Case Comparison of Label Efficiency with Active Learning Strategies in Satellite Image Archives

Abstract

Satellite image archives are expanding at an unprecedented rate, providing invaluable data for environmental monitoring, urban planning, and disaster management. However, the supervised machine learning models typically employed to analyze these vast datasets require massive amounts of annotated data. Generating these annotations is highly labor-intensive, time-consuming, and expensive, thus necessitating methods that maximize label efficiency. This paper assesses the label efficiency of various active learning strategies applied to satellite image archives through a comprehensive case comparison methodology. By evaluating uncertainty sampling, query-by-committee, and representative sampling strategies across diverse remote sensing datasets, the research elucidates the conditions under which specific active learning frameworks yield optimal performance. The comparative evidence demonstrates that hybrid active learning strategies significantly reduce the annotation burden while maintaining high classification accuracy, particularly in highly imbalanced or complex spatial environments. The findings present a rigorous methodological framework for optimizing annotation budgets in Earth observation studies. Ultimately, the paper provides theoretical and practical insights for deploying cost-effective machine learning pipelines in remote sensing applications, fostering improved scalability in the analysis of large-scale satellite image archives.