The microwave signal between cellular communication towers can be used to estimate rainfall, because the signal between two towers is (partially) attenuated by the rain that falls in between these two towers. However, it was so far unknown whether the attenuation signal can also be used to detect precipitation type, i.e. liquid rain versus solid precipitation such as snow, hail, and sleet. This project uses machine learning techniques to distinguish precipitation types from the CML signal.
Our study uses observations from an experimental CML set-up in Wageningen including five disdrometers that report the droplet fall velocity, size and precipitation type along the path length. These disdrometer data serve as reference to a machine learning process using almost 1.5 million precipitation type observations for our study period (August 2014 till December 2015). Although the amount of data seems enormous and well suited for a machine learning project, the identification of precipitation types introduces an additional challenge compared to the more ‘straightforward’ wet- or dry classification. It appears that it actually rains less often in the Netherlands than you might think. We find resampling to obtain a balanced amount of precipitation types is needed to apply machine learning technique. With rain cases just under 10% of the time, the dataset becomes quite unbalanced with mainly dry moments. Hail and snow occur even less frequently.
Moreover the study explores whether the attenuation signal from CMLs can be used to detect fog. Comparing the fog time series and the CML signal reveals a distinct drop of the CML signal which coincides with the fog occurrence. Once again we balance the dataset because the moments without fog far outnumber the fog events. It appears that the raw data for each individual time step is not enough to create a neural network that detects fog. However, the attenuation signal in the CML data shows a pattern associated with fog.