Hardware performance analysis for deep learning-based mistletoe detection in UAV imagery

Deep learning (DL) has become indispensable in applications like computer vision and ecological data analysis, where timely and accurate insights can significantly impact environmental conservation and agricultural productivity. Invasive species, such as mistletoe, pose a global threat to forestry and commercial agriculture by weakening host trees, reducing biodiversity, and impairing ecosystem services. This study evaluates the performance of various hardware configurations for training convolutional neural networks (CNNs) to detect mistletoe in Unmanned Aerial Vehicle (UAV) imagery. Using a ResNet-34 architecture, we compare four platforms: High-Performance Workstation (HiPWS), Mid-Range Desktop with Intel (MRD-Intel), Mid-Range Desktop with AMD (MRD-AMD), and Consumer-Grade Laptop (CGL). Metrics such as configuration, training, and validation times were recorded to assess computational efficiency across configurations, including the impact of hyperthreading (HT) and virtualization. Results indicate that disabling HT consistently reduced training times across all platforms, particularly for non-CGL setups. While HiPWS excelled in CPU-intensive tasks, CGL demonstrated superior GPU performance for short-term workloads. Importantly, cost-effective solutions like MRD-AMD proved capable of handling moderate training tasks without significant performance trade-offs. The findings offer actionable insights for optimizing DL workflows in agriculture and forestry, facilitating the development of scalable and sustainable solutions to detect and mitigate invasive species globally.