Reliability-Aware Citizen Science for Environmental Machine Learning

August 2026

Citizen science (CS) campaigns based on crowdsourcing pro- vide scalable labeled data for applications such as environmental moni- toring, but volunteer-generated annotations often suffer from noise and inconsistencies. Prior work has proposed statistical filtering techniques, including outlier detection and entropy-based agreement measures, to improve annotation reliability, yet their impact on downstream machine learning performance remains unclear. To address this gap, we conduct a case study using two CS campaigns from the ForestEyes project, focused on deforestation detection in tropical rainforests. Our approach is eval- uated against standard campaign-based and PRODES-based (expert- labeled) baselines for two different satellite (Landsat-8 and Sentinel-2). The results indicate that reliability-aware filtering increases balanced ac- curacy by up to 3.0% (relative gain) over unfiltered campaign data and surpasses the PRODES-based oracle by up to 0.8%, while requiring sub- stantially fewer training samples (at least ∼ 30% fewer). These findings highlight the potential of optimized citizen science data pipelines for deforestation detection in data-scarce regions.

Authors in the community:

• 

  F

  Fabio Augusto Faria

  ist430285

VoR - Version of Record

International Conference on Pattern Recognition 2026

Lyon

computer-and-information-sciences - Computer and information sciences

eng - English