Ultrahigh-resolution boreal forest canopy mapping: Combining UAV imagery and photogrammetric point clouds in a deep-learning-based approach

• DOI: 10.1016/j.jag.2022.102686
• Publication Date: 2022-02-05T04:41:26Z
• Authors: Linyuan Li; Xihan Mu; Francesco Chianucci; Jianbo Qi; Jingyi Jiang; Jiaxin Zhou; Ling Chen; Huaguo Huang; Guangjian Yan; Shouyang Liu
• Keywords: Understory; Tree canopy

Accurate wall-to-wall estimation of forest crown cover is critical for a wide range ecological studies. Notwithstanding the increasing use UAVs in canopy mapping, ultrahigh-resolution UAV imagery requires an appropriate procedure to separate contribution understorey from overstorey vegetation, which complicated by spectral similarity between two components and illumination environment. In this study, we investigated integration deep learning combined data photogrammetric point clouds boreal mapping. The enables automatic creation training sets tree (overstorey) background (understorey) via combination images their associated expands applicability models with self-supervision. Based on different overlap levels 12 conifer plots that are categorized into “I”, “II” “III” complexity according environment, compared self-supervised learning-predicted maps original manual delineation found average intersection union (IoU) larger than 0.9 “complexity I” II” 0.75 III” plots. proposed method was then three classical image segmentation methods (i.e., maximum likelihood, Kmeans, Otsu) plot-level estimation, showing outperformance extraction against other methods. also validated pointwise estimates using LiDAR situ digital photography (DCP) benchmarking results showed model-predicted line (RMSE 0.06) deviate DCP 0.18). We subsequently new commonly used structure-from-motion (SfM) at varying forward lateral overlaps over all rugged terrain region, yielding method-predicted relatively insensitive (largest bias less 0.15), whereas SfM-estimated seriously affected decreased decreasing overlap. addition, mapping verified merits method, no need detailed model (DTM). recommended be various overlaps, illuminations, terrains due its robustness high accuracy. This study offers opportunities promote applications (e.g., leaf area index estimation) sustainable management deforestation).