Using Citizen Science Data as Pre-Training for Semantic Segmentation of High-Resolution UAV Images for Natural Forests Post-Disturbance Assessment DOI Open Access

Kamyar Nasiri,

William Guimont-Martin, Damien LaRocque

et al.

Forests, Journal Year: 2025, Volume and Issue: 16(4), P. 616 - 616

Published: March 31, 2025

The ability to monitor forest areas after disturbances is key ensure their regrowth. Problematic situations that are detected can then be addressed with targeted regeneration efforts. However, achieving this automated photo interpretation problematic, as training such systems requires large amounts of labeled data. To effect, we leverage citizen science data (iNaturalist) alleviate issue. More precisely, seek generate pre-training from a classifier trained on selected exemplars. This accomplished by using moving-window approach carefully gathered low-altitude images an Unmanned Aerial Vehicle (UAV), WilDReF-Q (Wild Drone Regrowth Forest—Quebec) dataset, high-quality pseudo-labels. accurate pseudo-labels, the predictions our for each window integrated majority voting approach. Our results indicate semantic segmentation network over 140,000 auto-labeled yields F1 score 43.74% 24 different classes, separate ground truth dataset. In comparison, only 32.45%, while fine-tuning pre-trained marginal improvements (46.76%). Importantly, demonstrate able benefit more unlabeled images, opening door learning at scale. We also optimized hyperparameters pseudo-labeling, including number assigned pixel in process. Overall, demonstrates auto-labeling greatly reduce development cost plant identification regions, based UAV imagery.

Language: Английский

Using Citizen Science Data as Pre-Training for Semantic Segmentation of High-Resolution UAV Images for Natural Forests Post-Disturbance Assessment DOI Open Access

Kamyar Nasiri,

William Guimont-Martin, Damien LaRocque

et al.

Forests, Journal Year: 2025, Volume and Issue: 16(4), P. 616 - 616

Published: March 31, 2025

The ability to monitor forest areas after disturbances is key ensure their regrowth. Problematic situations that are detected can then be addressed with targeted regeneration efforts. However, achieving this automated photo interpretation problematic, as training such systems requires large amounts of labeled data. To effect, we leverage citizen science data (iNaturalist) alleviate issue. More precisely, seek generate pre-training from a classifier trained on selected exemplars. This accomplished by using moving-window approach carefully gathered low-altitude images an Unmanned Aerial Vehicle (UAV), WilDReF-Q (Wild Drone Regrowth Forest—Quebec) dataset, high-quality pseudo-labels. accurate pseudo-labels, the predictions our for each window integrated majority voting approach. Our results indicate semantic segmentation network over 140,000 auto-labeled yields F1 score 43.74% 24 different classes, separate ground truth dataset. In comparison, only 32.45%, while fine-tuning pre-trained marginal improvements (46.76%). Importantly, demonstrate able benefit more unlabeled images, opening door learning at scale. We also optimized hyperparameters pseudo-labeling, including number assigned pixel in process. Overall, demonstrates auto-labeling greatly reduce development cost plant identification regions, based UAV imagery.

Language: Английский

Citations

0