A daytime surface rain-rate classifier, based on artificial neural networks (ANNs), is proposed for the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board the Meteosat-8 geostationary satellite. It is developed over the British Isles and surrounding waters, where the Met Office radar network provided the “ground precipitation truth” for training and validation. The algorithm classifies rain rate in five classes at 15 min and 5 km of time and spatial resolution, and is applied on daytime hours in a summer and winter database. A further ANN application is restricted to hours between 1200 and 1400 UTC for which the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) on board the Aqua polar-orbiting satellite scans the U.K. area: ANN-classifier algorithms for the SEVIRI and AMSR-E data have been developed and the results have been compared. A reliable validation procedure is adopted to quantify the performance in view of the operational application of the daytime classifier and to investigate the relative skills of passive microwave and visible–infrared radiances in sensing precipitation if processed with equivalent algorithms. The key statistical parameters used are the equitable threat score (ETS) and the bias for rain–no rain classes and the Heidke skill score (HSS) for rain-rate classes. The SEVIRI daytime classifier shows, for mean seasonal conditions, the best performance in summer, with ETS = 47% and HSS = 22%, and in winter ETS = 36% and HSS = 17% were found. The comparison between AMSR-E and SEVIRI noon classifiers reveals a similar overall skill: in detecting rain areas, SEVIRI is slightly better than AMSR-E, while the opposite is true for rain-rate classification.
The quantitative description of precipitation patterns from spaceborne passive sensors is a challenging task in meteorological remote sensing, given the elusive relationship between precipitation characteristics and the radiation detected by satellite sensors. Precipitation estimates at high temporal and spatial resolution are operationally used within forecasting systems in meteorological and hydrological services, with efforts underway to use them in assimilation chains (Davolio and Buzzi 2004; Errico et al. 2007; among others), while precipitation estimates on the global scale contribute to improving climatological and global water cycle studies (Smith et al. 2007 and references therein). Passive spaceborne sensors in the visible (VIS) and infrared (IR) and/or in the microwave (MW) part of the spectrum are used in dealing with such issues.
The advantage of using VIS–IR lies in the possibility of obtaining these measurements with high spatial resolution and, when the sensor is onboard a geostationary platform, also on a global scale and with high temporal resolution. The disadvantage is that VIS–IR radiation originates from the top cloud layers and does not directly carry information about the rain at the ground. On the contrary, the direct interaction with precipitating layers is the advantage of using MW. Unfortunately, MW radiation can be measured at a reasonable spatial resolution only if the sensor is onboard sun-synchronous satellites with poor temporal resolution.
Given the possibility of direct physical retrieval, most interest has been focused on quantitative precipitation estimates using MW radiation measurement. Several MW-based algorithms have been proposed in the literature for the different available sensors, based on statistical, or physical, approaches to quantify the relationship between rain layers and the detected radiation [see Smith et al. (1998) for intercomparison and Levizzani et al. (2007) for latest developments]. Most of the algorithms use window channels between 10 and 89 GHz, while recently the use of absorption channels has been under investigation (Cheng and Staelin 2003) and the use of millimetric radiation is currently under study (Surussavadee and Staelin 2007). Based on the concept of the polar-orbiting satellite constellation, the Global Precipitation Measurement (GPM) project (Smith et al. 2007) represents the future step toward the maximum exploitation of MW remote sensing, as this will reduce the time resolution problem.
At the same time, a new generation of geostationary sensors—such as the already operative Spinning Enhanced Visible and Infrared Imager (SEVIRI) and Geostationary Operational Environmental Satellite (GOES) I-M imagers—has improved the spatial and temporal resolution and significantly enhanced VIS–IR spectral measurements. In particular, channels in the near-infrared (NIR) and shortwave infrared (SWIR) part of the spectrum provide information on the size of hydrometeor at the cloud top (Nakajima and King 1990) and subtop (Chang and Li 2002), offering the resulting possibility of improving indirect ground precipitation estimation (Rosenfeld and Gutman 1994; Chen et al. 2007). For the current GOES-LM series (in which the important 1.6-μm channel is missing), precipitation algorithms have been proposed by Bellerby et al. (2000) and Ba and Gruber (2001), which only partially exploit the new capabilities.
To exploit to the full the characteristics of VIS–IR and MW satellite sensors, considerable interest has recently been focused on methods and algorithms that merge and combine the information from both kinds of measurements, in accordance with the earlier studies of Levizzani et al. (1992). In particular, the direct precipitation estimates from MW take advantage of the use of IR geostationary measurements in terms of spatial and temporal resolution (Turk et al. 2000; Tapiador et al. 2004; Kidd et al. 2003; Joyce et al. 2004).
Such procedures, however, do not properly consider the real potential of new geostationary channels, such as 3.9 and 1.6 μm, since only the IR channels are considered. Moreover, the quantitative difference between MW- and VIS–IR-based precipitation estimates has not yet been accurately established for midlatitude precipitation regimes.
This work aims, therefore, to contribute to the two following issues:
evaluating the expected performance of a daytime SEVIRI rain-rate classifier over a large dataset with view of its implementation on a real-time precipitation monitoring or nowcasting system;
investigating the difference between VIS–IR and MW precipitation estimation performances with a view of their proper use in a combined technique.
In addressing these two points, the focus is on the British Isles and surrounding waters, where a complete and reliable ground precipitation daytime dataset is available for summer and winter seasons. A robust validation scheme is also set up to evaluate the performance of the daytime technique and to compare quantitatively VIS–IR and MW results.
The present paper is developed through eight main sections. The introduction is followed by section 2, where a complete illustration of the dataset used is provided. The statistical approach, the architecture of the artificial neural network (ANN) rain classifier algorithm, and the ANN building phase are discussed in section 3. Section 4 is dedicated to setting up a general validation procedure based on radar rainfall estimates as “true” reference. Section 5 is devoted to point 1: daytime SEVIRI rain-rate classification schemes for summer and winter seasons are developed and validated along the previous lines, and an analysis of the rain classification seasonal variability is also carried out. Section 6 provides an analysis of the most important channels contributing to the rain-rate classification for the SEVIRI sensor. In section 7, noon-ANN rain classifier algorithms for both SEVIRI and the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) sensors (for summer and winter seasons and times between 1200 and 1400 UTC) are developed and compared in an attempt to find an answer to point 2 and are also compared with the AMSR-E precipitation global product. Section 8 outlines the conclusions.
2. Description of the dataset
The supervised dataset required to set up a statistical rainfall estimation technique consists of a series of satellite images with associated corresponding (collocated and simultaneous) surface rain-rate maps provided by an independent instrument (i.e., radar or rain gauge) and referred to as truth. For this work, the satellite images are provided by SEVIRI and AMSR-E sensors mounted on board the geostationary Meteosat-8 satellite and the polar-orbiting Aqua satellite, respectively, whereas the surface rain-rate truth maps are obtained from the radar composite from the Met Office Nimrod system (Golding 1998).
All the available SEVIRI images at 15-min time resolution for June, July, and August 2004 from 0800 to 1800 UTC (approximately 3600 images) and for December 2004, January, and February 2005 from 0900 to 1500 UTC (approximately 1750 images) were collected. The time intervals were selected to avoid twilight in the images.
AMSR-E images were collected for the same summer and winter months by considering the Aqua noon overpasses over the U.K. area between 1200 and 1400 UTC (92 overpasses for summer and 85 overpasses for winter).
For the two seasons collocated, nearly simultaneous Nimrod radar data were selected and associated with corresponding satellite data to form the daytime summer and winter supervised datasets and the noon summer and winter supervised datasets. An element of the daytime seasonal datasets, hereinafter referred to as case, is formed by a SEVIRI multispectral set of images and the corresponding rain-rate radar map, while for the noon datasets, each case is constituted by nearly simultaneous SEVIRI and AMSR-E multispectral images and the corresponding rain-rate radar map.
a. SEVIRI data
Meteosat-8 is the first spacecraft of the second generation of European meteorological satellites, launched on 22 August 2002 from the Kourou base in the French Guiana (Schmetz et al. 2002). The satellite follows a geostationary orbit at an altitude of about 36 000 Km. The SEVIRI onboard Meteosat-8 and -9 scans Earth’s surface every 15 min, with a sampling of 3 km at the subsatellite point. The upwelling radiation is measured at 12 channels (including the High-Resolution Visible channel), 9 of which were selected on the basis of preliminary tests, centered on the following wavelengths: 0.635 μm (VIS0.6), 0.81 μm (VIS0.8), 1.64 μm (NIR1.6), 3.92 μm (IR3.9), 6.25 μm (WV6.2), 7.35 μm (WV7.3), 8.70 μm (IR8.7), 10.80 μm (IR10.8), and 12.0 μm (IR12.0). The data were freely obtained from the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) archive service.
b. AMSR-E data
AMSR-E, aboard the National Aeronautics and Space Administration (NASA) Aqua satellite, flies in sun-synchronous near-polar orbit at a nominal altitude of 705 km. It measures the upwelling scene brightness temperature at 6.9, 10.7, 18.7, 23.8, 36.5, and 89.0 GHz (horizontally and vertically polarized) over an angular sector of ±61° centered on the subsatellite track, resulting in a swath width of 1445 km. Because of the conical scanning, these measurements come from an earth incidence angle of 55°. The spatial resolution ranges from 5.4 km at 89.0 GHz to 56 km at 6.9 GHz. The AMSR-E/Aqua L2A global swath spatially resampled brightness temperatures (Ashcroft and Wentz 2003) were freely downloaded from the online data pool at the National Snow and Ice Data Center (NSIDC) Distributed Active Archive Center (DAAC).
c. Radar data
Radar data are collected at 5-min intervals nominally starting at 00 each hour, from a network of 15 C-band (5.3-cm wavelength) radars across the United Kingdom as shown in Fig. 1 (Fair et al. 1990). The radar-derived precipitation rate product represents instantaneous rain rate averaged over a pixel area (5 km × 5 km) and the smallest nonzero rate is 1/32 mm h−1.
The measurement accuracy is difficult to quantify and varies significantly with the type of rainfall events. However, verification studies employing a rain gauge network (Harrison et al. 2000) show that the root-mean-square factor (RMSF) difference in the radar estimates can be significantly reduced by including several corrections for hardware or data transmission faults, permanent ground clutter, and attenuation. The use of an adjustment factor based on rain gauge data, and the vertical profile of reflectivity correction scheme, also help to refine estimates. Nevertheless, the remaining RMSF difference is still approximately a factor of 2 in the first 100-km range. Outside this range the difference is larger and therefore only the first 100 km of the radar range are considered here, as shown in Fig. 1.
The rain-rate estimate from radar is a continuous variable. However, because an indirect precipitation estimation algorithm for VIS–IR satellite measurements is pursued, it is convenient to consider classes of precipitation. The classes considered are the five used in the Nimrod system, which represent an appropriate choice for giving, for each class, a sufficient number of pixels for significant statistical analysis. They are class 0 (0–1/32 mm h−1), class 1 (1/32–0.125 mm h−1), class 2 (0.125–0.5 mm h−1), class 3 (0.5–2 mm h−1), and class 4 (>2.0 mm h−1).
The radar rainfall estimate distribution into the four raining classes of precipitation is shown in Fig. 2 for both seasons considered here. The distribution for the period 1200–1400 UTC (winter and summer) does not show a significant difference from the plots in Fig. 2.
As it will be useful to distinguish relatively wet cases from relatively dry cases in the following validation phase, it is worth establishing here a conventional way to define them: the most simple parameter depending on the rain occurrences is the ratio between dry and wet pixels in the scene [dry to wet ratio (DWR)]. For both winter and summer seasons, the mean DWR is around 7.2. By fixing a DWR threshold equal to 15, roughly double the climatological value, cases with DWR above the threshold are considered dry, while cases with DWR less than the threshold are considered wet, without any assumption on cloud and rain types.
The DWR distribution for the two seasons is reported in Fig. 3, where the DWR range is limited to values less than 70, while all the cases are considered in this study. The vertical black line represents the threshold discriminating wet cases (approximately 65% of the total) and dry cases (remaining 35% of the total).
d. Satellite data preprocessing
To build the supervised datasets, the Nimrod radar maps closest in time to each satellite observation are selected and the satellite data are remapped onto the Nimrod radar grid.
Nimrod rainfall maps are generated at 5-min intervals from sets of radar scans taken by the network radars during the previous 4 min. The SEVIRI scans the Nimrod area in roughly 1.5 min and the center of that area is viewed at 11, 26, 41, and 56 min past each hour. By selecting the corresponding radar composite maps at 15, 30, 45, and 60 min past each hour, the maximum temporal lag that might occur between radar and SEVIRI pixels is roughly 4.75 min.
The AMSR-E sensor scans the Nimrod area in approximately 5 min and the time corresponding to the Nimrod area center is variable. By selecting the temporally closest radar image, the maximum time lag that might occur between radar and AMSR-E pixels is around 6.5 min.
These time lag maximum values, whenever they take place, are acceptable for slow-moving, midlatitude cloud systems. Nevertheless, they have to be regarded as an intrinsic source of error in matching satellite and radar maps, such as parallax, different viewing angles, and different pixel size.
Because the pixel size in both SEVIRI and AMSR-E is larger than that of the Nimrod, the remapping procedure consists of determining the fraction of each satellite pixel in each Nimrod pixel, without considering the antenna pattern. Subsequently, the remapped satellite image is obtained by averaging the radiance values corresponding to the fractions of different satellite pixels within the same Nimrod pixel.
The parallax correction is applied only to SEVIRI data, because the SEVIRI IR channel provides a relatively reliable cloud-top height estimate: the cloud-top height product and the parallax correction routine are from the Nimrod system. As MW radiation comes from different layers of the cloud, a similar procedure turns out to be inappropriate, and a more sophisticated strategy is required (Bauer et al. 1998); thus the parallax problem could affect the quality of the spatial matching between AMSR-E and radar data.
After the two seasonal supervised datasets have been constructed, two supervised noon datasets are extracted by selecting the SEVIRI images corresponding to the noon AMSR-E overpasses over the U.K. area (from 1200 to 1400 UTC) and including the AMSR-E images to form the noon seasonal datasets. AMSR-E overpasses with at least 7000 pixels inside the area were selected, roughly corresponding at a coverage of five radars.
The SEVIRI images for the noon dataset were selected as close as possible to the AMSR-E overpass times, the maximum temporal lag being around 11.5 min. However, two radar datasets were defined, selecting for each sensor the closest radar observation. For the two radar datasets, 30% of the cases are exactly the same, while 61% have a temporal lag of 5 min and only 9% have a temporal lag more than 5 min. Moreover, the rain-rate distributions for the two datasets do not show marked differences and the two radar datasets are considered equivalent.
3. Artificial neural network scheme and building phase
The indirect relationship between VIS–IR cloud-top radiation and ground precipitation makes the definition of the SEVIRI rain classifier by means of a statistical approach a valuable choice. A statistical approach defines a correlation function between satellite radiances and precipitation at the ground by statistical techniques, such as neural networks, linear regression analysis, histogram matching, and clustering techniques, among others. In the case of satellite MW measurements, either physical or statistical approaches can be adopted, but in this work the AMSR-E rain classifier is built by considering a statistical approach similar to that used for the VIS–IR.
Artificial neural networks, widely used in precipitation remote sensing (Hsu et al. 1997; Bellerby et al. 2000; Grimes et al. 2003; Tapiador et al. 2004), are the statistical tool chosen to define the correlations between satellite measurements and classes of ground precipitation as estimated by weather radars. The satellite rain classifier scheme here adopted needs four ANNs used in cascade, as illustrated in Fig. 4. Starting from a satellite image, the first ANN is trained to separate the class 0 pixels (dry pixels) from the other rain classes' pixels (wet pixels). The second ANN is applied to all the wet pixels and extracts from them the class-1 pixels. The third ANN separates the class-2 pixels from the class-3 and class-4 pixels and, finally, the remaining pixels are classified into classes 3 or 4 by means of the fourth ANN.
Each neural network used in this scheme is a multilayer perceptron (Rosenblatt 1962), which consists of input nodes, hidden nodes, and one output node, the nodes being called perceptrons. The ANN configuration used is the single hidden layer (with 10 nodes) version employing a sigmoidal transfer function at the nodes. The input nodes are the numerical values from satellite channels: all the available channels or an appropriately selected set can be used. The output node provides a number between 0 and 1, which, with a threshold value, may be applied to distinguish the two classes of precipitation under consideration. The output of each node in one layer is connected to the inputs of all the nodes in the next layer. The strengths of the connections are represented by continuously variable weights. The process of training the ANN consists of adjusting and defining the values of these weights. Details of the ANN architecture and ANN building procedures adopted can be found in Pankiewicz et al. (2001) and Capacci and Conway (2005).
To build the four ANNs, four ensembles of supervised pixels (called “building ensembles”) have to be selected from the supervised datasets. Each building ensemble must contain pixels from the two precipitation classes (or sum of classes) that the corresponding ANN has to learn to recognize. These pixels are labeled 0 or 1, depending upon their belonging to the two classes, as defined by means of radar data. From each building ensemble, by a random splitting procedure, the training and testing ensembles are obtained, the first being used to train the ANN and the second to test it. During the training phase, different sets of weights (and different ANNs) are computed and adjusted by several presentations of the training ensemble (called “epochs”). By applying the trained ANNs to the testing ensemble, it is therefore possible to select the one that best reproduces the “a priori known” classification of the testing ensemble. The procedure is repeated for the four ensembles and, in the end, the four ANNs are determined. The satellite classification scheme, obtained with the four ANNs, is then ready to be validated over a dataset independent from the building ensembles.
4. Validation procedure
The first step in the setup of the validation procedure is the selection of a number N of cases not considered for the building phase, which form the validating dataset. The ANN classifier applied to the N satellite multispectral images produces N satellite precipitation maps, for comparison with the corresponding N radar precipitation maps, to establish the extent to which the satellite rain classifier scheme is able to reproduce the radar precipitation estimates.
While a visual inspection of the two series of maps (the so-called eyeball validation) may give a rather comprehensive impression of the performance of the technique, it is necessary to quantify the comparison by using some statistical skill indicators, chosen among a rather large range of statistical parameters. In the present work it was decided to consider the equitable threat score (ETS) and the bias when two classes of precipitation are considered, and the Heidke skill score (HSS) for multiple precipitating classes. The HSS, a statistical parameter well known in literature, especially useful for categorical classification (Nurmi 2003), is often used as a surrogate for the correlation coefficient (Barnston 1992). Other statistical parameters that will be considered (and widely used in the literature) are the probability of detection (POD) and the false-alarm rate (FAR).
By overlapping the observed and predicted rain maps, two kinds of contingency tables are produced: the rain–no rain contingency table and the four-class contingency table. For the latter, the dry pixels are not considered. The advantage of this choice is that the validation for the raining pixels is not affected by the easier rain–no rain classification, which, due to the large number of dry pixels, would have too great an impact on the HSS value. From the contingency tables the skill indicators are computed (Nurmi 2003).
The method of determining the contingency tables defines two validation procedures:
ensemble validation: all the pixels in the validating ensemble are used to build one contingency table and a single value for the statistical parameters is computed;
case validation: from the N cases in the validating dataset, N contingency tables are built and the statistical parameters computed. The N values are then averaged to provide the statistical parameter as a mean value among the N cases with associated standard deviation.
There is an important difference between the two procedures: the ensemble validation procedure is useful to infer general information on the satellite classifier performance for typical or seasonal precipitation characteristics. The case validation procedure is useful to evaluate the variability of satellite classifier performances by varying the precipitation characteristics, especially the DWR value.
The validation procedure is then applied to the whole validation dataset and to a relatively wet subset (defined in section 2c), mainly containing cases with a large raining area. As the radar suffers the attenuation problem and may have a limited detecting range, especially for these cases, an operational nowcasting system would need the support of satellite data. The validation for a dry dataset can be easily inferred from the total and wet validation results, considering that the amount of dry cases is around 35% of the total.
Since the performance of any rainfall estimation technique depends on the characteristics of the validating dataset [e.g., geographical region, season, time of the day, number of cases, pixel size, total number of pixels, rain–no rain threshold, dry to wet ratio, rainfall rate (RR) mean value and extremes], in any validation exercise an essential description of these characteristics should be provided to complete the information inferred from the validating parameters illustrated above. These characteristics can be given in the simplest way, as it is done here, or can provide more detailed information, such as cloud/rain types. This will favor a more reliable comparison between the different techniques.
5. Daytime SEVIRI rain-rate classifier
a. Validation and performance analysis
Two SEVIRI rain classifiers are built: the daytime summer SEVIRI ANN and the daytime winter SEVIRI ANN, trained on the proper building ensembles. All the nine selected SEVIRI channels are used as input data plus the solar zenith angle and the building procedure is the one described in section 3. In the supervised dataset, the pair of radar–SEVIRI data is available at 00, 15, 30, and 45 min past each hour. All the cases at 00 min of each hour are used for the building ensembles, whereas all the cases at 15, 30, and 45 min past each hour (with at least one raining pixel: roughly 94% of the cases) are used for validating datasets. In this way, roughly 900 cases for summer and roughly 400 cases for winter enter into the building phase, whereas 2500 cases for summer and 1250 cases for winter are used for the validation. Are these validating cases really independent from the building ensembles? Since some building pixels may come from precipitating events with a time lag of only 15 min with respect to the precipitating event used in the validation, full independence is not guaranteed. On the other hand, the pixels actually processed (for both training and testing) in the building phase are only 15% of the whole building ensemble (Capacci and Conway 2005). This part, moreover, is roughly 6% of the validating dataset size, meaning that most of the validating pixels might be considered independent from the training pixels. Further evidence of the negligible impact of dataset interdependency will be presented in the discussion of the noon classifier results. Once the ANNs are obtained from the building procedure, the cascade scheme can be implemented and validated. Following the recommendations in section 4, for the above-specified geographical area (UK), pixel size (5 × 5 km2), and rain–no rain threshold (1/32 mm h−1), the description of the validating dataset and of the wet validating dataset used is reported in Table 1.
The validation results, from both the ensemble and case validation procedures applied to the datasets described in Table 1, are summarized in Table 2. The parentheses contain the computed standard deviation corresponding to the mean values obtained from the case validation. Although the overall performances of the technique in discriminating rain areas are expressed by ETS and bias, it is also worth looking at the POD and FAR values resulting from the ensemble validation. They are, respectively, around 68% (70%) and 31% (30%) for the summer total (wet) ensemble, and 55% (58%) and 37% (36%) for the winter total (wet) ensemble, reflecting the difficulties of the technique in detecting wet pixels during the winter.
Analyzing the validation results, the following main remarks can be made:
the summer ANN rain classifier works better than the winter ANN classifier;
relative to the rain–no rain classification, the winter ANN classifier markedly underestimates the raining pixels;
relative to the case validation, the results from the total dataset are always worse than the results from the wet dataset.
The last result confirms that the satellite precipitation estimates are more reliable for wet events than for dry ones. This is further highlighted by the fact that the case validation, in which the DWR may vary considerably, shows a worse ETS with respect to the ensemble validation in which the DWR is around 7.
To retrieve more detailed information on the rainfall classifier performances, the case validation procedure is applied to each day of the summer and winter months under examination. Here, the number N, introduced in section 4, is the collection of cases for each day. As the validating times considered are at 15, 30, and 45 min past each hour, the summer N value is roughly 30, whereas the winter N value is roughly 18. The corresponding daily number of pixels is, respectively, 450 000 and 270 000.
The results for summer are reported in Figs. 5 and 6, and those for winter in Figs. 7 and 8. The ETS values, DWR−1, HSS, and the rainfall rate are plotted daily for summer and winter months to demonstrate how the performances depend upon the precipitation characteristics of the group of cases examined.
This is particularly evident when observing Figs. 5 and 7, in which the ETS parameter follows the DWR variability quite closely. Although less marked, also in Figs. 6 and 8, it is possible to note that the high and low peaks of HSS values sometimes correspond to high and low peaks of RR values. To explain this observation, two factors have to be considered. First, it is intuitive to point out that cases with few precipitating pixels and low mean rain rate are more difficult to predict from satellite images. Second, the evaluating statistical parameters have an intrinsic dependence upon the occurrences of pixels in the considered classes. This is particularly true for the ETS, which assumes low values for high values of the DWR parameter, as shown in Porcù and Capacci (2007). For these reasons, the validation results, expressed in terms of the statistical parameters ETS and HSS, must be integrated with information about the datasets used.
b. Seasonal variability
Table 2 shows that the ETS and HSS decrease in the shift from summer to winter. In particular, for the ensemble validation, the percentage decrease is quantifiable in 23% and 25%, respectively. As shown by Porcù and Capacci (2007), for VIS–IR rain–no rain classification, this might be because of two main causes: different cloud characteristics and different solar zenith angles in summer and winter. The present section attempts to quantify roughly the impact of such factors on the seasonal rain classification variability. The availability of a large dataset makes it possible to determine ensembles of pixels at nearly fixed solar zenith angles: 5° zenith angle intervals from 30° to 85° for summer, and from 55° to 90° for winter. For each of these intervals, the building and validation phase are carried out following the already established procedures. The validation results from all of the ensembles are reported in the plots of Fig. 9.
As shown in Fig. 10 the range of angle values common to both the seasons is roughly between 60° and 80°. The low number of summer pixels in the [80°, 85°] interval makes the corresponding analysis unreliable, and therefore the previous range excludes this interval.
The ETS and HSS variability due to the cloud seasonal variability only can be assessed by computing the difference between summer (solid lines) and winter (dashed lines) performances at the same solar zenith angle in the range [57.5°, 77.5°] as shown in Fig. 9. The computation shows that, on average, in the shift from summer to winter, the percentage decrease of ETS and HSS is about 11% and 12%, respectively.
On the other hand, the role of the solar zenith angle can be assessed in the following way. As shown in Fig. 10, the maxima of the summer and winter solar zenith angle distributions are, respectively, at 40° and 75°. Focusing attention on the summer results, it is then possible to evaluate the difference in performance between the analyses at angles around 40° and around 75°. The computation shows that by moving from about 40° to about 75°, the ETS and HSS values decrease by about 6% and 19%, respectively.
Bearing in mind that the many winter cases with solar zenith angles higher than 80° might still decrease performance, it can be inferred that the ETS seasonal variability is due to cloud and zenith solar angle variability in roughly equal measure.
For the HSS parameter it seems that the solar zenith angle plays a greater role in the deterioration of winter performance than seasonal cloud variability.
These conclusions also apply when the “portability” of the technique to other regions outside the calibration area is considered. A variation of the performance is expected because of possible differences in cloud types, rain climatology, and mean solar zenith angle, which could be of the same order as the one estimated by the above analysis.
6. SEVIRI input channel analysis
Though all nine channels are used as input data for the SEVIRI daytime algorithm, it is important to understand which channel, or channel subset, provides the most important contribution for rain-rate classification. This is particularly useful for simplified satellite rain-rate estimation schemes requiring a limited set of input data (Francis et al. 2006), or in addressing more in-depth studies (Capacci et al. 2004) on the physics of the precipitation formation process.
Information on this issue can be inferred by focusing on the noon summer and winter datasets and considering only the building phase. In the latter, the building ensemble, determined from all noon cases, is randomly split into two ensembles: one is used to train the ANN, with a defined set of input channels, and the other to estimate its skill. This procedure cannot be considered a validation, as the ensemble of pixels used is not entirely independent of the one used in the training: by randomly splitting the above ensemble, two pixels belonging to the two ensembles obtained might come from the same meteorological event. However, two factors favor the use of the said procedure:
it is very fast in terms of computing time, thus allowing the necessarily very high number of programs required;
only relative indications are pursued with this discussion.
The channel analysis is carried out by considering any combination of the one to nine channels as ANN input (along with the solar zenith angle) for a total of 511 combinations and for each of the four ANNs within the rain-rate classification scheme. Therefore, for each channel combination, the skill in classifying pixels in the four pairs of classes [class 0/class (1 + 2 + 3 + 4), class 1/class (2 + 3 + 4), class 2/class (3 + 4), and class 3/class 4] is evaluated by computing the ETS parameter. In the analysis, the ratio between the number of pixels belonging to the two different classes is always fixed at 1.
Figure 11 illustrates, for both summer and winter data, how the SEVIRI rain classifier performances are improved by increasing the number of input channels.
A rapid improvement in performance is attained in passing from one- to two-channel ANN, prompting an investigation to find out which pairs of channels have the best skill in separating each pair of classes or sum of classes. The analysis has been carried out for the 36 possible combinations of two channels, and the best four results obtained are reported in Table 3.
The table shows that the best pairs of channels are always VIS08–NIR1.6 and VIS06–NIR1.6, while similar performances are obtained by combining a visible channel with IR3.9.
These results constitute an important confirmation of several studies and results already reported in the literature (Lensky and Rosenfeld 1997; Ba and Gruber 2001; Chang and Li 2002) concerning the cloud-top microphysical information concealed in NIR1.6 and IR3.9 measurements, which can be exploited to infer ground precipitation.
The plot in Fig. 11 shows that the four-channel ANN provides performances quite close to the nine-channel ANN. The analysis for the 126 combinations of four channels has been carried out and the best results are reported in Table 4. The four channels indicated in the table, for each of the two raining classes considered and for the two seasons, are suggested for a simplified four-channel rain classification scheme.
Finally, the SEVIRI VIS06–IR10.8 combination does not appear among the best results reported in Table 3. As those two channels are the most similar to the VIS and IR channels onboard the Meteosat first generation, by comparing those results against the results from the four channels in Table 4 (very close to the results from all nine channels in input), it is possible to quantify the improvement due to the introduction of new channels. By carrying out that comparison, averaged over the four precipitation classes, the improvement is about 32% for summer and 47% for winter.
7. Noon-ANN classifier for SEVIRI and AMSR-E
The noon-ANN classifiers are defined for both SEVIRI and AMSR-E sensors following the procedure indicated in section 3. The AMSR-E classifier makes use of all 12 channels available as input nodes. The validation phase needs a dataset independent from the one used in the building phase. Because the cases in the supervised noon dataset come from different days, their independence is guaranteed by randomly splitting them in two groups. The first group is used to build the ANN (training and testing) and the second to validate it. Subsequently, in reverse, the second group is used to build the ANN and the first to validate it. In this way all the available cases enter into the validation. The final validation result will therefore be expressed as the mean value of the two validations.
While SEVIRI images cover all the area observed by the radar network, the AMSR-E images, given the sun-synchronous orbit of Aqua spacecraft, sometimes do not: only radar pixels that are common to the two sensor images are selected and considered in Table 5.
The noon-ANN SEVIRI and AMSR-E classifiers are validated by applying the ensemble and case validation procedures, considering first all the cases and subsequently the relatively wet cases. The results are reported in Tables 6 and 7.
First, we compare the results of the daytime (Table 2) and noon (Table 6) SEVIRI classifier, to assess the impact of the dependence between training and validating datasets, which can be expected for the daytime but certainly not for the noon classifier. Results are very similar in terms of all the statistical parameters: the noon classifier is slightly better than the daytime classifier when ETS and HSS are considered, while the opposite is true if the bias parameter is considered. This allows one to suppose that a likely interdependency between the small training dataset and the validating datasets for the daytime classifier, whenever present, has very little impact on the overall results.
To visualize what the statistical parameters obtained represent, it is worth examining some of the precipitation images obtained by the three instruments: radar, SEVIRI, and AMSR-E sensors. The two cases reported in Figs. 12 and 13 have been selected so as to have the summer and winter climatological DWR, respectively, and ETS and HSS values as similar as possible to those shown in Table 6 and 7 for the ensemble validation.
For the rain–no rain classification, the ETS parameter of Table 6 shows that the SEVIRI classifier performs slightly better (especially for winter, when the occurrence of low rain rates, underestimated by MW, is more likely) than the AMSR-E classifier, though the latter shows a better bias. In the rain-rate classification, the AMSR-E classifier shows slightly higher HSS values compared to the SEVIRI classifier, though in the latter case the standard deviation values computed in the case validation are lower. More synthetically, the SEVIRI and AMSR-E rain classifiers perform in a comparable way.
To better interpret these results, the following additional considerations must be made:
the AMSR-E images are not parallax corrected and are derived from original brightness temperature data that are at a lower spatial resolution with respect to SEVIRI images;
the mean rainfall rate of the dataset is around 1 mm h−1, whereas it is known from the literature that MW precipitation retrieval works best at higher rainfall rate conditions;
usually, MW-based precipitation retrieval performs better over the sea.
The latter remark has been confirmed by performing the validation separately over three types of background terrain: sea, land, and coast, selected as in Fig. 1. The results are reported in the plots of Figs. 14 and 15 and show that over the sea and for raining classes the difference between SEVIRI and AMSR-E performances is more marked, the AMSR-E sensor doing better. This is not true just for the winter cases validation for which it seems that SEVIRI is slightly better than AMSR-E. Occurrences of cases with light precipitation might justify this fact.
To verify the reliability of the ANN approach applied to AMSR-E data, the rain–no rain maps obtained with the present noon-ANN AMSR-E classifier have been compared with the corresponding rain–no rain maps obtained using the global algorithm (Wilheit et al. 2003). The latter are available online in the DAAC archive and are referred to as AMSR-E L2B products (Adler et al. 2007). The AMSR-E L2B rain estimates corresponding to the same summer and winter 1200–1400 UTC dataset used in the paper have been retrieved online from the EOS/NASA Web site. The validation results from both the local and the global algorithms are shown here. As for the four raining precipitation classes, the global algorithm showed very poor skill; only the comparison for the rain–no rain classification (with a rain–no rain threshold of 0.1 mm h−1) is reported in Figs. 16 and 17.
The main aims of the paper were to propose a daytime artificial neural network SEVIRI-based rain classifier for operational use over the U.K. area and to investigate the difference between VIS–IR and MW precipitation statistical retrievals for a better exploitation of them in combining–merging algorithms. A validation procedure was introduced, stressing the importance of describing the precipitation dataset used in the validation phase: the present work proved the sensitivity of performance to the seasonal cycle and number of precipitating pixels in the scene. Moreover, a distinction between validation over the total ensemble of pixels or over a series of cases is also proposed as a different way to compute and express the bias, ETS, and HSS parameters.
The daytime ANN SEVIRI rain classifier algorithm has been defined and validated for both summer and winter seasons. Performances expected for summer cases are ETS = 36%, with a standard deviation of 16% and HSS = 16%, with a standard deviation of 9%. Performances for winter are ETS = 24%, with a standard deviation of 15% and HSS = 11%, with standard deviation of 7%. The bias values show that there is a slight overestimation of rain for summer and an underestimation for winter. If only wet cases (DWR < 15) are considered, performance is increased by about 20%, while for dryer cases (DWR > 15) performance decreases by about 35%–40% for both seasons. Considering the time series of the daily mean ETS values, the dependency of the performances on DWR is confirmed, while higher HSS daily mean values are usually reached when the averaged rain rate is higher.
The marked difference between summer and winter performances was investigated. For the rain–no rain classification, the difference is found to be equally determined by seasonal cloud variability and solar zenith angle changes, whereas for the rain-rate classification the analysis indicates that solar zenith angle variability plays a greater role in determining the difference.
The SEVIRI channel analysis showed that the VIS–1.6-μm combination has the best skill in distinguishing the five classes of precipitation considered here, and that any SEVIRI rain classifier scheme should consider that combination as the main contributor to the expected performances.
The VIS–IR and MW rain statistical retrieval comparison, restricted to noon hours, shows that SEVIRI and AMSR-E sensors perform in a very similar fashion. The bias shows an overall overestimate of precipitation areas to the VIS–IR retrieval and an equivalent underestimation to the MW algorithm. ETS values are very similar, except in winter cases, when VIS–IR is markedly better, while slightly higher HSS is obtained by the MW algorithm in all seasons and conditions. The background slightly influences the MW performances, showing higher ETS and HSS over the sea, especially in summer.
These results highlight the important role that multispectral VIS–IR observation may have in high-resolution precipitation retrieval, at least during daytime, and also in the coming GPM era, when MW data will be available at higher temporal frequency. Even if MW retrieval can be improved to some extent by using physical algorithms and a synergistic use of satellite-borne radar, as envisaged by the GPM strategy, VIS–IR retrieval may contribute, in the frame of a multiplatform algorithm, especially when high spatial and temporal resolutions are required.
This work is jointly funded by the MIUR program AEROCLOUDS (L. 204 05/06/98) in the frame of FISR funding, within the Project “Protezione Civile dalle Alluvioni: il Nowcasting” funded by the Italian Space Agency (ASI). The radar precipitation maps, cloud-top height data, and parallax correction routine were kindly provided by the Met Office.
Corresponding author address: Federico Porcù, Dept. of Physics, University of Ferrara, via Saragat 1, 44-100 Ferrara, Italy. Email: firstname.lastname@example.org