Forecasting vegetation condition in pastoral communities for disaster prevention

Drought is a slow occurring natural hazard that is known to be very complex. On average, drought events affect the livelihood of approximately 55 million people worldwide, with a large proportion from Africa. Addressing the challenges associated with drought requires understanding the drought categories and the factors that influence their occurrence. These categories include Meteorological drought, Hydrological drought, Agricultural drought, and Socio-Economic drought. Of all these, agricultural drought stands out due to its direct impact on people’s livelihoods. This category of drought can adversely affect wildlife habitats, agriculture production, food security and the economy of the affected country or region. Governments and policymakers have explored early warning strategies that speculate the onset of drought and its severity. When in place, these strategies will help meet the United Nations’ Sustainable Development Goals (SGDs) relating to food security. A major challenge with these strategies was that they were expensive to maintain, and drought forecasts were mainly based on expert judgement. Addressing these challenges requires cost-effective approaches that use easy to access satellite Earth observation data with machine learning methods to forecast drought. Recent advances in high-performance computing and storage have enabled the development and implementation of robust early warning systems via machine learning. This PhD research aims to develop agricultural drought forecast models using satellite-based Vegetation Condition Index (VCI) and other agricultural drought indicators like precipitation and soil moisture. Data sampled from Landsat and MODIS satellite images were used to develop a Gaussian Process model to forecast VCI. An Auto-regression modelling method was also used in this study for comparative analysis. The forecast models were very skilful for forecasting VCI for 2 to 6 weeks lead time. To extend the forecast range of VCI beyond 6 weeks, we used information from additional hydro-climatic factors within a Bayesian Auto-Regressive Distributed Lags (BARDL) model. The BARDL approach improved the forecast range by approximately two weeks. Finally, a Hierarchical Bayesian Model (HBM) was used to model agricultural drought in regions with diverse land cover types and agro-ecological zones. Forecasts from the HBM were more accurate than the BARDL approach, with an approximately one-week improvement in the forecast range.