Can Machine Learning be used to help rural communities adapt to Climate Change?

16th June 2023

Celia Petty

Co-founder and Director of Operations

Dai Clegg

Senior Associate and director of open-source software projects

William Shields, Celia Petty & Dai Clegg

Evidence-for-Development (EfD) aims to reduce poverty by improving the quality and effectiveness of livelihood information available to decision makers at every level, from local to global, and to  researchers working across multiple disciplines .

Focusing currently on sub-Saharan Africa, EfD combines high tech innovation with on-the-ground experience. We build our work around three pillars: action-oriented research at community level; capacity building with national partners, including governments, NGOs and universities; and creating new tools to reduce the costs of data collection.

Our project with DataKind UK, described in detail below, is designed for precisely this purpose: to reduce the costs of data collection, and make it more accessible and available to the widest possible user groups.

As part of its wider mission and supporting this project, Evidence for Development has also partnered with researchers at Reading University’s Walker Institute in efforts to create and accurately model climate impact scenarios. This work is increasingly urgent as climate change both increases the number of extreme weather events, which in turn erodes the resilience and adaptive capabilities of people in rural communities, and simultaneously forces those same communities to try to adapt to changing typical conditions.

Gathering information about people’s livelihoods is a critical part of this process – without it, it’s not possible to predict how they will be affected economically by events like droughts or floods. We use the Household Economy Approach ‘HEA,  to collect, model, and analyse this information. The first step involves dividing the country into Livelihood Zones (LZs), the second step is doing the survey, leading to the third step data which is data analysis – resulting in high quality information for better policy decisions. The HEA has been adopted as a method of famine early warning by many governments and humanitarian agencies across sub-Saharan Africa.

LZs are agro-ecological areas, based on land use, climate, rainfall, markets and other economic information. For example, fishing is likely to make up an important part of the livelihoods of lakeside communities, while livestock may play a more important role in semi-arid areas. Identifying distinct livelihood zones is currently carried out through multiple (up to 50) in-person regional workshops, with extensive preparation work, and is time consuming and costly. Needless to say, it is time consuming and costly. However, it is only possible to move on to the next stage of information collection – the detailed drill down into the specific vulnerabilities of different sections of the population with their various livelihood strategies (eg vulnerability to drought among farmers who rely on vegetable sales for cash income)  – when this stage has been completed. Automating this initial livelihood zoning activity, the focus of our recent project with DataKind UK, will save the national and international organisations that EfD works with time and money, and free up resources to implement programmes that will actually strengthen resilience to climate change and protect lives and livelihoods.

The Project

EfD’s goal was to use machine learning (ML) techniques combined with geospatial datasets to create LZs, and so greatly lower the time and cost spent on this part of the process. The idea to use satellite data to predict (LZs) began several years ago, with demonstrations of its uses at Reading’s Walker Institute. Automating or speeding up Livelihood Zones mapping could save thousands of person-hours on every project, dramatically reducing the cost of applying these advanced surveying techniques. EfD presented Datakind UK, a charity that supports other social sector organisations to make responsible use of data and data science,  with the challenge we faced. Datakind recognised the potential impact of this work and enthusiastically took up the challenge.

Working with DataKind UK

We were excited to engage with Datakind’s network of like-minded highly-competent professionals to have a real shot at solving our problem. Their case studies are a smorgasbord of causes for good, supporting vulnerable lives across the planet so we knew we were talking to the right people.

The initial meetings were exactly as we’d hoped: high energy, exploratory and curious, with wide-ranging debate on how the problem might be solved and how we might collaborate to solve it. The feedback we received was positive, but we were also told our project didn’t quite fit the DataKind UK typical mould. They had something slightly different in mind for us: a six month plausibility study to see if it really could be done, before we moved on to a more formal long-term programme known as a DataCorps project.

And with that we were off. A diverse volunteer team of project managers, software engineers, data scientists  and geospatial experts sprang into action in their spare time, helping us to begin sourcing data, applying rigour in our data storage and processing, and introducing automated tests in our software stack. Before we knew it, we had a rich dataset covering temperature, rainfall, altitude, land cover, and others. The automated pipeline pre-processed the raw data to produce datasets in the same format, tested and ready to support modelling activity. Example heatmaps for Malawi, depicting the aforementioned geophysical variables can be seen in Fig. 1 below.

Fig. 1: Malawi heatmaps depicting various geophysical parameters sourced from satellites.

We debated different approaches and then split off to consider both supervised and unsupervised machine learning techniques. The initial results, an example of which is shown in Figure 2 below, were promising from both techniques, but ultimately did not reach the level required to prove the approach would always work in all cases. Particularly disappointing was the application of some techniques from one country (Malawi) into another (Uganda), but, following discussion with experts, reasons linked to the nature of rural activities, as well as market trends, became clear.

Fig. 2: Supervised model example

Yet despite these setbacks, we had a starting point. In particular it gave us a rough method to take back to Livelihood Zoning experts who could begin to drill down into the areas where the models had failed, and why this might be. The richness of these conversations led us to new understandings, and a host of next steps that will take us beyond our plausibility study into full implementation.

HEA experts working in different contexts and with different agencies (humanitarian NGOs; USAID’s Famine Early Warning System Network (FEWSNET) together with EfD’s own Africa-based associates were consulted as part of this project and shown initial results. They all stressed its importance in delivering strategic insights into vulnerable rural communities and  in cutting the cost of assessments. Their observations on final ‘on the ground’ verification has also been invaluable.

New datasets and modelling steps have now been identified. And it seems clear that our methods can add real value, even if we didn’t find a silver bullet in the first place we looked. Research is, of course, not just about learning where to look, but where not to look. And when there’s this much at stake, the motto of “if at first you don’t succeed try, try again” could never be more applicable.

We at EfD, along with our partners at the Walker Institute and our Livelihood Zoning colleagues across the planet, look forward to continuing to work with DataKind UK as we drive this project into its next exciting phase. The real work may in the end only truly begin once we have an approach that works and it becomes time to prove it in the field!

If you are interested in anything you’ve read in this article, or keen to learn more about EfD  or the Walker Institute please contact us here  and we’ll be happy to tell you more.

We are extremely grateful to Datakind UK and its volunteer ambassadors for all their contributions and support.

Categories: The organisation


No comments yet.

The comments are closed.