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

• Published in: International journal of applied earth observation and geoinformation Vol. 107; p. 102686
• Main Authors: Li, Linyuan, Mu, Xihan, Chianucci, Francesco, Qi, Jianbo, Jiang, Jingyi, Zhou, Jiaxin, Chen, Ling, Huang, Huaguo, Yan, Guangjian, Liu, Shouyang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2022; Elsevier
• Subjects: boreal forests; Canopy mapping; coniferous forests; Crown cover; deforestation; forest canopy; image analysis; Image overlap; landscapes; leaf area index; lidar; lighting; overstory; photogrammetry; Self-supervised deep learning; SfM point cloud; spatial data; statistical analysis; tree crown; UAV imagery; understory; UAV imagery; CNN; nDSM; SNFP; AGL; LiDAR; DL; CHM; Self-supervised deep learning; MVS; UAV; Image overlap; FOV; Canopy mapping; IoU; DCP; OA; SfM point cloud; Crown cover; GSD; SfM; DTM; DSM
• DOI: 10.1016/j.jag.2022.102686
• ISSN: 1569-8432

•Self-supervised deep learning for forest canopy mapping using UAVs is proposed.•2D images and 3D SfM point clouds are combined for automated training set generation.•The method is evaluated by annotated images and classical segmentation algorithms.•The method is compared with UAV LiDAR and DCP benchmarking methods.•The method shows better consistency at varying image overlaps compared with UAV SfM. Accurate wall-to-wall estimation of forest crown cover is critical for a wide range of ecological studies. Notwithstanding the increasing use of UAVs in forest canopy mapping, the ultrahigh-resolution UAV imagery requires an appropriate procedure to separate the contribution of understorey from overstorey vegetation, which is complicated by the spectral similarity between the two forest components and the illumination environment. In this study, we investigated the integration of deep learning and the combined data of imagery and photogrammetric point clouds for boreal forest canopy mapping. The procedure enables the automatic creation of training sets of tree crown (overstorey) and background (understorey) data via the combination of UAV images and their associated photogrammetric point clouds and expands the applicability of deep learning models with self-supervision. Based on the UAV images with different overlap levels of 12 conifer forest plots that are categorized into “I”, “II” and “III” complexity levels according to illumination environment, we compared the self-supervised deep learning-predicted canopy maps from original images with manual delineation data and found an average intersection of union (IoU) larger than 0.9 for “complexity I” and “complexity II” plots and larger than 0.75 for “complexity III” plots. The proposed method was then compared with three classical image segmentation methods (i.e., maximum likelihood, Kmeans, and Otsu) in the plot-level crown cover estimation, showing outperformance in overstorey canopy extraction against other methods. The proposed method was also validated against wall-to-wall and pointwise crown cover estimates using UAV LiDAR and in situ digital cover photography (DCP) benchmarking methods. The results showed that the model-predicted crown cover was in line with the UAV LiDAR method (RMSE of 0.06) and deviate from the DCP method (RMSE of 0.18). We subsequently compared the new method and the commonly used UAV structure-from-motion (SfM) method at varying forward and lateral overlaps over all plots and a rugged terrain region, yielding results showing that the method-predicted crown cover was relatively insensitive to varying overlap (largest bias of less than 0.15), whereas the UAV SfM-estimated crown cover was seriously affected by overlap and decreased with decreasing overlap. In addition, canopy mapping over rugged terrain verified the merits of the new method, with no need for a detailed digital terrain model (DTM). The new method is recommended to be used in various image overlaps, illuminations, and terrains due to its robustness and high accuracy. This study offers opportunities to promote forest ecological applications (e.g., leaf area index estimation) and sustainable management (e.g., deforestation).