Predictive model for spray drift behavior from plant protection UAVs

Droplet drift can be generated during pesticide application by plant-protection Unmanned Aerial Vehicles (UAVs) in agricultural production. Excessive spray drift can reduce pesticide utilization efficiency, leading to a series of environmental and ecological risks, such as environmental contamination, off-target deposition, and phytotoxic effects on non-target organisms. It is often required to accurately characterize and then predict the droplet drift behavior under diverse operations, particularly for safe and sustainable spraying. International standards, such as ISO 22866:2005, have also been established to quantify the spray drift assessment in the field measurement protocols. However, conventional drift prediction models of manned aircraft or ground sprayers cannot be extended into the UAV spraying scenarios, due to the aerodynamic characteristics of multi-rotor UAVs—such as strong downwash airflow, low-altitude operation, and flexible flight parameters. In this study, the data-driven prediction models were developed to explicitly incorporate UAV operational features and environmental variability. The spray drift was measured on four plant-protection UAVs under typical operation. 150 datasets were then collected after measurement. Spearman correlation analysis was used to effectively capture the complex coupling relationships between multiple influencing factors and droplet drift rate. Feature importance was ranked to identify key variables using Random Forest. A prediction model was constructed for the droplet drift in the plant-protection UAVs using a Backpropagation (BP) neural network. SHAP (SHapley Shapley Additive exPlanations) was introduced to interpret the outputs and then quantify the contribution rates of each influencing factor. Furthermore, a specific UAVDP (UAV Droplet Drift Prediction Platform) was developed to facilitate practical application. A systematic comparison was made using the physical drift prediction tool (AGDISPpro). The results demonstrate that the BP neural network model was achieved in the high prediction accuracy and generalization, outperforming several conventional machine learning models. The accuracy was also improved over AGDISPpro. SHAP interpretability analysis revealed that the wind speed and spray flow rate were dominant positive contributors to spray drift. While operational height and nozzle number exhibited the notable suppression. The drift risk level of plant-protection UAV operations was quantitatively assessed after prediction. The required downwind buffer distancebuffer distance downwind was generally less than 20 m under compliant wind speed conditions (<5 m/s). The drift risk can be found between manned agricultural aircraft and large ground boom sprayers. Overall, this finding can provide a data-driven modeling framework, interpretable insights into drift mechanisms, and a practical prediction platform. Robust technical support can also offer the precision pesticide applications, drift risk control, and the decision-making on safe UAV spraying.