Intercomparison of machine learning models for estimating leaf area index of rice using UAV-based multispectral imagery

Abstract

Leaf Area Index (LAI) serves as a crucial biophysical indicator, providing valuable insights into canopy vigor and water use. Accurate LAI estimation is essential for crop monitoring, and crop yield prediction. The present study assessed the efficacy of different machine learning (ML) algorithms in estimating LAI obtained from field experiment conducted in Roorkee, India, where rice was grown under two irrigation techniques, and three nitrogen levels. LAI was measured using a ceptometer and images were captured from an Unmanned Aerial Vehicle (UAV)-borne multispectral sensor. Nine ML models were built using 20 vegetation indices, which included Multiple Linear Regression (MLR), Ridge Regression, Lasso Regression, Elastic Net Regression, Extreme Gradient Boosting Regression (XGBoosting), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Artificial Neural Network (ANN) and Random Forest (RF). Among the vegetation indices used in the study, ENDVI and EG showed the highest positive correlation (r = 0.68) with LAI values. Other vegetation indices such as NDVI, ARVI, OSAVI, and NDI also had a positive correlation (r ≥ 0.60) with LAI values. During model testing, lower R² values were recorded for MLR (0.74), Ridge (0.75), Lasso (0.78), ElasticNet (0.74), and XGBoosting (0.77) models, while KNN (0.82), SVM (0.84), ANN (0.83), and RF (0.80) models performed better. Overall, the SVM outperformed other ML algorithms in predicting the LAI of rice under different treatments. Our study demonstrated that UAV-based multispectral images coupled with ML algorithms are capable of producing LAI of rice with reasonable accuracy.