Abstract
The reliability of data produced by citizen science projects is crucial for applications that depend on precise annotations, such as training machine learning models. This reliability can be improved through careful visual design of campaigns and statistical analysis after data collection to address inconsistencies. Launched in 2019, the ForestEyes project combines citizen science, image analysis, and machine learning to detect deforestation in tropical rainforests. Continuous improvements remain necessary to ensure more reliable data. These enhancements may include using different remote sensing images (e.g., Landsat-8 and Sentinel-2), modifying the visual layout (e.g., grid mode and flip view), and adding viewing options such as zoom. This study presents four consecutive ForestEyes campaigns, each designed with targeted enhancements to improve volunteer classification quality. The first three campaigns followed a continuous improvement strategy, with each iteration addressing limitations identified previously. This process led to the use of higher resolution Sentinel-2 imagery in the third campaign, resulting in a notable accuracy increase compared to Landsat-8 data. To further investigate the gains from this improvement, a fourth campaign tested a new “transfer segmentation” approach, which applies segmentation to Sentinel-2 images and transfers the results to Landsat-8 imagery. Volunteer classification accuracy rose from 57.77% in the first campaign (Landsat-8) to 93.33% in the third campaign (Sentinel-2). These findings highlight the value of citizen science in detecting deforestation and underscore the importance of iterative campaign design and collaborative efforts to protect tropical forests.

This publication has 29 references indexed in Scilit: