Pilot summary: Using AI to scale access to forest carbon markets
A blog by Ian Vickers, a Frontier Tech Hub coach
The Problem
Tanzania boasts Africa's third-largest forest cover but faces the world's fifth-highest deforestation rate. This situation presents a prime opportunity for Tanzania to access global carbon markets, securing vital climate finance while also addressing deforestation. However, local authorities and project developers encounter challenges regarding the cost, accuracy, and timeliness of generating forest carbon data.
Improved access to affordable, high-quality data could significantly support exploration of carbon credit markets as sources of revenue for the sustainable growth of Tanzania's forest carbon sector.
The solution
The pilot sought to develop a machine learning model for detection of deforestation, using satellite data of the Rufiji Delta, that harbours East Africa’s largest mangrove forest. This initiative aimed to determine if an accurate machine learning model could provide accurate data to support Tanzania’s forest carbon sector.
Digital Monitoring, Reporting and Verification of forest carbon data in Tanzania, and East Africa more widely, is often led by foreign experts, using models initially developed on data from alternative geographies. This pilot sought to explore whether local AI expertise could be developed and mobilised, to produce adapted models that were relevant for the Tanzanian geographical context, with a view that they might be maintained by local experts for longer term sustainability.
Goals of the pilot
Build a local pool of AI talent through practical experience, online learning, and knowledge transfer from international experts.
Support Tanzanian AI talent to develop a platform that provides real-time monitoring of mangrove forest cover in the Rufiji Delta, using open source software and freely available data.
Engage forest conservation stakeholders in Tanzania to understand the longer term needs and requirements for a monitoring tool, and check the potential for the solution to meet these requirements.
Key activities
The pilot conducted a range of key activities, over the series of four sprints. Broadly this consisted of:
Mobilising local AI talent through recruiting both skilled software developers, and newcomers into local ‘AI Chapters’, which were to function as peer support groups, and providing these chapters with an AI learning module.
Setting local AI chapters ‘Open Innovation Challenges’, whereby each chapter was tasked with delivering a new machine learning model that addressed a locally relevant issue. At this stage some chapters were tasked with developing a model for detecting deforestation.
Inviting top performers from the Open Innovation Challenges to a ‘Advanced Challenge’ where they were tasked with developing a more sophisticated model for monitoring deforestation in the Rufiji Delta.
Commissioning an independent expert to review the deforestation monitoring tool, and its potential to be utilised by different forestry actors within Tanzania.
Outcomes and key findings
The pilot was able to effectively mobilise a new team (made up of individuals with a mix of technical experience) within Tanzania to develop an AI model that was able to accurately detect deforestation events using satellite data of the Rufiji Delta.
The unique approach utilised by the pilot to mobilise local talent, through creating local chapters and providing them with online learning, peer support and opportunities to participate in group projects, proved an effective enabling model.
During the process of developing the model, the team found that existing machine learning models required considerable adaptation to effectively detect deforestation in the Tanzanian context, and that a lack of available ground truth data created considerable challenge. In the case of Tanzania, the pilot found that there is a need for technologists that have a long-term commitment to develop and adapt solutions for the local geographical context. In this regard, the strategy undertaken through this pilot - of nurturing and mobilising local AI talents, with an understanding of the local context, and interest in developing and sustaining locally relevant solutions - is likely to be particularly valuable.
An independent assessment of the tool created through the pilot identified that it did not yet have the full range of wider features required to meet any given forest monitoring use case, but that it did provide a critical piece of functionality that many actors require.
Further work is required to validate the assumption that the solution is significantly more accurate at detecting deforestation in the Tanzanian context than any other existing models (and therefore whether it holds a competitive advantage), and to identify how best to support its utilisation. Given the wide range of technology platforms in existence to support forest use cases (including DMRV platforms for carbon market actors), this could likely involve developing the tool so that it can be integrated into existing solutions.
