Report on the development of a machine learning forecasting system for NDVI: A case study for Kenya.

Livelihoods in the horn of Africa have long been vulnerable to the impacts of climate variability and

extremes. The multi-year droughts and floods beginning in 2016 have exacerbated the challenges

facing regional agricultural production. The unprecedented nature of these consecutive extremes has

sharpened focus on the development of early warning systems for food production. One of the

objectives of CONFER is to improve the accuracy and resolution of sub-seasonal and seasonal

prediction systems to contribute towards improvements in early warning systems.

In recent years, sophisticated monitoring of vegetation indices have become an integral part of early

warning systems in the Greater Horn of Africa. Simultaneously, machine learning algorithms have

been revolutionalising predictive capacity in fields ranging from weather prediction to healthcare.

The confluence between the availability of high resolution satellite data, its current use in

agricultural monitoring systems and advances in deep learning algorithms suggests an opportunity to

leverage the power of machine learning for improved early warning.

We developed a set of machine learning algorithms to predict the Normalized Difference Vegetation

Index (a measure of the health and density of vegetation) in both crop and grasslands using Kenya as

a case study. We demonstrated good skill with a lead time of one month and a spatial resolution of

5km in cropland and reasonable skill in grassland. In the case of cropland, the new model was able to

capture a practically useful share of inter-annual variability at pixel scale over the entirety of the

modern satellite record. In the case of grassland, further improvement is required to reach

practically useful levels of skill. The work done here uses publicly available data sets, and is designed

for scalability beyond the case study for Kenya presented in this report.

The proof of concept presented here demonstrates the feasibility of combining satellite data and

deep learning algorithms to improve early warning systems, and subsequently, adaptive capacity.

The cropland model developed in this workstream of the CONFER project presents a clear case for

investing in optimising the development of vegetation forecasting systems for early warning – early

action.