Satellite-based high-resolution mapping of rainfall over southern Africa

A spatially explicit mapping of rainfall is necessary for southern Africa for eco-climatological studies or nowcasting but accurate estimates are still a challenging task. This study presents a method to estimate hourly rainfall based on data from the Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI). Rainfall measurements from about 350 weather stations from 2010–2014 served as ground truth for calibration and validation. SEVIRI and weather station data were used to train neural networks that allowed the estimation of rainfall area and rainfall quantities over all times of the day. The results revealed that 60 % of recorded rainfall events were correctly classified by the model (probability of detection, POD). However, the false alarm ratio (FAR) was high (0.80), leading to a Heidke skill score (HSS) of 0.18. Estimated hourly rainfall quantities were estimated with an average hourly correlation of ρ = 0.33 and a root mean square error (RMSE) of 0.72. The correlation increased with temporal aggregation to 0.52 (daily), 0.67 (weekly) and 0.71 (monthly). The main weakness was the overestimation of rainfall events. The model results were compared to the Integrated Multi-satellitE Retrievals for GPM (IMERG) of the Global Precipitation Measurement (GPM) mission. Despite being a comparably simple approach, the presented MSGbased rainfall retrieval outperformed GPM IMERG in terms of rainfall area detection: GPM IMERG had a considerably lower POD. The HSS was not significantly different compared to the MSG-based retrieval due to a lower FAR of GPM IMERG. There were no further significant differences between the MSG-based retrieval and GPM IMERG in terms of correlation with the observed rainfall quantities. The MSG-based retrieval, however, provides rainfall in a higher spatial resolution. Though estimating rainfall from satellite data remains challenging, especially at high temporal resolutions, this study showed promising results towards improved spatio-temporal estimates of rainfall over southern Africa.

are dominated by local and short-term convective heavy showers mostly with thunder in the afternoon or evening hours. Rain from synoptic systems lasting up to several days also occurs. Snow and hail only contribute a neglegible amount to the overall 3 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2017-33, 2017 Manuscript under review for journal Atmos. Meas. Tech. Published: 6 February 2017 c Author(s) 2017. CC-BY 3.0 License. Figure 2. Map of the average yearly precipitation sums in the study area as estimated by WordClim (Hijmans et al., 2005). Points show the locations of the weather stations that were used as ground truth data in this study. Automatic rainfall stations (ARS) and automatic weather stations (ARS) are operated by the South African Weather Service (SAWS). Further stations are operated by SASSCAL WeatherNet as well as by the IDESSA project. precipitation sums. The inter-annual variability of rainfall is high for the arid areas. For a detailed description of Southern African rainfall characteristics see Kruger (2007); Kaptué et al. (2015).

Station data
Rainfall data for the years 2010 to 2014 were obtained from the South African Weather Service (SAWS). The data were 5 recorded from 229 automatic rainfall stations and 91 automatic weather stations (Fig. 2). They were complemented by 22 stations from SASSCAL WeatherNet (www.sasscalweathernet.org/) located in southern Namibia and Botswana. For the year 2014, data from an additional 15 stations in South Africa operated by the IDESSA project (An Integrative Decision Support System for Sustainable Rangeland Management in Southern African Savannas, www.idessa.org/) were available. All station data that provided sub-hourly information were aggregated to a temporal resolution of 1 hour. Though the station data is 10 4 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2017-33, 2017 Manuscript under review for journal Atmos. Meas. Tech. Published: 6 February 2017 c Author(s) 2017. CC-BY 3.0 License. not randomly distributed in the model domain, it covers the entire aridity gradient, from sites with very low (< 200mm) precipitation to sites in areas with highest (∼ 1500 mm) yearly precipitation sums.

Satellite data
MSG SEVIRI (Aminou et al., 1997) scans the full disk every 15 minutes with a spatial resolution of 3 by 3 km at subsatellite point. Reflected and emitted radiances are measured by 12 channels, three channels at visible and very near infrared 5 wavelengths (between 0.6 and 1.6 µm), eight channels ranging from near-infrared to thermal infrared wavelengths (between 3.9 and 14 µm) and one high-resolution visible channel. MSG SEVIRI data were preprocessed based on a Meteosat processing scheme that uses xxl technology and custom raster extensions which were designed to support OpenCL acceleration (see https://github.com/umr-dbs/xxl).  Stengel et al. (2014) for further information on CLAAS). All pixels that were classified as cloud contaminated or cloud filled were interpreted as cloudy. Pixels that were classified as cloud-free were masked from further analysis. MSG SEVIRI to train a Random Forest model that is able to spatially predict rainfall areas and rainfall rates over Germany.

Cloud mask
Based on this study, Meyer et al. (2016) have shown that neural networks outperform the initially used Random Forest algorithm. In these previous studies on the rainfall retrieval, the radar based RADOLAN product (Bartels et al., 2004) was used as ground truths to train the model. The high data quality and spatially explicit information allowed the model to be optimised without too many confusions caused by uncertainties in the training data. However, the goal of the retrieval was that it can 25 be applied to areas where spatially explicit data for rainfall are not available, as it is the case in Southern Africa. All steps of model training were performed using the R environment for statistical computing (R Core Team, 2016). the sun zenith were used as predictor variables during daytime. Meyer et al. (Submitted) tested different texture parameters as additional predictor variables and have shown that these spectral channels are sufficient as predictors. Since the VIS and NIR channels of MSG are not available during the nighttime, the dataset was split into a daytime dataset (all scenes where VIS and NIR were available) and a nighttime dataset (reliable VIS and NIR not available). The response variables (rainfall yes/no and rainfall quantities) were taken from the rain gauge measurements.

5
The years 2010 to 2012 were used for model training. The year 2013 was used for validation. The retrieval process was two-folded and consisted of (i) the identification of precipitating cloud areas and (ii) the assignment of rainfall quantities. All 2010 to 2012 data from the rain gauges that are masked as cloudy by the cloud mask products were used for training the rainfall area model. All recorded rainfall events were used for training the rainfall quantities model. The resulting training dataset comprised 917774 (daytime) and 1409072 (nighttime) samples for the rainfall area training and 69703 (daytime) and 10 129325 (nighttime) samples for training of rainfall quantities from 26243 individual MSG scenes.

Tuning and model training
The neural network implementation from the "nnet" package (Venables and Ripley, 2002) in R was used in conjunction with the "caret" package Wing et al. (2016) that provides enhanced functionalities for model training, prediction and validation.
Model parameters were tuned using a stratified 10-fold cross-validation where each fold has the same distribution of rainfall 15 areas, or rainfall quantities respectively, as the entire training dataset. The number of hidden units were tuned for each value between two and the number of predictor variables (Kuhn and Johnson, 2013). Weight decay was tuned between 0 and 0.1 with increments of 0.02. For training of rainfall areas, the threshold that separates rainy from non-rainy clouds according to the predicted probabilities was an additional tuning parameter. The optimal threshold was expected to be considerably smaller than 0.5 since the amount of non rainy samples was higher than the amount of rainy samples. Therefore, the range of tested

Validation
Model predictions and weather station records from the entire year 2013 were used as independent data for model validation.
For the validation of predicted rainfall areas, all pixels at the location of the weather stations that were classified as cloudy 25 by the cloud mask product were considered. Therefore the information from the weather stations about whether it was raining or not was compared to the model prediction for the respective MSG pixel. The validation data contained 403211 samples during daytime and 565415 samples during nighttime. Average hourly probability of detection (POD), Probability of false detection (POFD), False alarm ratio (FAR) and Heidke skill score (HSS) were calculated as validation metrics. The POD gives the percentage of rain pixels that the model correctly identified as rain (Tab. 1, 2). POFD gives the proportion of non-rain pixels 30 that the model incorrectly classified as rain. The false alarm ratio (FAR) gives the proportion of predicted rain where no rain is observed. The HSS also accounts for chance agreement and gives the proportion of correct classifications (both rain pixels and non-rain pixels) after eliminating expected chance agreement. HSS is independent of the bias in the classifications.
To evaluate the ability of the model to predict rainfall quantities, the correlation between the measured and the predicted hourly rainfall was calculated using Spearmans rho to account for a non-normal distribution of the data. Further, the root mean square error (RMSE) was used. All clouded data points (including non-rainy data points) were used for the validation of rainfall quantities. The rainfall quantities were further aggregated to daily, weekly and monthly rainfall sums to assess the performance of the model on different temporal scales.

Comparison to GPM
The results of the presented rainfall retrieval were compared to the rainfall estimates of the GPM mission. GPM, as a successor of the Tropical Rainfall Measuring Mission (TRMM), consists of an international network of satellites aiming at worldwide high resolution precipitation estimates (Smith et al., 2007). GPM provides data from March 2014 onwards. The GPM IMERG product estimates rainfall by combining all available passive-microwave instruments as well as microwave-calibrated infrared 10 satellite estimates and data from rainfall gauges. GPM IMERG is available in 6h, 18h and 4 month latency.
In this study the 4 month latency (final product) with 30 minutes temporal and 0.1 • spatial resolution (∼10km x 10km) was used. Due to different data availabilities of GPM IMERG, MSG as well as weather station data, the comparison was conducted for the overlapping time period late March 2014 to August 2014. GPM was aggregated from 30 minutes to 1h to match the temporal resolution of the MSG predictions. Both products were validated using the weather station data as a reference. The 15 performance metrics were compared between the MSG product and the GPM product on an hourly basis. The validation of the rainfall retrieval showed promising results but highlights also the difficulties of optical satellite-based rainfall estimates. The major problem was the overestimation of rainfall events leading to an overestimation of rainfall quantities. In this context, it is of note that the FAR can easily increase to elevated levels in dry conditions when there are just a few false alarms in the predictions and no rainfall was observed by any station. However, the FAR was still high for hours with a considerable number of rainfall events. This might be partly explainable by spatial displacement due to parallax shifts that  (Vicente et al., 2002) would be appropriate. Differences in spatial and temporal scale are also an important issue especially since a majority of rainfall events in Southern Africa are of small spatial and temporal extent. The aggregation to an hour as well as the assumption that the weather station observation is representative for the entire pixel, are also problematic though essential. The issue of scale especially affects the broader resolution GPM IMERG data where a several km sized pixel is validated by a single point measurement. Beside of the issue of scale and spatial displacement, the retrieval technique depends 5 on the quality of the rain gauge observations. Although the data was quality checked, common problems associated with rain gauge measurements e.g. wind drift or evaporation leading to errors in the ground truth data and affect model training and validation remain Kidd and Huffman (2011). However, no reliable alternatives are available and rain gauge measurements are still considered as most reliable source of rainfall data.
Despite the errors and uncertainties associated with the presented rainfall retrieval, the combination of MSG data and neural 10 networks are a promising approach. The model presented in this study outperformed the GPM IMERG product in terms of rainfall area detection where GPM IMERG considerably underestimated rainfall events. This behavior is partly explainable by scale because GPM IMERG has a coarser resolution of 0.1 • . This makes local processes difficult to capture which is an disadvantage considering that in Southern Africa especially small scale convective showers contribute to rainfall sums Kruger (2007). In terms of rainfall quantities, GPM IMERG and the presented retrieval did not show significant differences in view to 15 correlation. The sample predictions have shown that GMP IMERG has more differentiated rainfall estimates while the MSG based retrieval tends to predict the mean distribution.
The presented MSG based retrieval is an easy to use method and allows for time series at a relatively high spatial resolution.
Aside of the promising results compared to GPM IMERG, the daily estimates of the MSG based retrieval are at least compa-10 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2017-33, 2017 Manuscript under review for journal Atmos. Meas. Tech. Published: 6 February 2017 c Author(s) 2017. CC-BY 3.0 License. rable to other products incorporated in the IPWG validation study IPWG (2016). A detailed comparison could currently not be given since validation data and strategy were not identical. Incorporation of the presented retrieval to the IPWG validation study is intended by the authors for future assessment.

Conclusions
The rainfall retrieval developed in this study provides hourly rainfall estimates in high spatial resolution based on the spectral 5 properties of MSG SEVIRI data and neural networks. The retrieval showed promising results in terms of rainfall area detection and estimation of rainfall quantities. However, the results also showed that the estimation of rainfall remains challenging. The

11
Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2017-33, 2017 Manuscript under review for journal Atmos. Meas. Tech. Published: 6 February 2017 c Author(s) 2017. CC-BY 3.0 License. main weakness of the presented retrieval was the overestimation of rainfall areas. However, the retrieval could compete with the global GPM IMERG product in terms of rainfall quantity assignment and was even advantageous for rainfall area detection.
High resolution spatial datasets of rainfall are requested by a variety of research disciplines. The developed MSG based rainfall retrieval is able to deliver time series from the launch of MSG SEVIRI onward. An operationalization for near realtime rainfall estimates is intended. It can therefore serve as valuable dataset where high resolution rainfall for Southern Africa August 2014. Each data point is the average performance of one hour. The notches serve as a rough estimation of significant differences. mm Figure 9. Sample satellite scene from 2014/04/24 12:00 UTC represented as a VIS0.8-IR3.9-IR10.8 false colour composite according to (Rosenfeld and Lensky, 1998) where cloud optical depth is indicated by red colouration, cloud particle sizes and phases in green and the brightness temperature modulates in blue. The rainfall predictions for this scene are shown as well as the corresponding GPM IMERG product. Observed rainfall is depicted where weather station data were available. For visualization purposes, the spatial extent of the stations was increased. White background in the colour composite as well as in the MSG based retrieval and the GPM IMERG product represent no data due to missing clouds. In addition, white background in the representation of the observed rainfall is due to the absence of weather stations.