Rice biophysical parameter retrieval with optical satellite imagery: a comparative assessment of parametric and non-parametric models

Abstract

This article presents a comparison of parametric and non-parametric models in rice biomass and leaf area index (LAI) retrieval using optical satellite imagery. Four parametric models including the linear, quadratic, logarithmic and exponential models, and four non-parametric models that include RF, SVM, kNN, and GBDT were applied, respectively, on the optical satellite dataset. GBDT produced the most accurate biomass estimates (RMSE of 191.8 g/m2) before heading and the quadratic model produced the most accurate biomass estimates (RMSE of 364.7 g/m2) after heading. RF registered the most accurate LAI estimates (RMSE of 0.79 m2/m2) before heading, whereas the quadratic model recorded the most accurate LAI estimates (RMSE of 1.04 m2/m2) after heading. Non-parametric models outperformed their parametric counterparts at before heading, whereas the reverse is the case after heading. These findings provide a guide to the optimal choice of empirical models for rice biomass and LAI retrieval with optical imagery.