Real wildfire data beats synthetic images in UAV YOLO detection study
Researchers tested hybrid vs real data for drone fire detection—real won
A new paper accepted at ICCAS 2026, titled "Impact of Dataset Composition on Embedded Real-Time UAV Wildfire Detection Using Compact YOLO Models," tackles a critical bottleneck in drone-based fire detection: the scarcity of diverse real-world wildfire training images. The authors—Eduardo de los Santos, Andre S. Kelbouscas, Ricardo B. Grando, and Bruna V. Guterres—systematically tested four training configurations: real non-augmented, real augmented, hybrid (real + AI-generated) non-augmented, and hybrid augmented. All experiments used compact YOLO models as a controlled validation family, targeting resource-constrained embedded deployment on UAVs.
The results were surprisingly counter to common practice. The best overall operating point came from the real non-augmented dataset, which delivered the strongest balance between recall and mean average precision. Neither hybridization with synthetic data nor image augmentation produced a better final deployment choice. The authors suggest that for embedded UAV wildfire detection, dataset realism and domain alignment are more valuable than expanding the training set with synthetic images. This has practical implications for autonomous fire monitoring systems: practitioners should invest in collecting high-quality real imagery rather than relying on data augmentation or generative data to stretch limited datasets.
- Real non-augmented data beat hybrid synthetic and augmented sets for UAV wildfire detection with compact YOLO models.
- Four configurations tested (real/hybrid × augmented/non-augmented), with real non-augmented achieving the best recall vs mAP balance.
- Neither synthetic data mixing nor augmentation improved final deployment performance, highlighting domain alignment over dataset size.
Why It Matters
UAV wildfire detection systems can prioritize real-world data collection over synthetic expansion, saving compute and improving reliability in live deployments.