South Africa - Using machine learning to map bioclimatic zones and crop yields in water-scarce areas

Hillary Mugiyo, South Africa, Centre for Transformative Agricultural & food systems, University of KwaZulu-Natal
Mapping high-risk agricultural drought areas is critical for informing policy and decision-making to formulate drought adaptation strategies. This study used the Vegetation Drought Response Index (VegDRI). This hybrid drought index integrates the Standardised Precipitation Index (SPI), Temperature Condition Index (TCI), and Vegetation Condition Index (VCI) to delineate bioclimatic zones with both high rainfall variability and water scarcity for South Africa. Historical satellite climate data (1981-2019) was used with land use/cover maps of South Africa to generate five scales ranging from very severe to no drought. A machine learning algorithm, the Classification and Regression Tree (CART) in R and ArcGIS, was used for analysis and map graphics. Average sorghum yields obtained at the district level were used to validate results obtained from the mapping exercise. The VegDRI (74.1%), VCI (71.8%), TCI (66.2%), and SPI (59%) showed higher performance in explaining sorghum yield, respectively. Over 50% of South Africa's land experienced droughts of different magnitudes. The predictive accuracy of drought risk maps was computed from the pixel-by-pixel comparison. However, high accuracy values from Kappa of VegDRI with VCI (0.80-0.98) and TCI (0.72-0.90) do not necessarily indicate an accurate mapping of drought risk maps. VegDRI is a helpful index in designing climate-smart practices for improved food and nutrition security under increasing water scarcity.