Adaptive spraying decision system for plant protection unmanned aerial vehicle based on reinforcement learning

To solve the problem of lacking scientific guidance in aerial pesticide application, this study introduced an adaptive spraying decision system (ASDS) for Unmanned Aerial Vehicle (UAV) spraying to guide the operators of plant protection UAVs to set reasonable spraying parameters under complicated environment. The minimum applied volume rate, proper spraying velocity, spraying height, and initial droplet size were recommended by the ASDS. The key factor of the decision system is the decision model of reinforcement learning based on the actor-critic neural network. In specific, the field experimental data were used to train the critic and actor networks, which made the model adaptive to optimize the output of spraying parameters. Compared with the conventional spraying parameters, the spraying parameters recommended by the ASDS had a positive impact on wheat parcels. The decision results of the ASDS showed that the spraying volume rate was lower in the blocks with a small leaf area index. In addition, the spraying volume rate for the whole parcel was reduced by 14%. After UAV spraying, the uniformity of the droplet deposition in the ASDS parcel was better than that in the conventional parcel. Moreover, the penetrability of the droplets and the control efficacy for the brown wheat mite Petrobia latens (Muller) were similar in the two parcels. The ASDS can recommend the optimal spraying parameters to minimize pesticide application.

Automated summary

This study introduces an adaptive spraying decision system (ASDS) for Unmanned Aerial Vehicle (UAV) spraying to guide the operators in setting reasonable spraying parameters under complicated environment. The ASDS uses reinforcement learning based on the actor-critic neural network to optimize the output of spraying parameters. The recommended spraying parameters by ASDS have a positive impact on wheat parcels compared to conventional spraying parameters.