Evaluation of a CNN model to map vegetation classification in a subalpine coniferous forest using UAV imagery

Abstract

Unmanned aerial vehicle (UAV) remote sensing based on deep learning has increasingly been applied for forest vegetation classification. However, existing studies have focused mainly on simple woodlands, and accurately mapping the vegetation distribution in complex natural forests remains challenging. To address this, we conducted a study in a natural alpine forest in Southwest China, leveraging high-resolution UAV imagery and deep learning for vegetation classification. We systematically assessed the effects of patch size, spatial resolution, and rotation angle on the model performance, considering their interactions. Our results demonstrate that UAVs combined with deep learning techniques achieve high classification accuracy in natural forests, with a mean F1-score of 0.91. Patch size has a significant influence on accuracy, although its impact diminishes as the spatial resolution decreases. As the patch size increased from 128 × 128 to 256 × 256, the model F1-score improved by 18 % at a 5 cm resolution, whereas it improved by only 3 % at a 10 cm resolution. A higher spatial resolution does not necessarily enhance model accuracy, and the effect of patch size also needs to be considered. The Rotation angle, as a data augmentation strategy, is crucial when training data are limited and can significantly increase model performance. These findings highlight the potential of combining deep learning and UAV remote sensing for natural forests. This approach facilitates more reliable access to forest information in forest areas where access is difficult. Overall, this study provides an efficient and cost-effective method for monitoring and protecting natural forests, serving as a reference for selecting appropriate parameters in UAV-based deep learning remote sensing.