DOES SUPERPIXEL QUALITY MATTER FOR RELIABLE CITIZEN SCIENCE DATA?

August 2026

Tropical forests are essential to global environmental balance; however, in recent decades, these biomes have suffered increasing losses due to deforestation, for example, in the Brazil- ian Legal Amazon (BLA). In this context, the ForestEyes Project, which combines citizen science (CS) and machine learning (ML) for monitoring deforested areas, emerges as a complementary and low-cost alternative to official monitoring systems. In this project, volunteers label image segments derived from remote sensing imagery, which can subsequently be used to train ML models. Achieving reliable data in CS campaigns requires careful planning to ensure that volunteers clearly understand the target task, thereby improving the quality of their contributions. Within this framework, this work evaluates five superpixel segmentation methods (SLIC, RSS, ERGC, ETPS, and CRS) to assess their impact on labeling agreement in deforestation detection tasks. The results demonstrate the superiority of the CRS method, which achieved a classification match improvement of over 10% by producing more regular segments with higher internal class homogeneity (HoR), thus reducing volunteers’ cognitive effort. Furthermore, removing noisy tasks led to additional improvements of approximately 15% in classification matches. These findings highlight the importance of selecting appropriate segmentation methods and mitigating response variability to enhance the effectiveness of CS campaigns for deforestation monitoring.

Authors in the community:

• 

  F

  Fabio Augusto Faria

  ist430285

VoR - Version of Record

IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2026)

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