Near-real-time rainfall detection and estimation from commercial microwave links and Meteosat Second Generation data

Many operational applications, including agriculture and water resources management, local climate studies, flash flood prediction and forecasting that directly affect human livelihoods, need accurate high-resolution rainfall information. High-resolution rainfall is also vital in research applications such as evaluating satellite rainfall products to inform new satellite rainfall observation missions. However, despite its great value, rainfall’s intricate characteristics, such as intermittency and spatial variation, make accurate rainfall estimation challenging.
There are existing techniques to measure rainfall, but each has an inherent limitation. For instance, rain gauges measure accurate rainfall from a point and are thus limited in their ability to measure the spatial state of rainfall. The weather radars are capable of accurate spatial rainfall information, but among other limiting factors, they are expensive to acquire, operate and maintain. Satellites are capable of global rainfall observations, but their estimates are often inaccurate. Moreover, their observations, usually averaged over squared kilometres, are not always ideal for all applications, e.g., farmers whose plots are only a few hundred squared meters.
Over a decade ago, a new technique was introduced that utilises telecom service operators' extensive mobile phone infrastructure for rainfall observations. The microwave signals beamed from one telephone tower to the other are significantly affected by rain. For this reason, researchers studied the fluctuations in the signal level recorded at the receiving tower and showed that these systems can estimate accurate rainfall; suggesting that each microwave link (MWL)–a pair of connected towers–can serve to gauge rainfall along two connecting mobile phone towers in near real-time. Albeit opportunistic and low-cost rainfall estimation technique, the problem is that the estimation accuracy is affected by many factors, including the variation of raindrop sizes along the link and the fact that the link’s antenna wetting during and after rainfall introduces additional uncertainties in the retrieved rainfall estimates. Moreover, the density of the MWL network is arbitrary and often biased towards densely populated areas. This characteristic may complicate their use for retrieving spatial rainfall patterns, especially in areas with low network densities.
At the same time, geostationary (GEO) meteorological satellites, such as the Meteosat Second Generation (MSG), frequently observe the earth’s atmosphere with wide spectral range radiometers capable of detecting rapidly developing storms, raining areas and estimating rainfall intensities. Several studies have extensively studied the data with traditional rainfall measurement techniques such as rain gauges and weather radars for rainfall observation. However, combing the GEO satellite and MWL data for rainfall observation has gained very little attention, while such an alternative could benefit many places that lack traditional rainfall observation systems.
For this reason, the main objective of this study was to investigate the MWL and the MSG satellite data for high spatiotemporal resolution rainfall detection and estimation using data from Central and Western Kenya, where traditional rainfall-observing systems are often lacking or sparse. This main objective was achieved based on 4 specific objectives, which form the basis of 4 individual research studies.
Firstly, it was investigated if the MWL-MSG data combination could improve rain rate estimation and detection. We investigated the MWL’s capability to estimate accurate rain rates in the study area using gauge rainfall data. Then, the MWL’s rain rates were studied with the MSG satellite data using a conceptual model in which clouds with high cloud top optical thickness and particle effective radius have high rainfall probabilities and rain rates. Regarding the MWL’s rainfall estimation capability, the results confirmed the robustness of the MWL rainfall estimation technique concluded by many past studies. Studying the MWL’s rain rate with the MSG satellite data revealed unique spectral characteristics of daytime, night-time, raining and non-raining data. Eventually, descriptive statistics derived from the satellite spectral characteristics successfully detected rainfall on individual MWL in the study area. However, daytime rain detection, which uses reflectance data, was better than night-time detection, which uses infrared (IR) data, owing to the better rain information about cloud optical and the effective radius contained in the reflectance data.
Following the successful combination of the MWL-MSG data for rainfall observation, the subsequent investigation developed an improved rain area detection system to improve the MWL’s rainfall estimates and map rainfall in the study area. This investigation evaluated multiple parametric rain detection models derived from MSG’s reflectance and IR data using a conceptual model similar to the previous research. Additionally, we developed a new technique that uses rain area-specific gradient parameters to improve detected rain areas by correcting the number and sizes of the detected rain area. While comparing the rain area technique with one of the most successful satellite rainfall products, the Global Precipitation Measurement Integrated Multi-satellitE Retrievals for GPM (GPM IMERG), the results corroborate in terms of the spatial dynamics of the detected rain areas and rates. Next, we improved the MWL’s rainfall estimates for the first time in a new MSG technique that uses the rain area detection system developed in the previous study. In this investigation, we developed a new technique for wet-dry classification and baseline level estimation using the MSG-based rain area information. A new wet path length (wpl) parameter, representing the length of the raining MWL, was also developed based on the MSG-based rain area. This technique’s wet-dry classification and baseline level estimation results agreed well with conventional technique results. The wpl parameter also remarkably improved the high rainfall intensities, which suggests that spatial rainfall variability over the MWL remains essential information to be considered in the MWL’s rain rate retrieval.
Finally, a new technique was developed for high spatiotemporal rainfall estimation from MSG’s cloud top properties and MWL rainfall intensities. This technique trains the random forest (RF) machine learning algorithm with the MWL‘s rainfall estimates to estimate rainfall from the MSG data. The results are convincing and promising. When compared to gauge rainfall data, the techniques’ retrieval errors were comparable to errors found when comparing GPM IMERG and the European Organisation for the Exploitation of Meteorological Satellites Multi-sensor Precipitation Estimate (EUMETSAT MPE) rainfall intensities to gauge data. The spatial dynamics of the retrieved rainfall intensities also agreed well with these satellite products. The technique’s advantage lies in retrieving high spatiotemporal resolution rainfall intensities regardless of the rainfall type.
Overall, this study demonstrates the great potential of using the MWL-MSG data for rainfall detection and estimation. In particular, the benefit of this rainfall observation alternative to areas with sparse or lacking conventional ground rainfall monitoring systems but growing MWL network and geostationary satellite (with capabilities like MSG) coverage may be invaluable.