PhenoTruckAI: mobile laboratory for hyperspectral and molecular detection of "flavescence dorée"

The quarantine "flavescence dorée (FD)", associated with 16SrV-C and -D phytoplasmas, is threatening the wine growing areas of Germany which is so far regarded as FD-free. However, one single FD-infected grapevine plant has been detected in 2020 (Jarausch et al., 2021) which was immediately eradicated. This case as well as the new EU regulations highlighted the need for a large scale screening of FD. This is hampered by the widespread presence of "bois noir (BN)", associated with ‘Candidatus Phytoplasma solani’, which induces similar symptoms in grapevine like FD. Therefore, fast and reliable detection methods for FD monitoring in the field have to be developed. The concept of the PhenoTruckAI is based on three axes: large-scale screening of vineyards using remote sensing by drones (UASs), hyperspectral screening of leaf samples for phytoplasma infections and molecular identification of FD in a mobile laboratory. The mobile laboratory is a special vehicle with 4-wheel drive which allows autonomous laboratory work direct at the field. Drone image data will be automatically processed, and sample strategies developed. One compartment of the mobile laboratory is equipped with a dual hyperspectral camera system (VNIR+SWIR, wavelength range from 400 - 2500 nm). The spectra of leaf samples will be automatically analyzed for phytoplasma symptom presence. Ongoing research focus on the spectral discrimination of FD- and BN-infected leaves based on machine learning technologies. Rapid molecular identification of FD-infections will be achieved by LAMP assays. A case study with a first prototype of the forthcoming PhenoTruckAI was conducted in Trentino and South Tyrol in summer 2023. A total of 430 either FD- or BN-infected as well as asymptomatic leaf samples of the cultivars Chardonnay and Pinot Gris were analyzed with the hyperspectral camera system in a mobile laboratory. The same samples were extracted in the molecular compartment for identification of FD and BN by PCR. Later on, spectral data were processed and segmented leaf data were analyzed patch-wise by machine learning techniques with a leave-n-out cross validation. Phytoplasma infections were identified in VNIR spectra with 95% accuracy and in SWIR spectra with 98% accuracy compared to healthy leaves. Discrimination between FD- and BN-infected leaves was more challenging. Nevertheless, machine learning approaches achieved an accuracy of FD/BN distinction of about 80% in VNIR spectra. Further work is needed to improve the FD detection.