The First Generation of Machine Learning Applications for Tracking Climate Change Adaptation

In the Information Age, datasets are getting too large and diverse for conventional synthesis methods. This is especially true for climate change adaptation: projects are highly context-dependent and information sources are numerous but scattered. At the same time, tracking progress on adaptation is vital: it shows if sufficient progress is being made, enables practitioners to learn from prior experiences and highlights where resources are most needed. In this thesis, I contribute to an emerging literature which uses machine learning for adaptation tracking. I explore how a combination of methods like Structural Topic Modelling and various supervised learning models can be used to map and analyse adaptation evidence at scale. First, I use inquisitive evidence mapping to systematically assess the breath of adaptation-relevant evidence in the peer-reviewed literature, finding it has developed rapidly and shows signs of maturing. However, long-standing problems persist, including significant Global North/South biases. The findings closely align with the results of semi-structured expert interviews, supporting the validity of my approach. Second, I focus on adaptation policies, using a Transformers-based machine learning model to identify and classify policy studies in the scientific literature. Here too, I note substantial geographical differences; moreover, I see few signs of progress on policy implementation and structural reforms. Third, I investigate how political framings influence the executive summaries of country-level reporting to the United Nations Framework Convention on Climate change. I find evidence that countries highlight local priorities in their executive summaries; however, attention to adaptation or climate action has not meaningfully increased since the adoption of the Paris Agreement. Finally, I critically assess the first generation of machine learning applications for adaptation. I note that most efforts are encouraging but fall short of their transformative potential. I then provide suggestions for improvement and argue that the adaptation community should treat machine learning as a paradigmatic shift, rather than an extension of business as usual.