Effective post-planting seedling detection is essential for the success of forestry operations, yet traditional ground-based surveys often struggle with efficiency and accuracy in large-scale environments. This study develops a multimodal deep learning approach to enhance seedling detection in New Zealand’s plantation forests by integrating high-resolution RGB and multispectral UAV imagery. Using Faster R-CNN models, this research evaluates the performance of RGB-only, multispectral-only, and combined data models for detecting Pinus radiata and Pseudotsuga menziesii seedlings across diverse site conditions. The combined model demonstrated superior accuracy, achieving F1 scores of up to 95.8% for P. menziesii, highlighting the value of multispectral data in improving precision and reducing false positives in areas with dense weed cover.

Results reveal that while environmental factors such as vegetation density and seedling visibility impact model performance, multispectral data can mitigate some of these limitations, particularly for sites with minimal vegetation and flat terrain. This study emphasises the role of optimal flight parameters and suggests early-season imagery to further enhance detection accuracy. These findings contribute to advancing UAV-based automated seedling detection, paving the way for efficient and scalable survival surveys, weed control, and wilding conifer management. Future research should focus on refining model robustness, optimising detection algorithms, and establishing operational decision thresholds for forest management applications.