Wanting Yang, PhD 

Remotely Sensing Forest-Agriculture Frontiers  

Using deep learning and high-resolution images to map and characterize mosaic-like agriculture in the tropics 

 Understanding tropical forest-agriculture frontiers is essential for assessing how human activities transform natural ecosystems and for evaluating the consequences of these changes. The widespread narrative that agriculture is the leading driver of deforestation often oversimplifies the complex and context-specific roles that different agricultural systems play, particularly mosaic-like practices such as shifting cultivation and agroforestry, which incorporate trees within the production system.

While remote sensing has become a key tool for monitoring large-scale land cover and land use (LCLU), these mosaic systems remain underrepresented due to their fragmented and dynamic nature, compounded by challenges such as persistent cloud cover and the limited generalizability of existing methods. Yet, there is a significant research gap in mapping the fine-scale extent of such mosaic agricultural practices across the pantropical region and in quantifying their interactions with natural tree cover. Addressing this gap is crucial for improving our understanding of land-use patterns and informing sustainable land-use policy in tropical regions.

This thesis explores novel approaches for better mapping and characterizing mosaic-like agricultural systems with deep learning technique and new-generation remote sensing data. Through three first-authored studies, it demonstrates how artificial intelligence (AI) can be incorporated to detect and analyze these complex agricultural systems in tropical regions with high-resolution PlanetScope satellite imagery. By mapping the spatial extent of those mosaic-like agriculture systems and designing a framework to quantify land-use intensification, this thesis offers new insights into the dynamics of tropical forest–agriculture frontiers.

This thesis benchmarked forest-agriculture frontiers mapping focusing on mosaic-like agricultural systems. This is achieved by the combined use of advanced deep learning algorithms together with recent development in nano-satellite remote sensing technology. The studies deepen our understanding of shifting cultivation and agroforestry in terms of their spatial distribution and their underrepresentation in existing global land cover products. Moreover, this work also contributes to the broader field of Earth observation by demonstrating how deep learning can be effectively applied to complex and fragmented tropical landscapes.