Human pressure on nature has reached unprecedented levels, creating an urgent need to measure and monitor its impact. Accurately mapping this pressure is crucial for environmental monitoring and conservation. Traditional methods rely on heuristic combinations of static geo-indicators (e.g., land cover, population), but they are limited by their dependence on potentially inaccurate and low-resolution data, leading to inconsistent assessments. Earth Observation offers a wealth of satellite data to address this task, yet it introduces its own set of challenges. It is often unclear which data modalities are most effective for a given task. Furthermore, the deep learning models applied to this data must be tailored to understand the unique intricacies of satellite imagery and the specific demands of naturalness mapping - a fundamentally ambiguous task, as the concept can only be approximated through imperfect geo-proxies. Moreover, ensuring these models generalize reliably and remain robust when encountering the inevitable distribution shifts inherent in real-world satellite data presents a significant hurdle. Addressing these challenges is key to producing machine learning-derived map products at a large spatial scale, where they can be most impactful. This thesis presents a complete, end-to-end methodology for mapping human influence from space - a task termed here naturalness mapping, representing the absence of modern human influence - from dataset curation to the deployment of a machine learning-derived large scale map product. The work first establishes the task by introducing a rule-based Naturalness Index as an annotation source, accompanying a globally-sampled dataset named MapInWild. It then details an interpretable machine learning methodology for identifying the most influential input modalities for the task. Subsequently, it tailors a model architecture suited to the unique demands of naturalness mapping from satellite imagery by incorporating spatial-contextual awareness and geographic location priors. The resulting model is then deployed at a continental scale. Acknowledging the ambiguity inherent in mapping naturalness, alongside the critical challenge of model robustness and generalization, the deployment pipeline integrates a distribution shift detector capable of identifying near-distribution covariate and semantic shifts. This tool produces a companion Reliability Map product that provides end-users with a crucial layer of confidence for the continental-scale, machine learning-derived Naturalness Map product.     «

Human pressure on nature has reached unprecedented levels, creating an urgent need to measure and monitor its impact. Accurately mapping this pressure is crucial for environmental monitoring and conservation. Traditional methods rely on heuristic combinations of static geo-indicators (e.g., land cover, population), but they are limited by their dependence on potentially inaccurate and low-resolution data, leading to inconsistent assessments. Earth Observation offers a wealth of satellite data to...     »