RainRunner - Machine Learning and Earth observation for reliable rainfall information in West Africa

West Africa’s economy is mainly sustained on agriculture and over 70% of crops are rain-fed. Economic growth and food security in this region is therefore highly dependent on the knowledge of rainfall patterns. According to the IPCC, the Global South will seriously suffer from climate change. As traditional rainfall patterns shift, accurate rainfall information becomes crucial for farmers to optimize food production.

The scarce rain gauge distribution and data transmission challenges make rainfall analysis difficult in these regions. Satellites could offer a solution to this problem, but present satellite products do not account for local characteristics and perform poorly in West Africa. For example, comparing the widely used TAMSAT and CHIRPS satellite rainfall products with ground data in our pilot area in the Northern Region of Ghana, we found a very poor correlation with TAMSAT and CHIRPS grossly overestimating the number of rainy days, while underestimating the amount of rainfall per event.

The RainRunner rainfall retrieval algorithm, developed within the Schools and Satellites (SaS) project, aims to overcome the lack of ground data and good rainfall satellite products through Earth Observation and advanced Machine Learning (ML). SaS is being funded by the European Space Agency as one of the pilot projects of CSEOL (Citizen Science and Earth Observation Lab). It is being developed in a cooperation between TU Delft, PULSAQUA, TAHMO Ghana, Smartphones4Water and the Ghana Meteorological Agency (GMet).

Research suggests that local characteristics are the reason for traditional rainfall retrieval algorithms to perform poorly in West Africa, where the land surface temperature and the concentration of atmospheric aerosols are higher than in other regions in the world. Hence, RainRunner will utilize information relevant to the rain process other than the traditionally used cloud top temperature, namely, cloud amount, atmospheric aerosols, soil moisture and land surface temperature. These data are derived from diverse sensors onboard ESA’s Sentinel satellites (S1, S2, S3 and S5P), as well as MSG’s Aviris. The satellite products, together with a Digital Elevation Model, will be pre-processed into datacubes to be fed to a Convolutional Neural Network (CNN) to estimate precipitation for a certain geographic point.

CNNs have shown to achieve better results when modelling complex natural processes than other ML algorithms, when provided with big amounts of data and well-designed architectures that represent the physical process knowledge. Furthermore, they have the main advantages of computing efficiency and the ability to represent processes beyond numerical simulations. The latter is essential for understanding the complex interactions between variables, therefore resulting in not only improving rainfall estimates but also in increasing our understanding of processes in poorly measured regions.

The Proof-of-Concept algorithm will be trained and validated with TAHMO and GMet ground measurements. Eventually, the training and validation dataset will incorporate data acquired by a rainfall observation network combining low-cost sensors and Citizen Science data collected by schoolchildren in Ghana.

Once operative, the RainRunner will guide agricultural extension agents, support crop insurance and ultimately contribute to economic growth and food security in the Global South.