- Date: Wednesday, September 10, 2025 between 9 a.m. and 12 p.m.
- Room: D117
- Speaker: Zhenghang Yuan
- Title: Vision-Language Understanding in Remote Sensing
- Abstract: Advances in remote sensing have greatly expanded Earth observation capabilities, yet the complexity of such data remains challenging for non-expert users. Natural language offers a powerful medium to bridge this gap. In this talk, I will explore key questions in vision-language understanding for remote sensing: (1) How can we design effective learning strategies for visual question answering (RSVQA)? (2) How can RSVQA be extended with change detection to capture temporal dynamics? (3) How can we achieve more fine-grained and interpretable outputs, such as pixel-level visual explanations? (4) How can recent progress in large language models (LLMs) enable the creation of large-scale, high-quality image–text datasets?
- Date: Wednesday, September 10, 2025 between 9 a.m. and 12 p.m.
- Room: D117
- Speaker: Yuanyuan Wang
- Title: Uncertainty quantification in machine learning for Earth observation
- Abstract: Machine learning is widely applied in Earth observation to derive geoinformation including critical products like biomass estimates, informing vital decisions for humanity. However, many ML models remain as black boxes, necessitating methods to quantify prediction uncertainties. I will explain the cause of the uncertainty in machine learning and explore uncertainty quantification techniques, such as Bayesian methods and ensemble approaches, as well as demonstration of uncertainty quantification in different machine learning tasks in EO.
- Date: Thursday, December 11, 2025 at 2 p.m.
- Room: A102
- Speaker: Manon Béchaz
- Title: Monitoring a Changing World: Towards Adapting Change Detection in Remote Sensing
- Abstract: The increasing availability of high-resolution satellite and aerial imagery offers unprecedented opportunities to monitor a rapidly transforming world. Yet the types of changes that are interesting to detect – urban expansion, deforestation, flood impact, seasonal variations, etc – are highly contextual and evolve over time. Conventional change detection pipelines, which rely on extensive labeled datasets and are trained for a fixed, predefined set of change categories, are therefore fundamentally limited. They lack the capacity to adapt to new environments or new types of change without costly data labeling and retraining. This motivates the development of adaptive change detection models that can continuously improve from incoming, largely unlabeled data while remaining flexible to evolving tasks. We will explore in this presentation different methods to achieve such adaptive change detection models. We will begin with 2Player, a cooperative self-supervised framework that can transform any existing supervised change detection model into an unsupervised one, enabling adaptation of existing architectures to new data and tasks while addressing the problem of label scarcity. Building on this, we will investigate how a model’s notion of change can be enriched and expanded as new semantic categories or finer-grained distinctions become relevant. We will discuss ongoing work on incremental semantic change detection, motivated by the hierarchical structure of remote sensing imagery, highlighting pathways toward continuously improving change detection systems.