Assessing uncertainties related to vine damage by Eotetranychus carpini via machine learning technique
Abstract: The control of the yellow spider mite Eotetranychus carpini (Oudemans) remains an important challenge on grapevine (Vitis vinifera L.) in Southern Europe, especially in Tuscan viticulture. The frequency of damage due to E. carpini has escalated by the lack and/or unsuitable timing in control interventions. This is largely attributable to the difficulty in correlating the onset of damage with E. carpini populations. Here, we apply a non-parametric machine learning technique, i.e., Random Forest (RF), to quantify the variability of leaf area damage as affected by the abundance and structure of populations of tetranychids and phytoseiids, and by vine cultivar and phenological stage. The highest predictive power was achieved by the RF built considering the damage class 0 (healthy leaf) and 1 (< 20% leaf area affected), with 86.73% of matched observations in the out-of-bag dataset. The performances of the RFs decreased when considering the other damage classes. The number of immature tetranychids was the top-ranked variable explaining leaf damage variability, followed by vine phenology, cultivar and phytoseiids. These results confirm the high correlation between E. carpini abundance and early leaf damage, suggesting the need of its timely detection on vine leaves, as well as the relevance of both cultivar and natural antagonists in modulating the impact of the yellow spider mite.