Decoding tropical forest change at fine scale with multi-sensor satellite data and artificial intelligence

Tropical forests are of crucial importance but are seriously threatened by human-driven deforestation and forest degradation. To better manage and protect tropical forests, we need fast and precise satellite-derived information on forest change, its causes, and its impacts. Current remote sensing-based monitoring systems commonly focus on generic detections of tree cover loss only, without thematic distinction of direct drivers, and often have a critical blind spot for selective logging activities, which are widespread across the tropics. This limits our understanding of the ecological and climate impacts of forest changes, and hampers the options to avert illegal or unsustainable activities in a timely manner. In this thesis, deep learning methods were developed based on multi-sensor Sentinel-1 radar and Sentinel-2 optical imagery for driver-specific tree cover loss monitoring across the pantropics, and for detailed monitoring of road development and associated logging activities in the Congo Basin forest. Beyond method development, these approaches were applied wall-to-wall to obtain a deeper understanding and assessment of recent forest changes.The results revealed that tree cover loss is driven by a wide range of processes across the tropics, with local-scale complexities, spatiotemporal interactions, and strong seasonality. While large-scale agriculture and wildfire were confirmed as the dominant drivers of tree cover loss, the fine-scale monitoring approach brought a novel perspective on the critical role of road development, selective logging, mining, small-scale agriculture, and natural disturbances. Furthermore, the near real-time delivery of these driver attributions to tree cover loss detections presents a critical step forward to advance the actionability of alerting systems as tools for rapid responses against illegal or unsustainable forest activities. A deeper look specifically into road development and associated timber harvests revealed the extent and impact of selective logging activities in the Congo Basin forest. The results showed that up to 10,000 km of new roads are opened up annually in the region. The majority is located inside concessions and built for logging purposes, with notable encroachment into intact forests. Based on the lengths of logging roads and surrounding satellite-detected canopy disturbances, it was estimated that approximately 14.7 million m³ of timber is harvested annually. This results in estimated emissions of 88.2 TgCO₂, of which 31.1 TgCO₂ can potentially be avoided with the adoption of reduced-impact logging for climate change mitigation practices (RIL-C).In conclusion, this thesis presents timely and detailed satellite-based monitoring, characterizing, and assessing of tropical forest changes to enable a better understanding of the causes and impacts of deforestation, forest degradation, and logging activities. This aids the implementation of climate change mitigation initiatives and policies, and can assist local forest management, forest conservation, and law enforcement to avert illegal or unsustainable activities.