Land cover mapping is the process of spatially labelling the Earth's surface according to its biophysical land cover type, and has uses including land-use planning and natural resource management that can be used to inform public policy. Remote sensing offers the potential for efficient provision of continuous sensor data, at a range of spatial and temporal scales, enabling detection of change. This allows land cover mapping to be applied to address a variety of environmental risks, including monitoring deforestation, forest degradation, sea ice extent, urbanisation and cropland expansion. The insights gained from land cover analysis can be used to help monitor 14 of the 17 UN Sustainable Development Goals. To date, the dominant paradigm for land cover mapping has been to use supervised learning in combination with models such as random forests and convolutional neural networks. While these approaches can accurately segment land cover, they suffer from three key drawbacks: (1) Annotation cost. To achieve good performance, large quantities of high-quality manually annotated semantic segmentation data are required; however, this can be extremely expensive to collect (e.g., 90 minutes per image ). (2) Robustness. Existing supervised approaches struggle to generalise beyond their narrow training distribution, restricting their use to small geographic regions and limited land cover types. (3) Fixed category labels. Supervised learning models can only predict labels from the fixed set of categories they are trained on, and retraining is needed to predict additional categories Due to the interconnectivity and scale of the environmental challenges we currently face, a 'big-picture' view encompassing multiple domains is needed to effectively address them. This PhD intends to provide a flexible and searchable interactive land cover investigation tool that can be used by researchers working to overcome these issues. The tool could be used to look at land cover in a (1) generalised setting (e.g., deforestation/urbanisation); (2) more specific setting (e.g., food security/forest degradation); and (3) object detection and analytics setting (e.g., locating renewable assets). Current approaches are insufficient to feasibly create this functionality; however, we believe leveraging vision-language models offer promising potential. Aims The two key aims of this project are: Aim 1. Develop computer vision models for land cover segmentation with an accuracy approaching that of a human expert. Aim 2. Develop a tool (similar in style to Google Earth1, Maps2 and electricityMap3) that provides an interactive land cover visualisation, language-driven object search, and analytics dashboard