Maurice Mugabowindekwe

Remote Sensing of Trees Integrating deep learning and high-resolution images to map individual trees in East Africa

The world is halfway through the UN Decade on Ecosystem Restoration aiming to “Prevent, halt and reverse the degradation of ecosystems”. Key agreements, such as the Sustainable Development Goals, Paris Agreement, and Bonn Challenge emphasise the restoration of forests and tree-dominated ecosystems due to their critical ecological, social, and economic services, particularly to mitigate climate change and biodiversity loss. Planting new trees has been central to these efforts, with global pledges to plant at least one trillion trees since the Paris Agreement in 2015. Trees offer essential benefits, including carbon sequestration, biodiversity support, income generation, and improved livelihoods.  

Despite these efforts, most restoration initiatives lack robust systems for consistent monitoring at the individual tree level across large scales. Reliable data on tree location, count, size, biomass, and carbon stock are crucial for optimising restoration efforts, resource allocation, and evaluating impacts. Such a monitoring system could also identify areas for new trees and predict the potential restoration outcomes prior to implementation.  

This PhD thesis developed methods for large-scale, tree-level monitoring to inform restoration strategies, using deep learning and high-resolution remote sensing images. It contains four studies which demonstrate these approaches across East African ecosystems, including forests, savannas, farmlands, wetlands, and urban environments. The first study focused on Trees Outside Forests (TOF), showing how these often-overlooked trees can be mapped using deep learning and satellite imagery, unlocking their potential as natural climate solutions. The second study extended this methodology to Rwanda, producing comprehensive maps of individual tree location, crown size, biomass, and carbon stock. It revealed that about 38% of trees containing 25% of the national carbon stock are found outside manually delineated forest boundaries. 

The third study highlighted the role of individual tree mapping in achieving climate goals, such as national “net zero by 2050” in Rwanda. It found that smallholder farmers planted about 50 million new trees over a decade, and presented scenarios for optimising agroforestry and forest restoration to enhance ecological and socio-economic benefits towards the national carbon neutrality by 2050. The fourth study developed a deep patch-level regression approach, integrating PlanetScope, Sentinel-1, and Sentinel-2 imagery to estimate tree counts across Rwanda, Burundi, and Tanzania. This method accounts for young and small trees, addressing gaps in conventional monitoring. The study estimated 6.9 billion individual trees in the region, revealing that only 53.3% fall within "tree cover" areas as classified by global land cover datasets, underscoring the critical need for tree-based restoration monitoring.