Detection of oak decline using radiative transfer modelling and machine learning from multispectral and thermal RPAS imagery

Oak trees are declining at an unprecedented rate due to the interaction of many factors, such as pests, diseases, droughts, pollution and flooding. Such abiotic- and biotic-induced stress produces anomalies in plant physiological and functional traits (PTs) that may be spectrally detected, serving to quantify trees' health status and condition. Previous studies have demonstrated that PTs' dynamic response can be tracked with hyperspectral and thermal images acquired via aerial platforms. However, the ability to detect the decline at different stages of severity among distinct oak species by using high-resolution multispectral images acquired via miniaturised cameras located aboard unpiloted airborne platforms is still unknown. This cost-effective approach offers improved operability to perform missions with greater continuity and replicability, which is critical to assess the decline progression. In this work, we evaluated the use of airborne multispectral and thermal imagery coupled with a 3-D radiative transfer modelling and machine learning approach for detecting Phytophthora-infected holm oak and cork oak trees. The field study included 2299 trees classified into disease severity classes with a gradient in levels of disease incidence located in Portugal (Ourique and Avis) and Spain (Huelva and Alcue ' scar). The classification model achieved an overall accuracy of 76 % (kappa = 0.51) in detecting decline for both species, successfully identifying up to 34 % of declining trees that were not initially detected by visual inspection and confirmed in a reevaluation six months later. When compared against airborne hyperspectral imagery, results yielded comparable accuracy, with a relative decrease of ca. 4 % in overall accuracy and an average Cohen's kappa decrease of 7 %. The results further showed that classification using only hyperspectral imagery is slightly lower but equivalent to using combined multispectral and thermal data, and those derived from these sensors independently are not adequate to classify the different severity stages. The proposed model has enabled us to effectively discern various stages of decline in cork and holm oak forests across diverse geographical areas. Our study, therefore, demonstrates that the tandem use of multispectral and thermal sensors onboard a remotely piloted aircraft platform, together with a radiative transfer modelling and machine learning approach, helps us to predict the impact of this particularly damaging disease on oak trees. This capability facilitates the detection and swift mapping of disease progression, ensuring a proactive approach to forest management.