Keynote 1: Gilberto Camara, “Geo-ontologies, Ecology and Machine Learning: Building an Interdisciplinary Understanding of Remote Sensing”

Abstract: This talk examines the limitations of current Machine Learning for Remote Sensing (ML4RS) methods—including recent foundational models—in representing the full richness of information contained in satellite imagery. Most existing ML4RS approaches rely on predefined object hierarchies and assume that class concepts are fixed and unambiguous. Earth Observation (EO) foundational models seek to integrate data sources with diverse temporal, spectral, and spatial resolutions. Such assumptions and practices, however, stand in tension with decades of Remote Sensing research and with the established epistemological foundations of Earth Science. Ontologies that describe the geographical world are inherently polysemic, with meanings that depend on context and perspective. For instance, what qualifies as a forest varies depending on who defines a particular landscape as such. A further challenge arises from the diversity of physical properties recorded by satellite sensors: not all images represent collections of discrete objects. Whereas high-resolution imagery may permit object-based interpretation, medium- and low-resolution data are better conceptualized as continuous fields describing spatially varying phenomena. Moreover, time series of satellite images reveal dynamic processes, allowing the detection of events that correspond to temporally bounded changes in the landscape. The talk will therefore argue that advancing ML4RS requires closer engagement with experts from GIScience, Remote Sensing, and Ecology. Such interdisciplinary collaboration is essential to develop analytical methods capable of capturing the complexity inherent in large-scale Earth Observation data.

Bio: Gilberto Câmara (Keynote 1) is a distinguished Brazilian computer scientist and geoinformatician whose work bridges spatial data science, remote sensing, and artificial intelligence. He explores how machine learning can be applied to satellite image time-series analysis and geospatial big data—contributing to open-source tools (e.g., the sits R package) for Earth observation. His research agenda encompasses land-use change, environmental modeling, and the development of semantic models (geo-fields) to enhance the interpretation of AI outputs in spatial contexts. Through his advocacy of open data policies, his leadership in national and international programs (including as former director of INPE and former Secretariat Director of GEO), and his innovative integration of AI into geospatial research, he continues to shape how we monitor and understand environmental change.