Explainable machine learning models for corn yield prediction using UAV multispectral data

Abstract

• Ensemble ML models achieved higher accuracy in corn yield prediction. • Model-agnostic tools transform regular ML into explainable ML models. • Permutation-based variable importance method is robust. • SHAP model, PDPs, and ICE plots are useful in explaining ML models. • Explainable ML models are recommended over regular ML models. Accurate and reliable corn ( Zea mays L.) yield prediction is essential for optimizing corn production management practices for closing yield gaps. Remote sensing data integrated with Machine Learning (ML) models have been widely used in crop yield prediction. However, conventional ML models are often criticized for their ‘black-box’ nature, impeding transparency into their internal processes and the rationale behind their predictions. To address this limitation, this study develops explainable ML models to predict corn yield using high-resolution multispectral data captured by Unmanned Aerial Vehicles. Among thirty variables derived from multispectral data, Simplified Canopy Chlorophyll Content Index and Transformed Chlorophyll Absorption Reflectance Index were found to be the most suitable inputs. To accurately predict corn yield, we evaluated the Generalized Linear Model (GLM), K-Nearest Neighbor (KNN), Principal Component Regression (PCR), Random Forest (RF), Support Vector Machine (SVM), and Bayesian Regularized Neural Networks (BRNN). We further evaluated various ensemble models combined using GLM and RF. Among individual models, SVM outperformed others with R 2 = 0.67 and RMSE = 1.56 Mg ha -1 . Ensemble models combined using RF produced significantly higher prediction accuracy (R 2 = 0.90 and RMSE = 0.86 Mg ha -1 ) than those combined using GLM (R 2 = 0.70 and RMSE = 1.52 Mg ha -1 ). Explanation tools provided valuable insights into three key aspects of ML modeling: (a) the contribution of variables to model performance, (b) the influence of variables on model predictions, and (c) the distribution of residuals. These insights offer a comprehensive understanding of how ML models accurately predict corn yield, fostering trust in ML models, and promoting their adoption in precision agriculture.

Topics & Concepts