Ex-ante assessment of herbicide reduction by implementing early precision weed control in spring crops

Use of herbicides needs to be properly managed to balance benefit of reducing harvest loss due to weeds against potential environmental risks. Precision weed control is expected to reduce the quantity of applied herbicide by better targeting the spraying. This can involve several components of weed control, ranging from strategic to tactic decision-making, up to equipment management. To this end we tested the coupling of a weed emergence predictive model (AlertInf) for spring row crops with mapping of early weed distribution. Our aim was to achieve both timely herbicide application and precise spatial weed distribution to ensure the highest herbicide efficacy. Airborne images produced by Unmanned Aerial Vehicles provides the best trade-off for retrieving the spatial distribution of the weed infested areas. Three series of images were acquired in early June 2019 over a maize field in the experimental farm of the University of Padova. Additional field surveys allowed to identify the presence of the target weed Sorghum halepense (L.) with an emergence percentage of 96%, according to the AlertInf model, at the date of the flight. Weed spatial distribution was assessed by the exploratory comparison of two algorithms: Artificial Neural Networks (ANN) performed within the SAGA software, and Visible Atmospherically Resistant Index (VARI) within ArcGIS Pro. Classification performances were trained and evaluated against a dataset issued from visual on-screen labelling three classes, i.e. crop, bare soil, weed. ANN provided more precise weed classification in respect to VARI, therefore VARI missed a part of infested area compared to ANN. Based on the technical specifications of different models of sprayers, several maps were created slitting the field into cells of 3m2, 2m2 and 0.25m2 and with three different thresholds for treatment decision depending on each cell infestation (i.e., >1%, >5% and >10% of pixel labelled as weed). Prescription maps showed that sprayed area potentially reduced from 42% to 87% using the ANN classification and from 65% to 93% at 1% using the VARI classification for pixel with 10% of cell infestation. Altogether, the results suggest that site-specific mapping informed on emergence models can enable a substantial herbicide reduction through both timely and site-specific weed control with ordinary section control sprayers. Further research is required to evaluate the contribution of a weed emergence model, such as AlertInf, to herbicide usage reduction, performing additional tests in different fields and including more weed species.