Invasive species detection in heritage sites using UAV-based hyperspectral imagery and patch-based deep learning

Detecting invasive species is essential for preserving native vegetation and maintaining the identity of cultural heritage sites. Conventional on-the-ground surveys, however, face significant limitations in terms of time, cost, and accessibility. Accordingly, this study evaluated the detection effectiveness of the invasive species Robinia pseudoacacia L. at two cultural heritage sites using unmanned aerial vehicle (UAV)-based hyperspectral imagery (HSI) and patch-based deep learning architectures. Particular attention was given to methodological implications, especially regarding generalizability and practical applicability. Fourteen datasets acquired at different time points from the two sites were employed, with a large portion dedicated to evaluation to enable rigorous validation. We also examined the influence of patch size as a key parameter in model design. Furthermore, by using ground-truth data spanning a wide range of scales, we assessed detection performance across various morphologies, including small tree crowns that are difficult to detect with medium-resolution imagery. The results showed that the Convolutional Neural Network (CNN) model combined with PCA for dimensionality reduction achieved robust and stable performance across acquisition dates, despite class imbalance and variable vegetation and meteorological conditions. The model effectively detected target species of various sizes, with undetected targets largely confined to small crowns, representing less than 0.5% of the total ground-truth area. Performance improved with patch size up to 15 pixels, after which gains plateaued. Overall, this study highlights the potential of UAV-based HSI integrated with patch-based DL models as a practical framework for invasive species monitoring and management in cultural heritage contexts.