Investigating the potential for detecting Oak Decline using Unmanned Aerial Vehicle (UAV) Remote Sensing

Abstract

This PhD project develops methods for the assessment of forest condition utilising modern remote sensing technologies, in particular optical imagery from unmanned aerial systems and with Structure from Motion photogrammetry. The research focuses on health threats to the UK's native oak trees, specifically, Chronic Oak Decline (COD) and Acute Oak Decline (AOD). The data requirements and methods to identify these complex diseases are investigated using RGB and multispectral imagery with very high spatial resolution, as well as crown textural information. These image data are produced photogrammetrically from multitemporal unmanned aerial vehicle (UAV) flights, collected during different seasons to assess the influence of phenology on the ability to detect oak decline. Particular attention is given to the identification of declined oak health within the context of semi-natural forests and heterogenous stands. Semi-natural forest environments pose challenges regarding naturally occurring variability. The studies investigate the potential and practical implications of UAV remote sensing approaches for detection of oak decline under these conditions. COD is studied at Speculation Cannop, a section in the Forest of Dean, dominated by 200-year-old oaks, where decline symptoms have been present for the last decade. Monks Wood, a semi-natural woodland in Cambridgeshire, is the study site for AOD, where trees exhibit active decline symptoms. Field surveys at these sites are designed and carried out to produce highly-accurate differential GNSS positional information of symptomatic and control oak trees. This allows the UAV data to be related to COD or AOD symptoms and the validation of model predictions. Random Forest modelling is used to determine the explanatory value of remote sensing-derived metrics to distinguish trees affected by COD or AOD from control trees. Spectral and textural variables are extracted from the remote sensing data using an object-based approach, adopting circular plots around crown centres at individual tree level. Furthermore, acquired UAV imagery is applied to generate a species distribution map, improving on the number of detectable species and spatial resolution from a previous classification using multispectral data from a piloted aircraft. In the production of the map, parameters relevant for classification accuracy, and identification of oak in particular, are assessed. The effect of plot size, sample size and data combinations are studied. With optimised parameters for species classification, the updated species map is subsequently employed to perform a wall-to-wall prediction of individual oak tree condition, evaluating the potential of a full inventory detection of declined health. UAV-acquired data showed potential for discrimination of control trees and declined trees, in the case of COD and AOD. The greatest potential for detecting declined oak condition was demonstrated with narrowband multispectral imagery. Broadband RGB imagery was determined to be unsuitable for a robust distinction...