Speaker
Description
Rain forecasting in Africa is essential in areas such as agriculture or flood and drought prevention. However, as rain gauges network are expensive to set up and maintain, unfortunately, a very limited number of them are operational in most African countries. The consequence of the lack of in-situ data negatively impact rain forecasting spatial and temporal accuracy. In this work, we present an alternative method for analyzing and nowcasting rain based on mobile telephone networks data, coupled with artificial intelligence (AI).
Our study focuses on nowcasting rain prediction in the city of Yaounde in Cameroon, using data from the Rain Cell Africa project supported by Institut de Recherche et Développement Toulouse (IRD) combined with satellite imagery from Meteosat Seconde Génération (MSG). Our data spans 8 months (05/2016 to 03/2017) and covers the entire country, with average rain values every 15 minutes.
At first, the number of inconsistencies in the Rain Cell Africa dataset proved particularly challenging. This issue was overcome by applying several statistical pre-treatments, by, for example, using daily rain patterns to reduce the number of missing data in time series. Secondly, through statistical analysis, we determined the spatial and temporal correlations inside each rainy period in different cities, allowing us to identify the movement of rain clouds. These results informed our data selection and model design. We benchmarked several AI modeling techniques to forecast rain, including Random Forest, SARIMA for time series, ElasticNet regression and Neural Networks. All these models provided rain forecasts that proved better than the persistence model. In particular, recurrent neural networks of the LSTM type yield very satisfactory results, especially for predictions after 45 minutes, with a mean squared error (MSE) of 0.0085.
We attempted to improve our models by incorporating cloud information thanks to satellite imagery, however, this has so far shown minimal improvements in the forecast accuracy. We present our approach to combining Rain Cell data and satellite imagery, and our prospectives for future development.
