Toward an Operational System for Automatically Detecting Xylella fastidiosa in Olive Groves Based on Hyperspectral and Thermal Remote Sensing Data

Xylella fastidiosa (Xf) is a pathogenic bacterium which causes severe damage to plants and has been detected in various countries, including Italy, France, Portugal, Spain, Lebanon, Iran, and Israel. In Europe, the first outbreak was observed in olive plants in Apulia, Italy, in 2013. The ease of its transmission, coupled with its ability to remain latent within plants for extended periods, has facilitated its rapid expansion, causing severe damage to the regional olive industry. The early detection of Xf infections is therefore crucial for the containment of its spread and, thus, to minimize crop yield losses. Recent studies described in the literature have assessed the potential of remote sensing for monitoring Xf through applicable machine learning models. In particular, high-resolution hyperspectral and thermal remote sensing imageries acquired by airborne platforms have demonstrated an ability to detect the early symptoms of Xf infection in olive trees. However, further analyses are needed to address technical challenges and validate their effectiveness in vast areas. In this paper, we propose to answer some of these crucial questions, which are also relevant to the future task of setting up an operational system to detect Xf on a large scale. First, we assess whether the size of a data set, composed of a limited number of labelled examples, is sufficient to train accurate classifiers. Then, we evaluate whether a classifier that is trained on data from a specific area can detect infected trees in other places, which are potentially different in terms of cultivars and overall agricultural management. The obtained results demonstrate that with as few as 200 labelled data points (even unbalanced between the two classes of interest of “infected” and “not infected”), it is possible to train classifiers to support the detection of Xf, also across a wide area, obtaining overall classification accuracies greater than 74%.