Farming Smarter with AI in Africa
Can Artificial Intelligence be used for successful drought prediction in Sub-Saharan Africa?
Droughts — what we describe as regional periods of water deficits in some stage of the water cycle — are natural disasters that have become increasingly common over the last 30 years as a direct result of anthropogenic climate change [1]. Yet, the impacts of droughts are not evenly spread across the globe. With over 50% percent of world-wide drought occurrences between 2001 and 2011 taking place in Africa [2], and droughts leading to more deaths worldwide than any other physical hazard [3], mounting evidence suggests that they will likely have disproportionate impacts on the African population and economies.
For the estimated 33 million smallholder farmers in Sub-Saharan Africa[4], this is a prime example of how many of the most vulnerable people worldwide are experiencing the heaviest effects of climate change — right now. Advances in applying artificial intelligence (AI) to help improve the accuracy and accessibility of drought predictions may be able to change the depth of these impacts, as ongoing innovations are suggesting in Sub-Saharan Africa.
Far-Reaching Impacts: The Western Cape of South Africa
Take a recent case: Between 2017 and 2019, the Western Cape of South Africa experienced a period of drought leading to dam levels dropping to below 20% of their full capacity. Subsequent significant regional water restrictions decreased food production and led to the economic loss of billions of Rands in the Western Cape [5].
This, in turn, had a massive effect on the overall economy, given nearly a quarter of the total agricultural GDP of South Africa comes from the Western Cape. Food pricing, too, was impacted by the drought: The cost of staple crops such as maize rose, leading low-income households to spend up to 34% of their total income on food [6]. In addition to these direct economic costs, 25,000 jobs also disappeared from the national agriculture sector [7], leading to widespread, acute food insecurity across the country. The impacts, however, do not stop there.
Instances of drought like this also force increased spending on imports and disaster relief at a government level. Such spending results in depleted public funds for tackling other critical issues including inequality, poverty and unemployment, all of which are further exacerbated by drought. For example: Smallholder farmers in rural communities are heavily reliant on subsistence economies. Instances of drought and erratic weather conditions cause catastrophic effects on their livelihoods and food security as described in the recent case of the Western Cape and wider South Africa. These people are disproportionately affected as their livelihoods are particularly vulnerable to changes in climate and any existing food insecurity is made worse by inadequate infrastructure in rural areas [8].
Predicting Droughts: From Dynamical Models to AI
In searching for innovative, integrated solutions, Artificial intelligence (AI) presents one possible pathway forward. AI shows promise as a particularly useful tool in helping to combat the effects of drought as it can be used in the development of Drought Early Warning Systems (DEWS).
DEWS take in data that AI uses to predict future instances of drought, enabling farmers to make more informed cropping decisions. Before the implementation of these AI solutions, dynamical and statistical models were used that output seasonal predictions of either above-, below- or close to normal rainfall, but with no indication as to the details of the event (such as how long it will last or how severe it is predicted to be) [9].
On top of the forecasts from the dynamical models, the Effective Drought Index (EDI) would be used as a classifier to quantify instances of drought by indicating the extremity of a drought or a flood. While the EDI gives accurate details for past and current instances of drought, including their severity, duration and the available soil moisture in a given location, they are not useful on their own for drought prediction and thus provide little benefit to farmers.
Here is where AI enters the equation. Artificial Neural Networks (ANNs) are algorithms that mimic the working of human neurons in the brain. ANNs operate by taking in training data at the input, classifying it and learning trends within hidden layers of neurons which are then passed to an output. Once they have been trained, they can then be used to make predictions by inputting unseen data. As such, they are extremely effective in taking in large amounts of data to solve complex physical systems.
For the purpose of drought prediction, the EDI can be used as the training data. The EDI is calculated using historic precipitation data and fed into an ANN in order to predict future EDI values and thus future instances of drought. This method was implemented in Kenya using 30 years’ worth of daily precipitation data from four weather stations, amounting to a total of 45260 records, to calculate the EDI input data. Obtained results from this study had accuracies ranging from 75%-98% [10]— an impressive improvement over previous prediction models.
Although the results obtained in the Kenya study were accurate — and this method for drought prediction is much more reliable and informative than previous dynamical models — there are still potential challenges. Most significantly, the data used for training the Kenyan model were only collected from four weather stations in Kenya. Such selective data inputs could result in the model predicting accurately for these specific areas but performing poorly when applied elsewhere. This is similar to only deciding whether to bring your umbrella on a walk by looking outside through one window: It may be sunny where you are, but you don’t know if it is raining in other areas.
A potential solution to this issue would be to input data into the model from a larger range of areas or ensure that data from the area where the prediction is being made is included in the training data for the model. However, due to a lack of data availability in many places across Sub-Saharan Africa, this is often not possible. This lack of data is a systemic problem for drought prediction in Africa given the ongoing sparse availability of climate data [11].
Prediction Roadblocks: Weather Data Availability and Accessibility
The precipitation data required for use in DEWS are sourced from classical weather stations, located across Africa. However, the coverage from these stations across the continent is sparse, creating an incomplete picture of Africa’s overall climate at a local level. These weather stations are also largely located in cities or along large roads. This results in what available coverage does exist being uneven and with nearly none available in rural areas where the largest number of smallholder farmers are located. To apply AI drought prediction with a higher degree of accuracy in these rural areas where it is most needed, this lack of data must first be addressed.
In addition to the availability of climate data disproportionately impacting those most vulnerable to the effects of climate change, problems also exist around weather data accessibility, resulting in even less data being available for use in DEWS.
Many weather data sets, especially in Africa, are hard to access due to legal restrictions, high access costs, and, in some cases, the historical data remaining in non-digital forms [12]. Currently farmers across Africa rely on indigenous knowledge of weather indicators and mitigation strategies, which have been used successfully and passed down generations for centuries, to predict weather and inform farming decisions. While these are considered a valuable source of data that could be used more widely in drought prediction — especially when coupled with AI-based drought predictions — they presently remain an under-researched area [13].
These indigenous indicators alone, however, are insufficient to cope with the increasingly variable climate and instances of drought witnessed in recent years. Between questions of availability and accessibility, more local-level climate data across Africa is urgently needed especially given the continent’s susceptibility to extreme variability in climate conditions and subsequent negative effects.
Promise for the Future: Community-Driven AI Solutions?
The ITIKI Drought Prediction tool, founded by Muthoni Masinde based on her PhD research from 2012 and officially launched in 2019 [14], is one recent innovation seeking to combine indigenous knowledge and computer science methods to provide an effective and affordable drought prediction tool to smallholder farmers.
As a community driven solution, the project is designed to enable farmers in rural areas to get reliable drought forecasts with the goal of better informing their cropping decisions. Masinde’s ITIKI Drought Prediction tool uses the same method of calculating EDI to train ANNs in order to forecast drought, but also augments data from weather stations with localised weather sensors located in rural areas and additional indigenous knowledge. Examples of indigenous indicators include the sighting of migratory birds or the blooming of certain flowers may inform a prediction of weather events such as precipitation.
Unlike using AI or dynamical models alone, making the ITIKI Drought Prediction tool work requires the connection and participation of whole communities to provide the essential indigenous knowledge as well as set up and maintain remote weather sensors. A core premise of ITIKI is this focus on community relevance and involvement — something that too often goes missing when the technology itself is prioritised over its environment.
The results from the ITIKI prediction tool are sent out via SMS and updated on a web portal, addressing previous access issues for farmers in remote locations. Results so far suggest that the accuracy of ITIKI’s forecasts is 70–98% [15], giving farmers additional, highly accurate information they can use to make their cropping decisions. This not only leads to increases in agricultural productivity and incomes, but also allows communities to develop better resistance to climate change and aids in other important tasks like improving school attendance, nutrition, and access to clean water. The ITIKI drought prediction tool is currently being used by thousands of farmers in Kenya, Mozambique, and South Africa — 70% of whom are women, who comprise a large number of resident smallholder farmers [16].
While a newer innovation, ITIKI serves as an example of how AI can assist community driven drought prediction solutions that prioritize the needs of local communities, valuing their indigenous knowledge as a solution and strengthening it through the tools of AI. The success of ITIKI indicates that AI tools have the potential to greatly aid drought prediction in Sub-Saharan Africa and warrant further interest from researchers and governments. However, it is important to keep the needs of those on the front-line at the forefront of any advances, ensuring they have a stake in and are benefitting from the new technology whilst still enabled to retain their traditional knowledge and autonomy.
Key takeaways (TL;DR):
Artificial intelligence (AI)-based drought prediction models have promise, but two key issues need to first be addressed: (1) The lack of local weather data in the rural areas where most smallholder farmers are located, and (2) the ongoing need for community involvement. Since the data produced from these AI models is a tool, it is still up to farmers themselves to decide how and when to use it to make more informed cropping decisions.
About the Author: Katie Green
Katie Green is a first year MRes+PhD candidate in the multi-disciplinary AI4ER programme at the University of Cambridge. Her current research interests focus on how Artificial Intelligence can be applied to the problem of environmental risk through helping build resilience to environmental hazards and managing environmental change. Katie previously studied Physics at Durham University (graduating in 2020), and is a research analyst at GDI.
