Building globally generalized Earth observation models is challenging because landscapes, land-use practices, and human pressures vary widely across regions, cultures, climates, and scales. My PhD project focuses on using computer vision and machine learning to develop a renewed cartography of human impact on landscapes from remote sensing imagery. Rather than treating land cover mapping as a purely visual classification problem, the project aims to capture two broader dimensions of landscape change: a gradient from intact to highly human-modified environments, and a second axis describing landscape heterogeneity. The long-term goal is to produce a map of naturalness that can support biodiversity monitoring, ecosystem restoration, and improved accounting of carbon sources and sinks across managed and natural lands.
A central scientific challenge is that Earth observation data are affected by strong geographic distribution shifts, meaning that patterns learned in one region often do not transfer reliably to another. This work therefore explores machine learning approaches that can separate globally stable signals from region-specific patterns, so models remain robust while still accounting for local context. In particular, we are interested in methods inspired by stable learning, domain prediction, and geographically aware architectures that explicitly disentangle invariant and location-dependent information. By combining these ideas with remote sensing and ecological knowledge, the project aims to build models that are both scientifically meaningful and practically useful for mapping human influence on landscapes at the global scale.
