for MICRA, measurements from both IR and microwave sensors collocated in space and in time. The calibration period has been fixed for this study to 24 hours and the validation has been carried on for the following day. In particular, the calibration periods are the 24 hours of the Jan. 23, 2003, while Jan. 24, 2003 has been retained for the validation. There is not any constraint in choosing a calibration period. The one day time step has been selected due to the length of the precipitation event that has been observed for at least tree days. In case of a short rain event the time step could be reduced to few hours.

Fig. 2. Neural Network architecture for rainfall estimation.

regression methods and artificial neural networks. De-tails on the background and foreground processes aregiven Sect. 4.

3 Neural rainfall estimation algorithm using microwaveand infrared sensors

The NEREMIS techniques shares the same methodology ofMICRA, but uses a Multiply Layer Feed Forward NN archi-tecture as shown in Fig. 2 (Grimes et al., 2003; Tapiador etal., 2004b). It is a four layers network using as input the IRTB from GEO-IR sensor and latitude and longitude informa-tion.

The NN is calibrated using input data derived from theGEO-IR sensor and output rain-rate retrieved from avail-able LEO-MW measurements. This requires, as for MICRA,measurements from both IR and microwave sensors collo-cated in space and in time. The calibration period has beenfixed for this study to 24 h and the validation has been carriedon for the following day. In particular, the calibration periodsare the 24 h of the 23 January 2003, while 24 January 2003has been retained for the validation.

There is not any constraint in choosing a calibration pe-riod. The one day time step has been selected due to thelength of the precipitation event that has been observed forat least tree days. In case of a short rain event the time stepcould be reduced to few hours.

The validation of NEREMIS has been carried on in thesmaller area of central Italy due to data availability prob-lem. The performances of the trained neural network willbe shown in Sect. 4.

4 Case studies and preliminary validation

The MICRA technique has been tested on a few case stud-ies characterized by intense and persistent rainfall over theMediterranean basin. The LEO-MW data are provided bySpecial Sensor Microwave Imager (SSM/I) on board of thethree Defense Meteorological Satellite Program (DMSP)platforms. LEO-IR data are provided by the Visible InfraredSpinning Scan Radiometer (VISSR) on board of the opera-tional European geostationary satellite METEOSAT.


July 11, 2005 3

The validation of NEREMIS has been carried on in the smaller area of central Italy due to data availability problem. The performances of the trained neural network will be shown in section 4.

Fig. 3. Top: Rain rate estimated by MICRA for 24 January 2003, 23:30 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of MICRA rain rate estimates for 24 January 2003. 4. Case studies and preliminary validation The MICRA technique has been tested on a few case studies characterized by intense and persistent rainfall over the Mediterranean basin. The LEO-MW data are provided by Special Sensor Microwave Imager (SSM/I) on board of the three Defense Meteorological Satellite Program (DMSP) platforms. LEO-IR data are provided by the Visible Infrared Spinning Scan Radiometer (VISSR) on board of the operational European geostationary satellite METEOSAT. The first case study we have analyzed occurred at the end of January 2003. In the left half of Fig. 3 (left) is shown rain rate over the Mediterranean basin as estimated by the current implementation of MICRA at 23:30 UTC of Jan 24, 2003. It is evident a large convective cell developing over Tuscany. Such a product is available every time a METEOSAT-VISSR image is released. By simply integrating a day-long time series of these estimated rain rates, we are able to provide 24-hour accumulated rain over the same area, as illustrated in the right half of Figure 3. In

Fig. 4. Top: Rain rate estimated by NEREMIS for 24 January 2003, 23:30 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of NEREMIS rain rate estimates for 24 January 2003. the 24 hours of Jan. 24, 2003, a large amount of rain has fallen near the costs of UK. The effects of the just mentioned convective activity over Tuscany are still visible. For the same case study the NEREMIS estimates are shown. In Fig. 4 the instantaneous rain rate for 23:00 UTC and the 24 hours accumulated rain are show for central Italy. Comparing Figure 2 and Figure 3 for the 24 hours accumulated rain the shape and the intensity of the rain fields are in fairly good agreement, although the spatial domain is limited. The instantaneous rain rate field estimated for the 23:30 UTC by MICRA and NEREMIS show a similar shape, but the intensity is slightly different. In the top panel of Fig. 5 is presented the 02:00 UTC product for the case study of May 20, 2003. Two areas with intense precipitation are distinguishable, the larger over Poland, while the other over Germany. A map of 24-hour accumulated rain for this case study is illustrated in the bottom of Figure 5. As the meteorological situation has developed south-eastward, the largest amount of accumulated rain was found in Northern Italy. Indeed, flash floods were reported near the cost of North-eastern Italy

Fig. 3. Top: Rain rate estimated by MICRA for 24 January 2003,23:30 UTC. Bottom: 24-hour accumulated rain as computed from1-day time series of MICRA rain rate estimates for 24 January 2003.

The first case study we have analyzed occurred at the endof January 2003. In the top of Fig. 3 is shown rain rate overthe Mediterranean basin as estimated by the current imple-mentation of MICRA at 23:30 UTC of 24 January 2003. Itis evident a large convective cell developing over Tuscany.Such a product is available every time a METEOSAT-VISSRimage is released. By simply integrating a day-long time se-ries of these estimated rain rates, we are able to provide 24-hour accumulated rain over the same area, as illustrated inthe bottom of Fig. 3. In the 24 h of 24 January 2003, a largeamount of rain has fallen near the costs of UK. The effects ofthe just mentioned convective activity over Tuscany are stillvisible.

For the same case study the NEREMIS estimates areshown. In Fig. 4 the instantaneous rain rate for 23:30 UTCand the 24 h accumulated rain are show for central Italy.Comparing Figs. 2 and 3 for the 24 h accumulated rain the



July 11, 2005 3

The validation of NEREMIS has been carried on in the smaller area of central Italy due to data availability problem. The performances of the trained neural network will be shown in section 4.

Fig. 3. Top: Rain rate estimated by MICRA for 24 January 2003, 23:30 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of MICRA rain rate estimates for 24 January 2003. 4. Case studies and preliminary validation The MICRA technique has been tested on a few case studies characterized by intense and persistent rainfall over the Mediterranean basin. The LEO-MW data are provided by Special Sensor Microwave Imager (SSM/I) on board of the three Defense Meteorological Satellite Program (DMSP) platforms. LEO-IR data are provided by the Visible Infrared Spinning Scan Radiometer (VISSR) on board of the operational European geostationary satellite METEOSAT. The first case study we have analyzed occurred at the end of January 2003. In the left half of Fig. 3 (left) is shown rain rate over the Mediterranean basin as estimated by the current implementation of MICRA at 23:30 UTC of Jan 24, 2003. It is evident a large convective cell developing over Tuscany. Such a product is available every time a METEOSAT-VISSR image is released. By simply integrating a day-long time series of these estimated rain rates, we are able to provide 24-hour accumulated rain over the same area, as illustrated in the right half of Figure 3. In

Fig. 4. Top: Rain rate estimated by NEREMIS for 24 January 2003, 23:30 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of NEREMIS rain rate estimates for 24 January 2003. the 24 hours of Jan. 24, 2003, a large amount of rain has fallen near the costs of UK. The effects of the just mentioned convective activity over Tuscany are still visible. For the same case study the NEREMIS estimates are shown. In Fig. 4 the instantaneous rain rate for 23:00 UTC and the 24 hours accumulated rain are show for central Italy. Comparing Figure 2 and Figure 3 for the 24 hours accumulated rain the shape and the intensity of the rain fields are in fairly good agreement, although the spatial domain is limited. The instantaneous rain rate field estimated for the 23:30 UTC by MICRA and NEREMIS show a similar shape, but the intensity is slightly different. In the top panel of Fig. 5 is presented the 02:00 UTC product for the case study of May 20, 2003. Two areas with intense precipitation are distinguishable, the larger over Poland, while the other over Germany. A map of 24-hour accumulated rain for this case study is illustrated in the bottom of Figure 5. As the meteorological situation has developed south-eastward, the largest amount of accumulated rain was found in Northern Italy. Indeed, flash floods were reported near the cost of North-eastern Italy

Fig. 4. Top: Rain rate estimated by NEREMIS for 24 January 2003,23:30 UTC. Bottom: 24-hour accumulated rain as computed from1-day time series of NEREMIS rain rate estimates for 24 January2003.

shape and the intensity of the rain fields are in fairly goodagreement, although the spatial domain is limited. The in-stantaneous rain rate field estimated for the 23:30 UTC byMICRA and NEREMIS show a similar shape, but the inten-sity is slightly different.

In the top panel of Fig. 5 is presented the 02:00 UTCproduct for the case study of 20 May 2003. Two areaswith intense precipitation are distinguishable, the larger overPoland, while the other over Germany. A map of 24-houraccumulated rain for this case study is illustrated in the bot-tom of Fig. 5. As the meteorological situation has developedsouth-eastward, the largest amount of accumulated rain wasfound in Northern Italy. Indeed, flash floods were reported


July 11, 2005 4

Fig. 5. Top: Rain rate estimated by MICRA for 20 May 2003, 02:00 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of MICRA rain rate estimates for 20 May 2003. during 20-21 May 2003. Although qualitative results are encouraging, a systematic validation would be needed. A way to pursue this goal is by comparison with ground based remote (radar) and in situ (raingauge networks) measurements available for the areas under investigation, as illustrated in Fig. 6. Unfortunately, both data sets were not available during the presented case study. A preliminary example of ground validation with radar measurements is presented in Fig. 7. Radar data (top) were collected by the S. Pietro Capofiume unit, owned by the Regional Environmental Agency of Emilia Romagna (ARPA-SIM), and processed by the Regional Meteo-hydro Service (SIM). At radar original resolution (order of 1 km), differences in rate rates of one order of magnitude are found with respect to MICRA estimates (bottom). However, a meaningful comparison has to take into account the fact that rain rate estimates by MICRA are limited by the relatively poor spatial resolution of both GEO-IR and LEO-MW measurements. Thus, radar spatial resolution needs to be degraded to the resolution achievable by satellite (order of 10 km). If we average radar measurements in Figure 7 to match the satellite pixel resolution, and we consider only the area covered by the radar beam, we obtain what shown in Fig. 8. Now the qualitative comparison looks much better: areas of intense rain are fairly well detected and the difference of rain rate values is significantly reduced. Average errors are

of the order of 30% with an overall trend to underestimate the radar estimates.

Fig. 6. Available measurements from ground-based radar and raingauge networks are used in the validation of MICRA.

Fig. 7. A comparison between rain rates as measured by ground-based radar (top) and satellite-based MICRA technique (bottom). Original radar resolution, about 1 km, is shown. From these first results it seems that the present implementation of MICRA tends to underestimate rain rates values. Further analysis is needed in order to optimize the algorithm parameterization, such as choices of sub-region size, spacing and look-back time lag. Results for NEREMIS show a similar behavior.

Fig. 5. Top: Rain rate estimated by MICRA for 20 May 2003,02:00 UTC. Bottom: 24-hour accumulated rain as computed from1-day time series of MICRA rain rate estimates for 20 May 2003.

near the cost of North-eastern Italy during 20–21 May 2003.Although qualitative results are encouraging, a systematicvalidation would be needed. A way to pursue this goal isby comparison with ground based remote (radar) and in situ(raingauge networks) measurements available for the areasunder investigation, as illustrated in Fig. 6. Unfortunately,both data sets were not available during the presented casestudy.

A preliminary example of ground validation with radarmeasurements is presented in Fig. 7. Radar data (top) werecollected by the S. Pietro Capofiume unit, owned by the Re-gional Environmental Agency of Emilia Romagna (ARPA-SMR), and processed by the Regional Meteo-hydro Service(SIM). At radar original resolution (order of 1 km), differ-ences in rate rates of one order of magnitude are found withrespect to MICRA estimates (bottom).

F. S. Marzano et al.: Satellite radiometric remote sensing of rainfall fields 271


July 11, 2005 4

Fig. 5. Top: Rain rate estimated by MICRA for 20 May 2003, 02:00 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of MICRA rain rate estimates for 20 May 2003. during 20-21 May 2003. Although qualitative results are encouraging, a systematic validation would be needed. A way to pursue this goal is by comparison with ground based remote (radar) and in situ (raingauge networks) measurements available for the areas under investigation, as illustrated in Fig. 6. Unfortunately, both data sets were not available during the presented case study. A preliminary example of ground validation with radar measurements is presented in Fig. 7. Radar data (top) were collected by the S. Pietro Capofiume unit, owned by the Regional Environmental Agency of Emilia Romagna (ARPA-SIM ), and processed by the Regional Meteo-hydro Service (SIM). At radar original resolution (order of 1 km), differences in rate rates of one order of magnitude are found with respect to MICRA estimates (bottom). However, a meaningful comparison has to take into account the fact that rain rate estimates by MICRA are limited by the relatively poor spatial resolution of both GEO-IR and LEO -MW measurements. Thus, radar spatial resolution needs to be degraded to the resolution achievable by satellite (order of 10 km). If we average radar measurements in Figure 7 to match the satellite pixel resolution, and we consider only the area covered by the radar beam, we obtain what shown in Fig. 8. Now the qualitative comparison looks much better: areas of intense rain are fairly well detected and the difference of rain rate values is significantly reduced. Average errors are




Fig. 6. Available measurements from ground-based radar and rain-gauge networks are used in the validation of MICRA.

However, a meaningful comparison has to take into ac-count the fact that rain rate estimates by MICRA are lim-ited by the relatively poor spatial resolution of both GEO-IRand LEO-MW measurements. Thus, radar spatial resolutionneeds to be degraded to the resolution achievable by satellite(order of 10 km).

If we average radar measurements in Fig. 7 to match thesatellite pixel resolution, and we consider only the area cov-ered by the radar beam, we obtain what shown in Fig. 8. Nowthe qualitative comparison looks much better: areas of in-tense rain are fairly well detected and the difference of rainrate values is significantly reduced. Average errors are of theorder of 30% with an overall trend to underestimate the radarestimates.

From these first results it seems that the present implemen-tation of MICRA tends to underestimate rain rates values.Further analysis is needed in order to optimize the algorithmparameterization, such as choices of sub-region size, spac-ing and look-back time lag. Results for NEREMIS show asimilar behavior.

5 Summary and future developments

The basic concepts of the MICRA technique have been in-troduced together with those of the NN based algorithmNEREMIS. A couple of case studies over the Mediterranean


July 11, 2005 4

Fig. 5. Top: Rain rate estimated by MICRA for 20 May 2003, 02:00 UTC. Bottom: 24-hour accumulated rain as computed from 1-day time series of MICRA rain rate estimates for 20 May 2003. during 20-21 May 2003. Although qualitative results are encouraging, a systematic validation would be needed. A way to pursue this goal is by comparison with ground based remote (radar) and in situ (raingauge networks) measurements available for the areas under investigation, as illustrated in Fig. 6. Unfortunately, both data sets were not available during the presented case study. A preliminary example of ground validation with radar measurements is presented in Fig. 7. Radar data (top) were collected by the S. Pietro Capofiume unit, owned by the Regional Environmental Agency of Emilia Romagna (ARPA-SIM), and processed by the Regional Meteo-hydro Service (SIM). At radar original resolution (order of 1 km), differences in rate rates of one order of magnitude are found with respect to MICRA estimates (bottom). However, a meaningful comparison has to take into account the fact that rain rate estimates by MICRA are limited by the relatively poor spatial resolution of both GEO-IR and LEO-MW measurements. Thus, radar spatial resolution needs to be degraded to the resolution achievable by satellite (order of 10 km). If we average radar measurements in Figure 7 to match the satellite pixel resolution, and we consider only the area covered by the radar beam, we obtain what shown in Fig. 8. Now the qualitative comparison looks much better: areas of intense rain are fairly well detected and the difference of rain rate values is significantly reduced. Average errors are




Fig. 7. A comparison between rain rates as measured by ground-based radar (top) and satellite-based MICRA technique (bottom).Original radar resolution, about 1 km, is shown.


July 11, 2005 5

Fig. 8. Same comparison as Figure 7, where radar data (top) have been averaged to match the satellite pixel resolution, and only the MICRA estimates within the area covered by the radar beam are considered (bottom).

5. Summary and future developments The basic concepts of the MICRA technique have been introduced together with those of the NN based algorithm NEREMIS. A couple of case studies over the Mediterranean basin have been discussed, through the analysis of rain rate and accumulated rain fields as estimated by MICRA and NEREMIS. In our first attempt of validation of the MICRA technique, we found encouraging results concerning the capability of the algorithm to detect the raining cells. From a quantitative point of view, MICRA seems to underestimate rain rate values when compared to ground based radar measurements. Possible sources of underestimation might be related to algorithm parameterization, such as the choice of sub-region size and spacing or the look-back time lag, whose optimum values have to be investigated in deeper detail. A systematic validation of MICRA and NEREMIS with ground-based measurements is planned, but their intercomparison of estimated rain field shows encouraging agreements. Inclusions of data derived from other sensors, like MODIS and AMSR, are also foreseen. Once that the algorithms have been tested and validated, the resulting optimized version will be included in a monitoring and forecast operational framework, using the real-time data collected by the METEOSAT Primary Data User Station (PDUS). The exploitation of Meteosat Second Generation - High Rate Information Transmission (HRIT) data is finally the challenging aim of a further development and extension of the proposed techniques.. Acknowledgements. This work has been partially funded by the EU Fifth-Framework EURAINSAT project (EVG2-

2000-00522) and by EU Risk-AWARE project. Support by the Italian Space Agency (ASI), Ministry of Education and Research (MIUR) and by the GNDCI-CNR project is also acknowledged. Radar data were kindly provided by Dr. P.P. Alberoni (ARPA-SIM, Bologna, Italy). References Bellerby T., M. Todd, D. Kniveton and C. Kidd, “Rainfall Estimation from

a Combination of TRMM Precipitation Radar and GOES Multispectral Satellite Imagery through the Use of an Artificial Neural Network”, J. App. Met., vol. 39, pp. 2115-2128, 2000.

Ferraro R.R.,”SSM/I derived global rainfall estimates for climatological applications”, J. Geophys. Res., vol. 102, pp 16,715-16,735, 1997.

Grimes D.I.F., E. Coppola , M. Verdecchia, G. Visconti, A neural network approach to real time rainfall estimation for Africa using satellite data, J. Hydrometeorology, 6, 4, 1119-1133, 2003

Hsu K.L., X. Gao, S. Sorooshian and H.V. Gupta, “Precipitation estimation from remotely sensed information using artificial neural networks”, J. Appl. Meteor., vol. 36, pp. 1176-1190, 1997.

Kummerow C. and L. Giglio, “A method for combining passive microwave and infrared rainfall observations”, J. Atmos. Oceanic Tecnol., vol. 12, pp. 33-45, 1995.

Levizzani V., F. Porcù, F.S. Marzano, A. Mugnai, E.A. Smith, and F. Prodi, “Investigating a SSM/I microwave algorithm to calibrate METEOSAT infrared instantaneous rain-rate estimates” Meteor. Applications, vol. 3, pp. 5-17, 1996.

Marzano, F. S., M. Palmacci, D. Cimini, G. Giuliani and J. F. Turk, "Multivariate Statistical Integration of Satellite Infrared and Microwave Radiometric Measurements for Rainfall Retrieval at the Geostationary Scale", IEEE Trans. Geosci. Rem. Sens., vol. 42, n. 4, pp. 1018-1032, 2004.

Marzano, F. S., G. Giuliani, and D. Cimini, "Design of rapid update cycles of rain estimation: MICRA technique", 2nd-year Report of 5th-Framework EURAINSAT project, V. Levizzani Editor, Jan 2003.

Marzano F.S., A. Mugnai and J. Turk, “Precipitation retrieval from spaceborne microwave radiometers and combined sensors”, in Remote sensing of atmosphere and ocean from space, F.S. Marzano and G. Visconti Eds., Kluwer Academic Publ., Dordrecht (NL), pp. 107-126, 2002.

Marzano F.S., J. Turk, P. Ciotti, S. Di Michele, and N. Pierdicca, “Potential of combined spaceborne microwave and infrared radiometry for near real-time rainfall attenuation monitoring along earth satellite”, Int. J. Satell. Commun., vol. 19, n. 4, pp. 385-412, 2001.

Miller S.W., P.A. Arkin, and R. Joyce, “A combined microwave infrared rain rate algorithm”, Int. J. Remote Sensing, vol. 22, pp. 3285-3307, 2001.

Tapiador, F. J., C. Kidd, V. Levizzani, and F.S. Marzano, “A neural networks-based PMW-IR fusion technique to derive half hourly rainfall estimates at 0.1° resolution”, J. Appl. Meteor., vol. 43, pp. 576-594, 2004a.

Tapiador F.J., C. Kidd, K.L. Hsu, and F.S. Marzano, “Neural Networks in Satellite Rainfall Estimation”, Meteorological Applications, vol. 11, pp1-9, 2004b.

Turk J.F., G. Rohaly, J. Hawkins, E.A. Smith, F.S. Marzano, A. Mugnai, and V. Levizzani, “Meteorological applications of precipitation estimation from combined SSM/I, TRMM, and geostationary satellite data,” in Microwave Radiometry and Remote Sensing of the Environment, P. Pampaloni Ed., VSP Intern. Sci. Publisher, Utrecht (NL), pp. 353-363, 1999.

Vicente G.A., R.A. Cofield, and W.P. Menzel, “The operational GOES infrared rainfall estimation technique”, Bull. Amer. Meteor. Soc., vol. 79, pp 1883-1898, 1998.

Fig. 8. Same comparison as Fig. 7, where radar data (top) havebeen averaged to match the satellite pixel resolution, and only theMICRA estimates within the area covered by the radar beam areconsidered (bottom).

basin have been discussed, through the analysis of rain rateand accumulated rain fields as estimated by MICRA andNEREMIS. In our first attempt of validation of the MICRAtechnique, we found encouraging results concerning the ca-pability of the algorithm to detect the raining cells.


From a quantitative point of view, MICRA seems to un-derestimate rain rate values when compared to ground basedradar measurements. Possible sources of underestimationmight be related to algorithm parameterization, such as thechoice of sub-region size and spacing or the look-back timelag, whose optimum values have to be investigated in deeperdetail.

A systematic validation of MICRA and NEREMIS withground-based measurements is planned, but their intercom-parison of estimated rain field shows encouraging agree-ments. Inclusions of data derived from other sensors, likeMODIS and AMSR, are also foreseen.

Once that the algorithms have been tested and validated,the resulting optimized version will be included in a monitor-ing and forecast operational framework. The exploitation ofMeteosat Second Generation - High Rate Information Trans-mission (HRIT) data is finally the challenging aim of a fur-ther development and extension of the proposed techniques.

Acknowledgements.This work has been partially funded by theEU Fifth-Framework EURAINSAT project (EVG2-2000-00522)and by EU Risk-AWARE project. Support by the Italian SpaceAgency (ASI), Ministry of Education and Research (MIUR) and bythe GNDCI-CNR project is also acknowledged. Radar data werekindly provided by P. P. Alberoni (ARPA-SIM, Bologna, Italy).

Edited by: L. FerrarisReviewed by: C. M. Medaglia and another referee

References

Bellerby, T., Todd, M., Kniveton, D., and Kidd, C.: Rainfall Esti-mation from a Combination of TRMM Precipitation Radar andGOES Multispectral Satellite Imagery through the Use of an Ar-tificial Neural Network, J. App. Met., 39, 2115–2128, 2000.

Ferraro R. R.: SSM/I derived global rainfall estimates for climato-logical applications, J. Geophys. Res., 102, 16 715–16 735, 1997.

Grimes, D. I. F., Coppola, E., Verdecchia, M., and Visconti, G.:A neural network approach to real time rainfall estimation forAfrica using satellite data, J. Hydrometeorology, 6, 4, 1119–1133, 2003

Hsu, K. L., Gao, X., Sorooshian, S., and Gupta, H. V.: Precipita-tion estimation from remotely sensed information using artificialneural networks, J. Appl. Meteor., 36, 1176–1190, 1997.

Kummerow, C. and Giglio, L.: A method for combining passivemicrowave and infrared rainfall observations, J. Atmos. OceanicTecnol., 12, 33–45, 1995.

Levizzani, V., Porcu, F., Marzano, F. S., Mugnai, A., Smith, E.A., and Prodi, F.: Investigating a SSM/I microwave algorithm tocalibrate METEOSAT infrared instantaneous rain-rate estimates,Meteor. Applications, 3, 5–17, 1996.

Marzano, F. S., Palmacci, M., Cimini, D., Giuliani, G., and Turk,J. F.: Multivariate Statistical Integration of Satellite Infrared andMicrowave Radiometric Measurements for Rainfall Retrieval atthe Geostationary Scale, IEEE Trans. Geosci. Rem. Sens., 42, 4,1018–1032, 2004.

Marzano, F. S., Giuliani, G., and Cimini, D.: Design of rapid updatecycles of rain estimation: MICRA technique, 2nd-year Report of5th-Framework EURAINSAT project, edited by: Levizzani, V.,Jan 2003.

Marzano, F. S., Mugnai, A., and Turk, J.: Precipitation retrievalfrom spaceborne microwave radiometers and combined sensors,in: Remote sensing of atmosphere and ocean from space, editedby: Marzano, F. S. and Visconti, G., Kluwer Academic Publ.,Dordrecht (NL), 107–126, 2002.

Marzano, F. S., Turk, J., Ciotti, P., Di Michele, S., and Pierdicca, N.:Potential of combined spaceborne microwave and infrared ra-diometry for near real-time rainfall attenuation monitoring alongearth satellite, Int. J. Satell. Commun., 19, 4, 385–412, 2001.

Miller, S. W., Arkin, P. A., and Joyce, R.: A combined microwaveinfrared rain rate algorithm, Int. J. Remote Sensing, 22, 3285–3307, 2001.

Tapiador, F. J., Kidd, C., Levizzani, V., and Marzano, F. S.: A neuralnetworks-based PMW-IR fusion technique to derive half hourlyrainfall estimates at 0.1◦ resolution, J. Appl. Meteor., 43, 576–594, 2004a.

Tapiador, F. J., Kidd, C., Hsu, K. L., and Marzano, F. S.: NeuralNetworks in Satellite Rainfall Estimation, Meteorological Ap-plications, 11, 1–9, 2004b.

Turk, J. F., Rohaly, G., Hawkins, J., Smith, E. A., Marzano, F.S., Mugnai, A., and Levizzani, V.: Meteorological applicationsof precipitation estimation from combined SSM/I, TRMM, andgeostationary satellite data, in: Microwave Radiometry and Re-mote Sensing of the Environment, edited by: Pampaloni, P., VSPIntern. Sci. Publisher, Utrecht (NL), 353–363, 1999.

Vicente, G. A., Cofield, R. A., and Menzel, W. P.: The operationalGOES infrared rainfall estimation technique, Bull. Amer. Me-teor. Soc., 79, 1883–1898, 1998.

satellite radiometric remote sensing of rainfall ﬁelds ...€¦ · 268 f. s. marzano et al.:...

Documents