Introduction

Remote sensing data are now widely used for different purposes, and the use of satellite data for rainfall estimates has increased recently. Meteorological satellites are the only instruments capable to provide large amounts of rainfall measurements in remote areas where data are difficult or impossible to collect from the ground. For this purpose, rainfall estimates using meteorological satellite data [e.g. Meteosat Second Generation (MSG), National Oceanic and Atmospheric Administration (NOAA), Tropical Rainfall Measuring Mission (TRMM), TERRA, etc.] started since their commissioning. Several research works have been published in the literature (e.g. Stephens and Kummerow 2007; Levizzani et al. 2001; Levizzani 2003; Nauss and Kokhanovsky 2006). Programs such as Global Precipitation Climate Project (GPCP) (Ebert et al. 1996), Climate Prediction Center using Goes Precipitation Index (CPC) (Herman et al. 1997), Estimation of Precipitation using Satellite (EPSAT) (Berges et al. 2010; Jobard et al. 2011), Tropical Applications in Meteorology using Satellite data (TAMSAT) (e.g. Grimes et al. 1999), Global Atmosphere Research Programme (GARP) (Arkin 1979) have also been developed and applied to rainfall estimates. Techniques researching the relationship between thermal infrared and intensity of precipitation have been widely used (e.g. Mishra et al. 2011; Adler and Negri 1988; Tarruella and Jorge 2003) or Cold Cloud Durations (CCD) (e.g. Huffman et al. 2001; Grimes et al. 1999; Arkin 1979).

According to the precipitation processes in connection with extratropical cyclones, convectively dominated precipitation areas are characterized by a large vertical extension and a cloud top rising high into the atmosphere. As a result, the established relationship between cloud-top temperature (CTT) and rainfall probability and intensity can be applied for the detection and classification of these precipitation areas. However, it becomes also visible that the advective-stratiform precipitation areas, that form a major part of the precipitating cloud areas in connection with extra-tropical cyclones, are not necessarily connected to cold CTT. As a consequence, a CTT threshold does not seem to be effective for the detection and classification of these areas.

For that reason, some authors have suggested to use optical and microphysical cloud proprieties to identify the raining clouds even in stratiform systems from TRMM satellite (e.g. Lensky and Rosenfeld 2003a, b) or from Terra-MODIS satellite (e.g. Nauss and Kokhanovsky 2006; Platnick et al. 2003). These techniques are based on the use of information about droplet diameter, optical thickness and thermodynamical phase of clouds. A raining cloud requires a high optical thickness, a large effective particle radius and the presence of ice crystals in the top.

The arrival of MSG gave a breakthrough for the classification of clouds. The SEVIRI radiometer (Spinning Enhanced Visible and Infrared) on board MSG provides more condensed information and its frequency of observation changes from 30 to 15 min. Multispectral capacity increases to 12 channels. The spatial resolution changes from 2.5 to 1 km at nadir for broadband visible channel and 5–3 km for all other channels.

The simultaneous use of its spectral parameters provides information about microphysical and optical cloud properties. Several techniques have been therefore developed to identify the raining clouds from multispectral parameters of SEVIRI radiometer (e.g. Kobayashi 2007; Roebeling and Holleman 2009; Wolters et al. 2011; Nauss and Kokhanovsky 2006; Thies et al. 2008a, b). These approaches are not only applicable to the detection of mainly convective precipitation by means of the commonly used relation between infrared cloud-top temperature and rainfall probability but enables also the detection of stratiform precipitation (e.g. in connection with mid-latitude frontal systems).

Therefore, the objective of the present paper is to propose a new operational technique for rain area delineation in the mid-latitudes on a 15-min basis for MSG/SEVIRI daytime and nighttime data. It is applied to the complex situation of the Mediterranean climate of this region and takes into account the variation of the diurnal cycle of clouds. The technique is based on information about optical and microphysical cloud properties from MSG/SEVIRI. It is calibrated by instantaneous meteorological radar data using multilayers perceptron. Artificial neural networks (NNs), widely used in precipitation remote sensing (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.

Presentation of study region and datasets

Algeria is located on the South shore of the Mediterranean region; it is bordered on the East by Tunisia and Libya, on the South by Niger and Mali, South-West by Mauritania and Western Sahara and West by Morocco. This region has a particular orographic structure and special characteristics of the sea-land coast. Due to these geographical properties, its climate has a very complex spatio-temporelle feature (Lionello et al. 2006). Indeed, it is influenced by both the subtropical climate and the climate of mid-latitude systems (Trigo et al. 2006; Alpert et al. 2006). The spatial distribution of precipitation is characterized by a very marked North–South gradient and a very low East–West gradient. The rainy season extends from October to March, with maximum rainfall occurring during November–December. In the north, the climate is Mediterranean transit, marked by seasonal oscillations. The average annual rainfall is estimated at about 600 mm. The minimum rainfall is recorded in the southern regions. It is about 50 mm while the maximum is observed in the Djurdjura massif located in Kabylia and the massif of Edough located a little farther east, where it exceeds 1,500 mm. The study area in this work is located in the north of Algeria, on domain with a radius of 250 km (Fig. 1). In Fig. 1, the red circle shows the domain of radar which coincides with the study area.

Fig. 1
figure 1

The study area and the position of the weather radar of Setif. The red circle shows the radar domain with a radius of 250 km

Full size image

For this study, MSG/SEVIRI data together with corresponding ground-based radar data are required.

MSG/SEVIRI data

The dataset used in this work provided by the SEVIRI radiometer of Meteosat-8 in different frequency bands is collected from November 2006 to March 2007 and November 2009 to March 2010. The MSG is a spinning stabilized satellite that is positioned at an altitude of about 36,000 km above the equator at 3.4°W. The SEVIRI radiometer gives every 15 min 12 images in the 12 available channels. We selected the channels sensitive to optical and microphysical properties of clouds (optical thickness, droplet size, cloud phase) as well as to the temperature of cloud tops, and those located in the spectral absorption bands mainly affected by the water vapor. These channels correspond to bands: visible (VIS0.6), near infrared (NIR1.6), water vapor (WV6.2, WV7.3) and infrared (IR3.9, IR8.7, IR10.8 and IR12.0). The raw image (level 1.5) has a size of 3,712 × 3,712 pixels in each channel (Eumetsat 2004). This corresponds to a spatial resolution at the image center of about 3 km. Each pixel is coded on 10 bits. All pixels are geolocalized on a common grid in geostationary projection. The sub-satellite point corresponds to the pixel position (1,856, 1,856) on the image. We stored the raw data (level 1.5), i.e. the values of 3,712 × 3,712 pixels of the image, and the calibration coefficients to deduce the radiance for each pixel. For our case, we have predefined an area in the image of the Earth’s surface; it corresponds to our study region (Fig. 1).

Meteorological radar data

The radar data are provided by the ground-based C band radar network of The National Office of Meteorology (ONM). The radar of Setif is installed near to the town of Setif, at 36°11′N, 5°25′E and 1,700 m of altitude, is one of seven Algerian meteorological radars. This is a Radar AWSR 81C in Cband, its operational frequency is 5.6 GHz. The displacement in azimuth is between 0° and 360° in continuous and the movement in inclination is of −1° to 90°. Its polarization is linear and horizontal. The effective domain of radar is a radius of 250 km.

Meteorological radar data are collected at a temporal resolution of 15 min and a spatial resolution of 1 km in a format of 512 × 512 pixels. Each pixel is coded on four bits. Thus, it consist of 15 classes representing different reflectivity intensities which are all together considered as raining in the comparison with collocated satellite pixels and one class representing no raining. The physical parameter of the radar is the reflectivity factor, referred to as Z and expressed in (mm6/m3). The conversion of reflectivity factor Z into rainfall intensity R (mm/h) is obtained using the Eq. (1) adapted to our Radar and can also be converted into dBZ:

$$ Z = 300\,R^{1.5} $$
(1)

The scan interval for both data sets is 15 min. For the spatial comparison the radar data with an original spatial resolution of 1 × 1 km were projected to the viewing geometry of SEVIRI with a spatial resolution of 4 × 5 km in the study area.

Because of discrepancies between the SEVIRI data and radar data, due to differences in observation time, parallax errors and collocation errors (Vicente et al. 2002), the comparison of these types of data may be hampered. To reduce the imbalances mentioned above and find a better correlation, we performed a repositioning to SEVIRI data to coincide spatially with radar data. We also applied a resampling to radar data to have the same resolution as resolution of satellite data. The resolution is 4 × 5 km in the study region and is assumed constant due to low overlapped area observed by both sensors. Therefore, each SEVIRI pixel is collocated with 4 × 5 radar pixels. The time lag between the radar and the satellite is about 3 min. This small time difference does not require synchronization between the two data types.

Methodology

Both datasets (SEVIRI and Radar) are divided into a training and validation data set. The training data set used for the development of the technique is collected from November 2006 to March 2007. The validation data set considered for the appraisal of the proposed technique is recorded between November 2009 and March 2010.

Because of the differing information about optical and microphysical cloud properties between day-time and night-time scenes, both data sets are divided into day- and night-time scenes.

We have hence created two identical neural networks for identifying of rainfall. The first one is used during the daytime and the second one is applied during nighttime (see Fig. 5). Each neural network used in this scheme is a multilayer perceptron (Rosenblatt 1962). The target outputs of the neural network correspond to radar data. A neural network learns the input–output relationship through the training process. The learning process in a neural network is an interactive procedure in which its connection weights are adapted through the presentation of a set of input–output training example pairs.

The neural network algorithm

The kind of neural network used for this study is a feed-forward multiple-layer perceptron (MLP). The MLP has a relatively simple architecture in which each node receives output only from nodes in the preceding layer and provides input only to nodes in the subsequent layer. Thus, for the node j in the kth layer the net input l kj is a weighted average of the outputs of the (k−1)th layer :

$$ l_{kj} = \sum\limits_{i = 1}^{{N_{k - 1} }} {w_{{\left( {k - 1} \right)ij}} \,O_{{\left( {k - 1} \right)i}} } $$
(2)

where w(k−1) ij is the weight connecting the output of the node i in the (k−1)th layer to node j in the kth layer and Nk−1 is the number of nodes in the (k−1)th layer. The output of node j is a specified function of l kj :

$$ O_{kj} = f(l_{kj} ) $$
(3)

The eventual output is then a function of the weighted combination of the final hidden-layer output values.

The list of inputs is information extracted from SEVIRI. The full list is shown in Fig. 5a during daytime and in Fig. 5b during nighttime. Each architecture is consisted of four layers: input, output, and two hidden layers. The transfer function relating input to output was a sigmoid function:

$$ O_{kj} = \left[ {1 + \exp ( - 2l_{kj} )} \right]^{ - 1} $$
(4)

To ensure that the model has similar sensitivity to changes in the various inputs, all of the inputs were normalized to values between 0 and 1. For an input variable x with maximum xmax and minimum xmin we calculate the normalized value x A as:

$$ x_{A} = \frac{{x - x_{\hbox{min} } }}{{x_{\hbox{max} } - x_{\hbox{min} } }} $$
(5)

The optimum weights were determined by training against the target values. The Fig. 2 shows the structure of a neuron with the transfer function relating input to output.

Fig. 2
figure 2

Structure of a neuron of MLP

Full size image

Three-layer perceptron network for function approximation

The rainfall estimation problem can be viewed as a complex function approximation problem. The universal approximation theorem for neural network states that a two-layer feed forward perceptron network with nonconstant, bounded, and monotone-increasing continuous activation function can perform arbitrary nonlinear input–output relationship mapping (Haykin 1994). Therefore, a two-layer perceptron network can be used for the rainfall estimation problem. The above universal approximation theorem gives the theoretical justification for the approximation of an arbitrary continuous function by a two-layer (one hidden-layer) perceptron network. In practice, however, a three-layer (two hidden-layer) perceptron network works better than a two layer perceptron for the function approximation problem. This is because the interaction between neurons in a single hidden layer network makes it difficult to obtain a globally good approximation, while a two-hidden layer network isolates and thus reduces the interaction effects by solving the problem in two steps, i.e. the first hidden-layer extracts the local features of the input data whereas the second hidden-layer extracts the global feature, to make the approximations in different regions of the input space individually adjusted (Haykin 1994). Due to above reasons, multilayer perceptron networks have been chosen in this paper for the rainfall identification problem. The structure of this multilayer perceptron with two hidden layers is shown in Fig. 5.

Training and testing data generation

A representative training data set consisting of the SEVIRI data and corresponding Radar data are needed to develop a multilayer perceptron for the rainfall identification problem. SEVIRI data are applied to the network as the input and the corresponding Radar data are used as the target or desired output (Fig. 5). The training procedure for a multilayer perceptron includes two steps, namely forward propagation and backward propagation.

The connectional weights are updated during the backward error propagation according to the learning algorithm. This process is repeated until the error between the network output and desired output (radar measurements) meets the prescribed requirement. When the training process is complete, the network is ready for application. Rainfall identification can be obtained if SEVIRI data are applied to the network at this stage.

Application

The neural network MLP was created using five spectral parameters that were calculated from SEVIRI radiometer to discriminate raining cloud from no raining cloud. It was created with four layers (input, two hidden, and output) that consist of five input neurons, six neurons in the first hidden layer, five neurons in the second hidden layer and two output neurons in the output layer that represent the two classes (raining and no raining) corresponding to radar data (Fig. 5). The information about optical and microphysical cloud properties derived from SEVIRI used as input data to Artificial neural networks MLP are given as follows:

Information about “Cloud Water Path (CWP)” The effective droplet radius re and the optical thickness τ of clouds both represented by a single parameter referred to as CWP are directly related to rainfall probability of a cloud. The effective particle radius (re) defined by the ratio between the third and the second power of the droplet spectrum is taken in place of the actual droplet spectrum. The cloud optical thickness (τ) defined by the integration of the extinction coefficient integrated over the cloud geometrical thickness is considered representatively for the cloud geometrical thickness. The CWP represents the amount of water vertically integrated in the cloud and depends on the diameter of raindrops and the thickness of the cloud formed by these drops. The relationship is given by the following equation (Thies et al. 2008a, b; Nauss and Kokhanovsky 2006):

$$ {\text{CWP}} = \frac{2}{3}\,\rho \,\tau \,r_{\text{e}} $$
(6)

where ρ (g/m3) is the density of water in the clouds.

The CWP is related to the rainfall probability of a cloud and can therefore be used as a delimiter between the non-raining and raining cloud (e.g. Thies et al. 2008a, b; Nauss and Kokhanovsky 2006).

Information about “thermodynamic phase of cloud” To strengthen the above probability, some authors hypothesize that a cloud is more likely to produce rainfall for frontal processes if the water droplets and ice crystals coexist (phase of the cloud) (Thies et al. 2008a, b; Roebeling et al. 2007; Feidas and Giannakos 2010, 2011; Levizzani et al. 2001; Levizzani 2003). Indeed, the frontal systems are mainly associated with particles of ice in the upper clouds according to the Bergeron–Findeisen process (Houze 1993). Therefore, the intensity of precipitation in clouds is related to thermodynamic phase of cloud and CWP. This means that precipitating clouds are characterized by a large enough CWP and the existence of ice particles in the upper part of the cloud.

The technique we present for the identification of precipitating clouds is based on these assumptions. It is a technique based on optical and microphysical cloud properties that was originally developed and applied to SSM/I data (Special Sensor Microwave/Imager) (Wentz and Spencer 1998). We will apply it to the radiances measured in different bands of SEVIRI. Reflectances and brightness temperatures are used during the daytime (in the presence of solar radiation) and only brightness temperatures are used during the nighttime.

Identification of the precipitating cloud during daytime

The CWP (i.e. values of re and τ) considered for a rainfall intensity differentiation can be retrieved on a pixel basis during daytime using a combination of two solar channels (i.e. a VIS and a NIR channel). Radiances obtained in the band (0.4 and 0.8 μm) are usually used to estimate the optical thickness of clouds (Arking and Childs 1985). While the band between (1.6 and 3.9 μm) is a function that depends on effective droplet radius of cloud (Baum et al. 2000; Arking and Childs 1985). The combination of these two bands can provide useful information about optical thickness and effective droplet radius of cloud. Indeed, the information about the CWP can be derived implicitly using both reflectance RVIS0.6 of visible channel (VIS0.6) and reflectance RNIR1.6 of near-infrared channel (NIR1.6) from SEVIRI (e.g. Thies et al. 2008a; Roebeling et al. 2007; Roebeling and Holleman 2009). High values of RVIS0.6 correspond to high optical depth of cloud and low values of RNIR1.6 indicate large particles in the cloud. This means that a large CWP is obtained when high values of VIS0.6 coincide with low values of NIR1.6. However, because no operational retrieval technique is currently available for MSG SEVIRI, that is applicable to water and ice clouds, and that is fast enough concerning the 15-min scan cycle, the authors decided to use the original reflectance of cannel (VIS0.6) and cannel (NIR1.6) SEVIRI channels, instead of computed values of effective droplet radius and optical thickness of clouds. The Fig. 3 shows CWP as a function of RVIS0.6 versus RNIR1.63.

Fig. 3
figure 3

Cloud water path as a function of RVIS0.6 versus RNIR1.63

Full size image

To determine the cloud phase (CP), which is also an important parameter for the identification of precipitating clouds; some authors have used information obtained from the difference of brightness temperature ΔTIR8.7–IR10.8 (e.g. Strabala et al. 1994; Thies et al. 2008b). Others have defined a threshold of brightness temperature in the band IR10.8 to discriminate between cloud ice and cloud water (Rossow and Schiffer 1999). While Wolters et al. (2008) showed that the simultaneous use of TIR10.8 and ΔTIR8.7–IR10.8 gives more accurate identification of the thermodynamic phases of clouds. They found that the formation of ice crystals begin when TIR10.8 < 238 K and ΔTIR8.7–IR10.8 > 0.25 K. It should be noted that a cloud is likely to produce precipitation when ice crystals are present at the top of cloud.

Besides, the brightness temperature difference ∆TIR10.8–IR12.1, which is a good indicator of the cloud optical thickness, is very useful to discriminate optically thick cumuliform clouds from optically thin cirrus clouds (Feidas 2011; Inoue 1985, 1987a). Optically thick cumulus-type cloud shows the smaller ∆TIR10.8–IR12.1 due to their black-body characteristics, while optically thin cirrus cloud shows the larger ∆TIR10.8–IR12.1 due to the differential absorption characteristics of ice crystals between the two channels (Inoue et al. 2001). It is expected that optically thick and deep convective clouds are associated with rain (Inoue 1987b). Even though the split window technique is very effective in detecting and removing optically thin cirrus clouds with no precipitation, it sometimes incorrectly assigns optically thick clouds like cumulonimbus in place of optically thin clouds (Inoue 1997).

Five parameters were therefore chosen to identify raining clouds during the daytime (Fig. 5a): RVIS0.6, RNIR1.6, TIR10.8, ∆TIR10.8–IR12.1 and ΔTIR8.7–IR10.8. This is carried out during solar illumination over the study area. To avoid errors due to low solar radiation, we used the radiances only when the zenith angle of the sun is between 0° and 70° relative to the study area.

Identification of precipitating clouds during nighttime

The night-time technique is based on the same conceptual model as the presented daytime scheme. However, since no operational retrieval exists for MSG to compute the CWP during nighttime, suitable combinations of brightness temperature TIR10.8 and brightness temperature differences ΔTIR10.8–IR12.1, ΔTIR8.7–IR10.8, ΔTIR3.9–IR10.8 and ΔTIR3.9–WV7.3 are used to infer implicit information about microphysical and optical cloud properties.

The combination of the channel IR3.7 with the channel IR11 was already used to extract information about microphysical and optical cloud properties (e.g. Hutchison et al. 2006; Lensky and Rosenfeld 2003a).

Lensky and Rosenfeld (2003a, b) attempted to relate the brightness temperature difference ∆TIR3.7–IR11 of satellite TRMM to effective particle radius and to cloud optical thickness. They showed that for a cloud-top temperature higher than 260 K, precipitation is obtained when the ΔTIR3.7–IR11 is in the interval ]1 K, 4 K[. Values lower than the interval are obtained when the cloud is optically thick with small effective particle radius. Values higher than the interval correspond to a semi-transparent cloud. These two situations correspond to a non-raining cloud. Thies et al. (2008b) used the channel IR3.9 and the channel IR10.8 of SEVIRI radiometer on ice clouds and water clouds. They get the same conclusions as Lensky and Rosenfeld (2003a, b). Indeed, a precipitating cloud indicates mean values for ΔTIR3.9–IR10.8.

In general, ΔTIR3.9–WV7.3 should show similar characteristics as ΔTIR3.9–IR10.8. Because of the diminishing effect of the water vapor absorption and emission in mid- to low- tropospheric levels on the brightness temperature (BT) in the channel (WV7.3) (Schmetz et al. 2002), ΔTIR3.9–WV7.3 should be generally higher than ΔTIR3.9–IR10.8. Therefore, ΔTIR3.9–WV7.3 is expected to provide additional information about the CWP. For thin clouds with small or large particles, respectively (small or medium CWP), TIR3.9 is larger than TWV7.3 and ΔTIR3.9–WV7.3 reaches the highest values. Large particles together with a high optical thickness (high CWP) result in medium to high difference values, which are lower than for optically thin clouds. Thick clouds with small particles (medium CWP) lead to small ΔTIR3.9–WV7.3.

Therefore, we use suitable combinations of brightness temperature differences (∆T) between the thermal bands of MSG SEVIRI to infer qualitative information on CWP (ΔTIR3.9–WV7.3 and ΔTIR3.9–IR10.8) (Thies et al. 2008b; Feidas and Giannakos 2011). The CWP as a function of ΔTIR3.9–IR10.8 versus ΔTIR3.9–WV7.3 is given in the Fig. 4.

Fig. 4
figure 4

Cloud water path as a function of ΔTIR3.9–IR10.8 versus ΔTIR3.9–WV7.3

Full size image

In addition, ΔTIR8.7–IR10.8 and ΔTIR10.8–IR12.0 are particularly considered to supply information about the cloud phase. Similar to the daytime approach, rain area delineation is realized by means of the pixel-based rainfall confidence as a function of the respective value combination of the brightness temperature differences. To these parameters, we add TIR10.8, ∆TIR10.8–IR12.0 and ΔTIR8.7–IR10.8 to identify precipitating clouds as mentioned previously for daytime.

In summary, we have decided to use parameters ΔTIR3.9–IR10.8 and ΔTIR10.8–IR12.0 that provide information about the CWP, and we incorporate the ΔTIR8.7–IR10.8, ∆TIR10.8–IR12.0 and TIR10.8, which give information about the thermodynamic phase and thickness of cloud.

The parameters presented previously will be used as input data, and information about rain from meteorological radar as output data (Fig. 5). The MPL rain delineation algorithms during daytime were trained using five spectral parameters (Fig. 5a) and the MPL during nighttime were trained using five spectral parameters (Fig. 5b). These parameters were computed from SEVIRI dataset during the period November 2006 to March 2007.

Fig. 5
figure 5

Structure of multilayer perceptron rain area delineation algorithm (MLP) that combines spectral parameters from SEVIRI satellite images a during daytime and b during nighttime

Full size image

Rainfall identification results and performance evaluation

Models are validated against independent rainy days during November 2009 to March 2010, not used for training the rain area delineation algorithms. The evaluation was performed by comparison with instantaneous ground-based radar data collocated with SEVIRI data. The aim is to evaluate the potential of MLP algorithm in the identification of precipitation. The observation scenes made by the radar and satellite at a rhythm of 15 min are 15,880, most of which are non-raining situations.

To evaluate the potential improvement by the new scheme, the validation scenes were also classified by the Enhanced Convective Stratiform Technique (ECST) (Reudenbach et al. 2001) which is similar to the Convective Stratiform Technique (CST) (Adler and Negri 1988). The method CST originally developed by Adler and Negri (1988) uses a temperature threshold to identify convective cores. Adjacent pixels are identified as stratiform according to their levels of temperature and the area occupied by these pixels. This technique was applied successfully in tropical convective systems (Bendix 1997). However, it shows limits in the extratropical region (Negri and Adler 1993). Reudenbach et al. (2001) adjusted the technique CST to be applied to convective systems in the mid-latitude. The ECST additionally includes the water vapor channel temperature for a more reliable deep convective/cirrus clouds discrimination (Tjemkes et al. 1997). It is applied also in extratropical regions and is used for the identification of rain areas since these regions approximately represent the performance of many present optical rainfall retrievals.

The results were calculated at scale of pixel of SEVIRI images, in collocation with radar data. Each observation gives about 6,500 pair of pixels in co-coincidences. The evaluation parameters are determined from Table 1, in which a, b, c and d were values from a contingency table (Table 1). These parameters are calculated by using equations (7), (8), (9), (10), (11) and (12).

Table 1 Overview of the combinations in a contingency table
Full size table

  • The probability of detection (POD) measures the fraction of observed events that were correctly identified:

    $$ {\text{POD}} = \frac{a}{a + c} $$
    (7)

    The optimal value of the POD is 1.

  • The probability of false detection (POFD) indicates the fraction of pixels incorrectly identified by the satellite method.

    $$ {\text{POFD}} = \frac{b}{b + d} $$
    (8)

    The optimal value of POFD is 0.

  • The false alarm ratio (FAR) measures the fraction of estimated events that were actually not events:

    $$ {\text{FAR}} = \frac{b}{a + b} $$
    (9)

    Its optimal value is 0.

  • The frequency BIAS index (Bias):

    $$ {\text{Bias}} = \frac{a + b}{a + c} $$
    (10)

    The optimal value of the Bias is 1. It measures the overestimation or underestimation of the method. A Bias greater than 1 indicates an overestimation, while a Bias lower than 1 indicates an underestimation.

  • The Critical Success Index (CSI) measures the fraction of observed and/or estimated events that were correctly diagnosed:

    $$ {\text{CSI}} = \frac{a}{a + b + c} $$
    (11)

    The optimal value of CSI is 1.

  • The percentage of corrects (PC) is the percentage of correct estimations:

    $$ {\text{PC}} = \frac{a + d}{n} $$
    (12)

    The optimal value of PC is 1.

The statistical results of the verification for MLP and ECST are given in the following table:

The area percentage of raining cloud classified by the proposed algorithm is almost consistent with those detected by the ground-based radar during daytime (Table 2). The good POD value is accompanied by a very low POFD value and the relatively low FAR. The values are 0.78, 0.03 and 0.21, respectively. The CSI indicates a good degree of correctly classified pixels. Bias parameter shows that the technique slightly underestimates the precipitation. It indicates a value of 0.69. The accuracy score shows that large fractions (PC 98 %) of the pixels are correctly identified as rainy or non-rainy by MLP. However, it is recognized that this result is heavily influenced by the high occurrence of non-rainy pixels. All these results indicate that the best values are obtained for MLP compared to ECST (Table 2).

Table 2 Results of evaluation parameters
Full size table

The results obtained during nighttime are almost similar to those obtained during daytime. Indeed, the area fraction of raining cloud classified by the satellite-based technique is in good agreement with the corresponding area detected by the radar. The POD indicates that on average about 73 % of the pixels classified as raining cloud by the radar are consistently identified by the proposed scheme and it is more important than ECST technique. Compared to ECST technique, the FAR and the POFD are also lower and so indicate good performance. The high CSI value further supports the overall good performance of the technique with a better value. The slight underestimation by the ECST is a bit more pronounced than MLP.

For the total scenes, the identified fraction of raining cloud reveals a positive performance of the new differentiation scheme. The better performance of the daytime scheme for identifying precipitation area is probably due to the higher information content about the CWP inherent in the VIS0.6 and NIR1.6 channel compared to the four channel differences considered in the nighttime scheme.

The accuracy score shows that large fractions (97 %) of the pixels are correctly identified as rainy or non-rainy by MLP. The Bias score (0.95 for MLP and 0.85 for ECST) indicates that both MLP and ECST have a general tendency to underestimate rainy pixels. POD shows that 77 % of the rainy area is correctly detected by MLP, while ECST detects 64 % of the rainy area correctly. FAR reports that 24 and 29 % of the pixels detected as rainy by MLP and ECST, respectively, are false alarms.

To gain a visual impression of the performance of the introduced retrieval scheme, the classified rain area for a scene from Mars 09, 2011 (12:30 UTC) is depicted in Fig. 6. Figure 6a shows the brightness temperature in the channel IR10.8. Figure 6b shows the rain area classified by both the new developed scheme and the radar. Figure 6c shows the rain area classified by both the ECST and radar.

Fig. 6
figure 6

Delineated rain area for the scene from Mars 09, 2011 (12:30 UTC). a BT10.8 image, b rain area delineated by Radar and MLP c rain area detected by Radar and ECST

Full size image

The number of misclassified pixels is more important for the ECST method than for the MLP method. This visual results support the statistics results obtained previously. It should be noted that for scenes classified by the ECST, the results show that the identification of raining clouds still unsatisfactory in the mid-latitudes.

To summarize the results of the comparison study, an overall good performance of the proposed scheme can be stated, especially concerning the high temporal resolution of 15 min and the high spatial resolution of 3 × 3 km in sub-satellite (4 × 5 km in study area). Thus, a process-oriented separation of areas according to the conceptual model introduced in this study is possible.

Conclusion

A new consistent day and night technique for precipitation process separation and rainfall intensity differentiation using MSG SEVIRI data is proposed. It relies on information about the CWP, the CP in the upper regions, and the cloud top height (CTH). This technique is based on the new conceptual model that precipitation is favored by a large CWP and the presence of ice particles in the upper part of the cloud. The technique considers information about both parameters inherent in the channel differences ΔTIR3.9–IR10.8, ΔTIR3.9–IR7.3, ΔTIR8.7–IR10.8, ΔTIR10.8–IR12.1 and brightness temperature TIR10.8, to detect potentially precipitating cloud areas during nighttime and reflectances RVIS0.6, RNIR1.6, brightness temperature TIR10.8, brightness temperature difference ΔTIR8.7–IR10.8 and ΔTIR10.8–IR12.0 during the daytime. All parameters are used to gain implicit knowledge about the CWP and the cloud phase.

Together nighttime and daytime, the new algorithm offers the great potential for a 24-h technique for rain area delineation with a high spatial and temporal resolution.

The present study has demonstrated the great potential offered by the increased spectral resolution of MSG/SEVIRI. The spectral parameters are adjusted to geoclimatic conditions of the Mediterranean region. Indeed, the incorporation of multispectral information about optical and microphysical cloud properties has improved the identification of precipitating clouds. The results of identification by the MLP method were compared with radar data. The proposed technique is more efficient than existing techniques based only on information about the cloud-top temperature. The algorithm shows encouraging performance for the delineation of raining area in middle latitudes.

A further improvement in the identification performance of a neural network algorithm can be investigated if a wider set of input/output patterns, which are representative of different meteorological and geographical situations, is provided to the network during the training phase. Another possible way to optimize the estimate capability of a NN-based system is to use multisensory information at a higher sampling rate, using the MSG data from different multispectral channels.