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 Alcuéscar). 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.

由于病虫害、干旱、污染和洪水等多种因素的相互作用，橡树正在以前所未有的速度减少。这种非生物和生物引起的胁迫会导致植物生理和功能性状（PT）异常，这些异常可以通过光谱检测，从而量化树木的健康状况和状况。先前的研究表明，PT 的动态响应可以通过空中平台获取的高光谱和热图像进行跟踪。然而，通过使用位于无人驾驶机载平台上的小型摄像机获取的高分辨率多光谱图像来检测不同橡树品种在不同严重程度阶段的衰退的能力仍然未知。这种具有成本效益的方法提高了可操作性，以更大的连续性和可复制性执行任务，这对于评估衰退进展至关重要。在这项工作中，我们评估了使用机载多光谱和热图像以及 3-D 辐射传输模型和机器学习方法来检测受疫霉感染的圣栎和栓皮栎树。实地研究包括位于葡萄牙（Ourique 和 Avis）和西班牙（Huelva 和 Alcuéscar）的 2299 棵树，这些树被分类为疾病严重程度等级，疾病发病率水平存在梯度。该分类模型在检测两个物种的衰退方面达到了 76%（ kappa = 0.51）的总体准确度 ，成功识别了高达 34% 的衰退树木，这些树木最初未通过目视检查检测到，并在六个月后的重新评估中得到确认。与机载高光谱图像相比，结果产生了相当的精度，但相对降低了约。总体准确度降低 4%，平均 Cohen kappa 降低 7%。结果进一步表明，仅使用高光谱图像的分类稍低，但相当于使用组合的多光谱和热数据，并且独立地从这些传感器获得的分类不足以对不同的严重性阶段进行分类。所提出的模型使我们能够有效地辨别不同地理区域的软木和圣栎森林衰退的各个阶段。因此，我们的研究表明，在遥控飞机平台上串联使用多光谱和热传感器，再加上辐射传输模型和机器学习方法，有助于我们预测这种特别具有破坏性的疾病对橡树的影响。这种能力有助于检测和快速绘制疾病进展图，确保采取积极主动的森林管理方法。