Keynote 2: Dario Augusto Borges Oliveira, “From Scarce Labels to Real-World Impact in Earth Observation”

Abstract: Recent advances in machine learning for Earth observation have led to increasingly powerful models, yet their real-world impact remains constrained by a fundamental challenge: the scarcity and imperfection of labeled data. In many sustainability-critical applications, supervision is limited, noisy, or only available at aggregated spatial and temporal scales, creating a mismatch between available data and the level of detail required for decision-making. In this talk, I present a unified methodological perspective based on representation learning and deep clustering for learning from scarce and imperfect supervision in Earth observation. The core idea is to leverage self-supervised representations and prototype-based clustering to uncover meaningful structure in large-scale geospatial data, which can then be combined with weak supervision, secondary data sources, and domain knowledge to produce actionable predictions. I will illustrate this approach through three application domains. First, in tropical forest ecosystems, I will show how deep clustering of individual tree crowns enables (pseudo) species-level discrimination under extremely limited annotations. Second, in climate science, I will discuss how similar ideas can be used to automatically identify and categorize weather and climate regimes from spatiotemporal data. Third, in agriculture, I will demonstrate how clustering-based representations can be integrated with aggregated statistics to infer crop types and production at finer spatial scales. Together, these case studies highlight how a common set of methodological principles can bridge the gap between machine learning research and real-world deployment in sustainability applications. I will conclude by discussing open challenges in scaling these approaches and in building reliable spatiotemporal AI systems that operate under pervasive supervision constraints.

Dário Oliveira received his Ph.D. from Puc-Rio in 2013, doing part of his Ph.D. research at the Leibniz University of Hannover, Germany, and the Instituto Superior Técnico, Lisbon, Portugal. From 2014 to 2015, he worked as a research fellow at the Institute of Mathematics and Statistics from the University of Sao Paulo, and in 2015 he transitioned to the industry working at General Electric Global Research Center in Rio de Janeiro and at the IBM Research lab in São Paulo, Brazil. In 2021, he returned to the academy as a Guest Professor at the Technical University of Munich, Germany. Since 20222, he has worked as a Professor at the School of Applied Mathematics - Getulio Vargas Foundation, Rio de Janeiro.