Probabilistic Prediction of Agricultural Spray Drift and Deposition Using a LES Ensemble

A longstanding challenge with agricultural herbicide and pesticide application is anticipating and controlling the amount and direction of chemical drift beyond the target area. Such drift is typically sensitive to variations in meteorological parameters (e.g., wind speed, wind direction, surface layer stability, etc.) on space-time scales of minutes and tens of meters. At these scales, it is nearly impossible to predict these parameters deterministically with useful levels of skill. A potentially viable approach is to construct probabilistic drift predictions based upon an ensemble of Large Eddy Simulation (LES)-scale simulations or perhaps more efficiently a machine learning algorithm that is trained to emulate the key information provided by LES-scale ensembles.

Experiments have been conducted to test this concept on a spray application field at a site in the central United States. As part of this field experiment, extensive measurements were made of the application parameters and off-target drift and deposition. An LES-scale (30-m) ensemble was constructed on nested grids with the Weather Research and Forecast (WRF) system using input from a set of standard operational Numerical Weather Prediction (NWP) forecasts. The ensemble was created using variations in the source of the operational NWP data, initiation time for LES-scale simulations, structure of the nested grid configuration, and land surface and surface layer physics.

The meteorological output from each LES ensemble member was read into a community agricultural dispersion model (AgDisp). AgDisp predicts drift and deposition based on various inputs including meteorological conditions, drop size distribution, injection height, nozzle flow speed, local turbulence induced by the spray process, and several other user-selectable factors. The ensemble of drift and deposition predictions was then used to construct a probabilistic prediction of off-target drift and deposition.

The probabilistic predictions for the field test case were compared to drift and deposition measurements as well as deterministic predictions from single NWP forecasts and on-site meteorological observations. An analysis was also performed to estimate probabilistic forecast sensitivity to subsets of the LES ensemble. The conference presentation will summarize these results and conclude with recommendations on how to improve the state-of-the-science for agricultural spray drift and deposition.