A Study of Synthetic and Cross-Domain Real-Data Pre-Training for Aerial Farm Monitoring

• Juliana Ferreira UFV
• Lucas Silva UFV
• Antônio Gomes UFV
• Larissa Rodrigues UFV
• Thiago Gomes UFV
• Michel Silva UFV

Resumo

The use of Unmanned Aerial Vehicles (UAVs) for monitoring farms has increased in recent years. Consequently, it is essential to enhance Artificial Intelligence capabilities to leverage the imagery captured by these devices fully. Training semantic segmentation models requires a large number of labeled images, which are nonexistent in the agricultural context. To address the bottleneck of manual labeling, we investigate the use of different pre-training strategies utilizing synthetic data from the same domain (Synthetic Dataset), e.g., UAV flights over rural areas in a virtual environment, and real data from a slightly different domain (Cross-Domain Dataset), e.g., high-resolution remote sensing images. Both quantitative and qualitative results demonstrated that pre-training using the Synthetic Dataset performed better in the final training, leading to an increase of 3.1 p.p. in IoU, 6.4 in F1-Score, and 7.5 in Recall when compared to the Cross-Domain pre-training strategy.

Palavras-chave: Training, Semantic segmentation, Transfer learning, Virtual environments, Autonomous aerial vehicles, Data models, Artificial intelligence, Monitoring, Synthetic data, Testing