부산대학교 도서관

The field environment is complex and variable, and multiple factors constrain the effectiveness of UAV applications, and a single flow applications may result in over- or under-use of pesticides in plots with different requirements. Therefore, it is crucial to study a decision-making model of flow rate for plant protection UAVs under multi-factor interaction. In this paper, based on a large amount of experimental data, combined with Pearson correlation analysis and random forest variable importance score ranking, screening the features obtained from the experiment increases the correlation between input and output, making the output results more reliable. The model evaluation results showed that the GA-BP neural network model has a correlation coefficient of 0.99 between the true value, predicted value, and a coefficient of determination of 0.98, which is better than the general regression model. A validation test was conducted to test the effectiveness of the model for new data. The final result yields an error value within ±20% for the GA-BP model to predict the flow rate. At the same time, the BP neural network fluctuated more for some of the predicted values, which caused a 50% error in fitting results. It proves the feasibility of the BP neural network optimized based on feature screening and genetic algorithm in plant protection UAV flow rate decision-making, which can provide a reference basis and scientific guidance for precise variable spraying operation of plant protection UAVs.