From Satellite Data to Policy: Deep Learning for Identifying Tropical Deforestation and Degradation Drivers

This PhD research investigates how deep learning and Earth Observation (EO) data can be used to accurately identify the direct drivers of deforestation and forest degradation, defined as the specific land-use changes that immediately lead to forest loss, in data-scarce tropical regions. Focusing on Cameroon as a case study within the Congo Basin, the research addresses the central question of whether country-specific deep learning frameworks can provide reliable, policy-relevant classifications of deforestation drivers. To address this, the research first develops a novel, nationally relevant EO dataset that integrates satellite imagery with auxiliary biophysical and infrastructure data. This dataset provides the first systematic classification of fifteen direct deforestation and degradation drivers in Cameroon and forms the foundation for model development. Building on this, the study introduces and optimises a deep learning model, Cam-ForestNet, evaluating multiple satellite data sources, spectral configurations and transfer learning approaches. The results demonstrate that deep learning can reliably identify detailed drivers of forest loss when trained on country-specific data. The best-performing configuration, using Landsat-8 Top-of-Atmosphere imagery with NIR-R-G spectral bands, achieved a macro-average F1 score of 0.77 and outperformed higher-resolution alternatives. Transfer learning from other tropical regions did not improve performance, highlighting the importance of locally tailored datasets. Incorporating temporal information further improved classification accuracy, particularly for degradation processes and slow-developing land uses, while also revealing trade-offs between near real-time monitoring and more comprehensive analyses. Finally, the integration of calibrated confidence scores enhanced the interpretability and operational usability of model outputs, enabling more informed decision-making. By integrating methodological innovation with policy-relevant applications, this research helps bridge the gap between technological capabilities and actionable deforestation mitigation strategies, contributing to more effective conservation and sustainable land-use planning. It also provides a transferable methodological framework that can be adapted to other data-scarce regions in the Congo Basin and beyond, where the identification of deforestation and degradation drivers has often remained limited to broad categories without country-specific contextualisation.