【Team】Geoinformatics Team
【Date】2026/October/14(Wednesday) 15:00-16:00(JST)
【Speaker】Talk by Mitsuyoshi Otsu, Griffith University
Title: Learning with Noisy Labels for Robust Remote Sensing Image Analysis
Abstract:
Remote sensing (RS) image analysis relies on annotated datasets, yet obtaining accurate annotations is costly and labor-intensive. This is particularly challenging for semantic segmentation, which requires pixel-level labels. High inter-class similarity and spatial redundancy in RS imagery further complicate annotation and learning, making models vulnerable to imperfect supervision.
This seminar presents two complementary approaches to learning with noisy labels for robust RS image analysis. First, we investigate the adaptation of Co-Teaching to semantic segmentation. We reformulate the approach for pixel-level learning, incorporate spatial relationships through a conditional random field (CRF)-based loss, and introduce a Parameterized Mask Production Module (MPM) to identify and filter uncertain regions. Together, these components aim to improve learning from inaccurate annotations, and the approach is validated on three distinct RS datasets.
Second, we investigate meta-learning-based label purification without relying on a trusted clean subset. In practical applications, particularly in RS, annotation errors may persist even in reference datasets, while small clean subsets may inadequately represent the target domain. To address these challenges, we develop a multi-round meta-learning framework that iteratively refines noisy labels without requiring clean reference data. To mitigate the accumulation of confirmation bias across rounds, we theoretically derive a condition for determining the appropriate number of purification rounds and evaluate its effectiveness on general computer vision benchmarks and RS datasets.
Together, these studies address a fundamental challenge in learning from imperfect supervision: model learning and label reliability are mutually dependent. A model must learn from labels that may be unreliable, yet its predictions may also inform the identification and correction of those labels. How can this interdependence be managed when no trusted clean reference is available, and how can label refinement be stopped before it reinforces its own errors? Through these questions, the seminar explores robust learning with noisy labels in RS image analysis and its broader implications for learning from imperfect supervision.
