ML-CLIMATE

ML-CLIMATE - iMproving cLimate information serviCes for sustainabLe agrIculture by integrating scientific and indigenous forecasts using Machine leArning TEchniques

Research introduction

Small-scale agriculture across many regions in the global South (such as Africa, South Asia, and Latin America among others) depends mainly on rainfed agriculture to ensure food security1. The agricultural sector in the global South employs around half a billion smallholder farmers and provides income security to secure their livelihoods. At the same time, increased hydro-climatic variability is threatening farmers’ livelihoods and the situation is exacerbated due to their evident low adaptive capacity. For smallholder farmers, knowledge of precipitation timing and the amount is crucial for operational decisions, such as when to plant, when to fertilize, or when to harvest. Currently, farmers have limited prior knowledge and access to scientific weather forecasting (SF). They use different forecasting techniques that rely on their local or indigenous forecast (IF) knowledge, which is based on agro-meteorological indicators they observe in the field.

There are more than 1400 indicators based on meteorology, astronomy, plants, and animals in more than 65 regions of the world that farmers operationally make use of these to forecast weather patterns2. Many scientists emphasize the necessity of acknowledging IF and combining the IF and SF systems, here it is called hybrid forecasts (HF)3-5. The integration of IF and SF into HF yields more accurate and skilful forecasts compared to IF and SF alone6-7. Despite the higher skill that can be achieved in HF, the novel method of developing HF is still under researched. Machine learning approaches as a subset of Artificial Intelligence (AI), on the other hand, have been widely used to forecast climate and weather events including extremes8-10. These techniques, therefore, offer a new opportunity to be used for integrating the IF and SF knowledge into HF.

Research challenges

The objective of the ML-CLIMATE is studying the potential of machine learning (ML) techniques to improve climate services using HF derived from the IF knowledge and SF knowledge obtained from climate and weather prediction models. The integration of IF with SF will be performed by testing different ML techniques, such as random forest, neural network, and log regression to deliver a skilful HF system. The ML algorithms will be trained using the indigenous forecast indicators and scientific forecasts as predictors and the observed data as response variables (Fig. 1). The skill of the forecasts will be evaluated using statistical metrics. Figure 1 shows the conceptual framework of HF integration and the proposed methodology.

The ML-CLIMATE project is a collaboration between the Water Systems and Global Change group and Meteorology and Air Quality group. The project is funded by the WUR investment theme “Data-Driven Discoveries in a Changing Climate” D3-C2, part of Wageningen Climate Solutions