As a result of global warming, the upper limit of alpine forests is expected to shift to higher altitudes. This threatens the habitat of numerous endemic plant species located in the alpine grasslands above. Anticipating the timing and pace of treeline shifts is therefore crucial. In the Swiss Alps, forests have been shaped by centuries of agricultural land use, thus maintaining the realized treeline below its climatic potential. Since the decline of alpine farming in the mid-19th century, a rapid expansion of forests into abandoned pastures has been observed. As treelines are shifting due to both land-use and climate change, as well as local controlling factors, understanding treeline response to warming remains challenging. Aerial historical imagery captured over Switzerland since 1946 offers the opportunity to moni- tor treeline shifts across time and space, thereby complementing field-based monitoring and providing insights into spatiotemporal patterns of treeline dynamics. Recently, advances in remote sensing techniques and deep learning (DL) have allowed the development of powerful forest monitoring tools. Building upon these methods to process time series of aerial images, this thesis aims at monitoring treeline dynamics in the Swiss Alps over an almost 80-year period to provide insights into their controlling factors. Chapter 2 tackles DL methods for high-resolution forest mapping, focusing on how they can be designed to mirror the common reasoning underlying forest definitions. The proposed approach facilitates the understanding of the modelâ s predictions, usually hindered by the black-box nature of DL models. In Chapter 3, a multi-temporal DL approach is developed to map forest cover from time series of aerial images. To overcome the lack of training annotations and to improve temporal consistency, a training loss is designed using prior knowledge about forest cover dynamics, allowing to map forest cover from recent color images as well as historical, lower-quality grayscale images. The same methodology is adapted in Chapter 4 to map forest cover from 1946 over the entire Swiss Alps. Treeline shift rates, as well as the realized treeline and its outpost positions, are then extracted. Additionally, a climate-based potential treeline is estimated from historical temperature records. Results show that treeline shifts have been accelerating since the mid- 20th century. Acceleration is now localized at treeline outposts, a pattern that emerged in the early 2000s. However, despite reaching unprecedented rates â locally up to 1.25 m/year â treeline shifts rates remain considerably lower than the shift rate of their modeled climatic potential.

This thesis highlights the relevance of integrating DL methods, remote sensing data, and domain-specific knowledge to capture the spatiotemporal patterns of treeline dynamics. The results highlight the remaining uncertainties in treeline responses to accelerated warming, further emphasizing the need for long-term, large-scale monitoring.