Mapping understorey tree invasion in a drought-affected forest using multi-view UAV imagery and deep learning

Abstract

Advances in deep learning (DL) and structure from motion (SfM) photogrammetry combined with off-the-shelf unoccupied aerial vehicles (UAVs) and high-resolution cameras enable unprecedented plant species mapping accuracy. While these tools have been mainly applied to flat terrain or upper forest canopies, the forest understorey remains largely unexplored. Here, we present a method combining DL with multi-view UAV imagery to map the invasive tree-of-heaven (Ailanthus altissima) in the understorey of a drought-affected Central European forest. The raw UAV photographs were segmented with convolutional neural networks (CNNs). Resulting predictions were projected onto georeferenced point clouds using SfM. This novel approach revealed that more than 40% of the invasion was hidden beneath the canopy and would have been missed by conventional orthomosaic-based methods. To assess CNN generalization abilities, we altered training and prediction domains: lower-processing-level aerial images vs. higher-processing-level orthomosaic, both originating from a small training extent (420 m2 or 0.25% of the study area). For intra-domain predictions, aerial-trained models (F1 = 0.880) outperformed ortho-trained models (F1 = 0.805). When applied cross-domain, aerial models retained superior performance (F1 = 0.843) over ortho-trained ones (F1 = 0.750). Expanding the training extent eightfold raised the accuracy of the ortho-trained models to F1 = 0.836 within-domain and F1 = 0.846 cross-domain. In summary, integrating photogrammetry with DL is a promising route for utilizing overlapping aerial imagery efficiently and leveraging information that conventional orthomosaic-based analyses discard. The resulting richer 3D information and improved transferability may support decision-making and retrofitting this novel technique to upcoming and existing datasets opens new avenues in vegetation studies and beyond.