Development and comparative analysis of ANN and SVR-based models with conventional regression models for predicting...

Area of Science:

• Agricultural Engineering
• Environmental Science
• Data Science

Background:

• Spray drift monitoring is costly and labor-intensive, necessitating practical alternatives.
• Existing mechanistic drift prediction models are often complex, requiring extensive input data.
• The need for accessible and accurate pesticide drift prediction tools is critical.

Purpose of the Study:

• To develop and compare advanced machine learning models (ANN, SVR) for pesticide drift prediction.
• To evaluate the performance of machine learning models against conventional regression models (MLR, GLM, GNLS).
• To assess the potential of machine learning as a viable tool for spray drift modeling.

Main Methods:

• Developed artificial neural network (ANN) and support vector regression (SVR) models.
• Compared ANN and SVR with multiple linear regression (MLR), generalized linear model (GLM), and generalized nonlinear least squares (GNLS).
• Utilized fivefold cross-validation and external validation, assessing performance with R², RMSE, MAE, and MAB.

Main Results:

• ANN and SVR models achieved high predictive accuracy (R²=0.98 and 0.97, respectively).
• GLM and GNLS showed strong performance among conventional models (R²=0.96 and 0.95).
• The ANN model outperformed all other models, including previously published empirical models.

Conclusions:

• Machine learning models, particularly ANN, offer superior predictive power for pesticide drift compared to conventional methods.
• ANN and SVR models effectively handle the complex relationships inherent in pesticide drift variability.
• The ANN model presents a promising, data-driven approach for predicting ground drift, warranting further investigation.