In-Situ Aerial Mapping of New Zealand Myrtaceae Affected by Myrtle Rust (Austropuccinia psidii) Using Deep Learning

Abstract

Invasive fungal pathogens pose a significant threat to forest ecosystems worldwide and have far-reaching consequences for tree species. The rust fungus (Austropuccinia psidii) causes the disease commonly known as myrtle rust and threatens susceptible Myrtaceae populations on several continents. This includes Syzygium maire, a rare taonga (treasure) species endemic to New Zealand that is significant in Māori culture and ecologically important, but is now threatened with extinction. Accurate spatial mapping of threatened populations is essential for targeted management and conservation efforts, but traditional ground-based survey methods are logistically challenging and time-consuming.

The practical application of unmanned aerial vehicles (UAVs) in combination with deep learning was evaluated to detect S. maire in dense, species-rich native forests. High-resolution RGB (1.5 cm) and multispectral (2.5 cm) imagery were captured from four urban forest reserves on New Zealand’s North Island using consumer-grade imaging sensors. A fully convolutional neural network for semantic segmentation (U-Net) was trained to classify S. maire from background vegetation. Furthermore, dataset composition and hyperparameter configurations were systematically tested including loss functions, learning rates, and different spectral band combinations. Point cloud segmentation approaches using a UAV-mounted LiDAR system were also qualitatively evaluated to assess the potential for three-dimensional tree instance detection.

Site-specific models showed moderate to good detection performance (F1 scores: 0.46–0.81), with RGB images performing comparably or marginally better than multispectral images. Dice loss outperformed pixel-wise approaches in handling severe class imbalances, and an aggressive learning rate of 0.02 with adaptive scheduling led to significantly better performance. However, generalising models across multiple sites proved more difficult due to site-specific differences (best F1=0.51). LiDAR-based instance segmentation algorithms developed for managed forests have potential for the dense, structurally complex context of natural forests, but are insufficient without further development.

The results show that deep learning can successfully identify S. maire under optimal, site-specific conditions. At the same time, critical limitations for operational real-world use were identified. The limited availability of training samples, the severe class imbalance of 3.4±0.7% (mean±SE) target class and the insufficient radiometric calibration capabilities of low-cost multispectral sensors fundamentally limit the approach. These findings highlight the need for standardised frameworks governing multispectral data capture and radiometric calibration.

The findings expose significant challenges in translating remote sensing methods from simplified scenarios to operational conservation monitoring in structurally complex forests. The methodological insights regarding hyperparameter optimisation, spectral band selection, and calibration challenges can be transferred to analogous applications for species detection. As invasive pathogens increasingly threaten forest biodiversity worldwide, the development of robust detection methods for (rare) species in complex environments remains essential to support proactive conservation and biosecurity measures.