Advancing forest structure product validation with ground, space and unmanned aerial vehicle sensors

Abstract

Forests play a crucial role in the functioning of the Earth’s climate system, through their role in the carbon, energy and water cycles. The accurate description and quantification of their physical structure is essential to understand these roles, predict their behaviour under future climate change and adapt management practices accordingly. Remote sensing in particular from space-borne platforms is attractive for large area assessment of forest structure due to its cost-effectiveness, repeatability and objectiveness. However, the remote sensing signal is by nature ambiguous and needs to be interpreted with solid understanding of the underlying radiative mechanisms and uncertainties need to be rigorously quantified with independent ground data. The remote sensing community has produced a range of biophysical products describing vegetation and forest structure as well as best practice guidelines for their validation. However, the full implementation of anticipated products, including systematic repetition of validation across multiple sites (Committee on Earth Observing Satellites (CEOS) Land Product Validation (LPV) stage 4), is still to be concluded. A major challenge in this context is the provision of long-term validation data sets, which need to be cost-effective, repeatable and fast to acquire in the field. This thesis aims to investigate new ways of validation that meet the temporal and/or spatial scales of global forest structure products from space-borne missions with hectometric resolution. The particular focus is on Leaf Area Index (LAI) and Above-Ground Biomass (AGB) as metrics of physical forest structure. For the purpose of this thesis, the Speulderbos Reference site in the Veluwe forest area (The Netherlands) was established, where ground and Unmanned Aerial Vehicle (UAV)-borne sensors were tested. In Chapter 2, the automatic, passive optical sensor PAI Autonomous System from Transmittance Sensors at 57° (PASTiS-57) was tested for its suitability to monitor forest phenology and Plant Area Index (PAI), the total one-sided area of plant material per unit ground. For this, Radiative Transfer Model (RTM) experiments with turbid media and heterogeneous scenes were employed. PASTiS-57 generally meets the CEOS LPV requirement of 20% accuracy over a wide range of biochemical and illumination conditions for turbid medium canopies. However, canopy non-randomness in discrete tree models led to strong biases. In a field experiment, PASTiS-57 compared well in terms of phenological timing with Terrestrial Laser Scanning (TLS)-based PAI time series. PASTiS-57 represents a cost-effective way to continuously monitor PAI in forests. In Chapter 3, decametric resolution Sentinel-2 and Landsat 7/8 observations were analysed with hybrid LAI retrieval algorithms, which combine RTMs with Machine Learning Regression Algorithms (MLRAs). Several combinations of RTMs, MLRAs, and modifications to the processing chain were tested in order to assess their performance to predict a ground-based LAI time series, created from combined TLS and litter trap data. Most important for the success of the processing chain was the addition of a certain level of Gaussian noise to the RTM-produced database prior to MLRA training. With this processing chain, decametric resolution optical missions can produce reference LAI products for inter-comparison with hectometric products. Alternatively, the higher resolution can help to scale up small plot-based ground validation data. In Chapter 4, a novel Unmanned Aerial Vehicle Laser Scanning (UAV-LS), the RiCOPTER with VUX-1UAV laser scanner, was used to estimate canopy height and Diameter at Breast Height (DBH). TLS was used to derive reference datasets for both variables. Canopy height was comparable between both sensors with a slight underestimation for TLS, which was expected due to occlusion of the upper canopy when seen from below and hence lower TLS canopy heights. DBH was derived for the first time from UAV-LS data and compared well with TLS derived DBH. However, a part of the UAV-LS samples could not produce a meaningful estimate of DBH based on the extracted point cloud segment due to low point density. Repeated overpasses could counteract this to some degree. In this context, UAV-LS can support fast, plot-scale assessment of these two variables. In Chapter 5, the capabilities of UAV-LS are further explored in terms of explicit 3D modelling in order to estimate tree volume, which is the first step to retrieve tree AGB. For this purpose, 3D cylinder models were fitted to the segmented single trees with the TreeQSM routine. The resulting models were compared with TLS-based models and analysed separately for five different stands with varying architectures, including deciduous and coniferous species. UAV-LS was generally very successful in modelling large, deciduous trees, while coniferous trees with low branches and foliage as well as small trees proved more difficult. If successful, UAV-LS can provide the means to produce plot-scale assessment of woody volume and subsequently AGB at a fraction of time needed for TLS surveys. This thesis investigates new ways of forest structure product validation with techniques and sensors that meet the temporal and/or spatial resolution of hectometric space-borne missions.