Predicting the site-specific distribution of agrochemical spray deposition in vineyards at multiple phenological stages using 2D LiDAR-based primary canopy attributes

Predicting the dose to be applied on the basis of the structural characteristics of the plant canopy is a crucial step for the optimization of the spraying process. Mobile 2D LiDAR sensor data and local measurements of deposition rates from a face-to-face sprayer were made across eight fields in two Mediterranean vineyards at four dates in 2016 and 2017. Primary canopy attributes (height, width and density) were calculated from the LiDAR sensor data and the leaf wall area (LWA) determined. Multivariate models to predict the deposition distribution, as deciles, as a function of the primary canopy attributes were constructed and calibrated using the 2017 data and validated against the 2016 data. The prediction quality and uncertainty of these multivariate statistical models at various stages of growth was evaluated by comparison with a previously proposed univariate deposition models based on LWA at the same growth stages. The results showed that multivariate models can predict the distribution of deposits from a typical face-to-face sprayer more accurately (0.76 < R2 < 0.94), and robustly (10% < nRMSEp < 24%) than LWA-based univariate prediction models over the whole growing season. This improvement was especially clear for the lowest deciles (D1 to D5) of the deposition distribution. Results also demonstrated the importance of canopy density to provide relevant and complementary information to canopy dimensions when predicting deposition deciles with the multivariate models. The improved ability of multivariate models to predict underestimated deposition (-1.5% < bias < -3.2%) when compared to univariate models makes it possible to consider a reduction in the plant protection products while guaranteeing a safety margin for winegrowers when spraying. These predictive multivariate models could enable variable-rate sprayers to modulate doses at an intra-plot scale, which would allow a potential reduction in the quantities of plant protection products to be applied.

Automated summary

This study demonstrated the use of multivariate statistical models to accurately predict deposition distribution from a typical face-to-face sprayer based on the structural characteristics of the plant canopy, especially in the early stages of the growing season. Compared to LWA-based univariate prediction models, multivariate models showed higher accuracy and robustness throughout the growing season.