Seeing The Ground Below Dense Canopies: A Structure from Motion Workflow for Generating High Accuracy Canopy Heights in a Tropical Forest

Details

Supervisors

Van Oost, Kristof

Faculty

Faculté des sciences

Degree label

Master [120] en sciences géographiques, orientation générale, à finalité approfondie

Abstract

Forests play an active part in the carbon cycle, by storing large amounts of it as biomass. In recent years, deforestation and forest degradation have contributed to CO2 emissions, causing a global climate change. The Democratic Republic of Congo (DRC) is the second most forested country in the world, and is undergoing land use changes through deforestation, which cause CO2 emissions. Efforts to mitigate these emissions, and restore forest carbon stocks, such as REDD+ require monitoring of high spatial and temporal resolution. One way that this can be done is through measuring canopy heights in forests. However, current solutions such as satellite-based have too high an error and too coarse a resolution, while airborne LiDAR measurements fit precision requirements, but are too expensive for regular monitoring. Recent advances in uncrewed aerial vehicles (UAVs) and structure from motion (SfM) (a photogrammetry technique) technologies have enabled fast and cheap forest inventories for monitoring forest stands. These UAV-SfM inventories underperform in certain contexts, such as dense canopies (including the forests in the DRC), leading to underestimations of digital terrain models (DTMs) and canopy heights models (CHMs). Zhang et al. (2021) developed a high potential workflow with UAV-SfM point cloud for estimating tree heights in the dense forest of the DRC, by attaining estimations comparable to LiDAR estimations through machine learning and morphological filters to identify ground points and co-kriging to estimate a digital terrain model. In this study, we explored possibilities of this new workflow in 3 objectives. We first investigated the possibility of improving the ground detection filters by exploring the subtle structure of the point cloud. Secondly, we adapted the UAV-SfM data collection by adding oblique pictures to the photogrammetric reconstruction, to improve ground detection. Finally, we evaluated the potential for monitoring the state of forest structure through UAV-SfM by comparing our workflow to a 7-year-old LiDAR canipy height dataset. Thanks to the improved classification of ground points, the UAV-SfM DTM had a RMSE of 1.6 m and NSE of 0.84, compared to a LiDAR reference DTM. An external validation with GNSS ground sample measurements showed the error was high (RMSE = 4.8 m ; NSE = 0.68) and caused by the interpolation of ground points performing poorly in varied terrain. We also showed that oblique pictures improved ground detection, and lowered the RMSE from 4.45 m to 2.74 m on average, and, using two different angles for oblique pictures (70° and 40°), that low oblique (i.e. 70°) aided most in ground reconstruction. Finally, in evaluating the state of forest structure through canopy heights, we found that smaller trees overestimated tree heights and had a relatively high error (RMSE = 3.18 m) compared to the 7-year reference, and bigger trees (more than 40 m) had a more stable height (and lower error of RMSE = 2.12 m). Moreover, we found a RMSE of 2.04 m for the tallest tree in a plot, and deduced that the forest had remained in a similar structural state compared to 7 years ago due to absence of over/underestimation. In conclusion, thanks to an optimization of ground detection algorithms and data collection scheme (i.e., adding oblique pictures), we improved a workflow for estimating tree heights, and demonstrated the potential for monitoring forest structure with a cheap and easy-to-use UAV-SfM workflow in the challenging context of tropical forests.