Open-Set Spatiotemporal Segmentation for Phenological Novelty Detection

• Cristiano N. Rodrigues UFMG
• Frederico G. Guimarães UFMG
• Jefersson A. Dos Santos University of Sheffield

Resumo

Vegetation segmentation in natural ecosystems is challenged by class incompleteness, phenological ambiguity, and spatiotemporal variability. Annotated datasets often represent only a subset of temporal variations, leaving many phenological stages of known species unlabeled. Traditional segmentation models trained under closed-set assumptions struggle to generalize under such conditions. We present GeMOSS (Generative Models for Open-Set Spatiotemporal Segmentation), a novel extension of the GeMOS framework to pixel-level segmentation under open-set conditions. GeMOSS integrates a multibranch temporal convolutional network with shallow generative modeling to detect previously unseen temporal patterns within known vegetation classes. Rather than identifying taxonomically novel species, our approach targets phenological novelty within established classes, essential for ecological monitoring using partially labeled time series. Experiments on a real-world dataset from the Brazilian Cerrado demonstrate that GeMOSS achieves 96. 45% known class accuracy (compared to 94.36% for closed-set baselines), while filtering low-likelihood predictions potentially associated with atypical temporal behaviors, supporting its potential to flag underrepresented phenological instances in dynamic ecosystems.

Palavras-chave: Image segmentation, Accuracy, Biological system modeling, Ecosystems, Time series analysis, Vegetation mapping, Vegetation, Spatiotemporal phenomena, Monitoring, Anomaly detection