Estimating rice leaf area index at multiple growth stages with Sentinel-2 data: An evaluation of different retrieval algorithms

Abstract

Leaf area index (LAI), which is closely related to canopy physiological processes such as photosynthesis and water utilization, serves as an important biophysical parameter for monitoring rice (Oryza sativa L.) growth. However, due to significant variations in the water-soil mixed background during the rice growing season, the optimal LAI retrieval method for rice throughout the whole period remains unclear. Here, different LAI retrieval methods, categorized into vegetation index (VI)-, look-up table (LUT)- and machine learning (ML)-based groups, were evaluated for rice at multiple growth stages using Sentinel-2 images. Particularly, the performance of rice LAI derived from the optimal retrieval model was comprehensively analyzed to understand factors influencing LAI estimation during different growth stages. Results suggested that the Gaussian Process Regression (GPR) in the ML-based group achieved the best performance for rice LAI estimation in the whole growth period (R² = 0.75, RMSE = 0.60, RRMSE = 15.81 %), followed by the normalized difference VI (NDVI) in the VI-based group (R² = 0.74, RMSE = 0.61, RRMSE = 15.98 %) and the cost function K(x)=log(x)+1/x in the LUT-based group (R² = 0.70, RMSE = 0.69, RRMSE = 18.09 %). Notably, although the VI-based LAI retrieval method incorporated ground measurements to build empirical LAI-VI relationships, a single VI with parametric regression cannot well capture variations in rice LAI across growth stages compared to the ML-based method using simulation dataset from the physical model. The optimal ML-based method also exhibited better performance in rice LAI estimation than similar studies, with R² increasing by 0.14 and RMSE decreasing by 0.18. Furthermore, based on ground LAI measurements, the water-soil mixed background is a primary influencing factor for rice LAI estimation in the tillering, jointing, and booting stages (RMSE: 0.39–0.79), whereas the saturation effects should be considered in the full heading stage (RMSE = 0.68). Overall, this study indicates that the GPR-based retrieval strategy is the optimal method for generating rice LAI datasets over the whole growth period, providing valuable reference for precision agriculture application such as field irrigation, fertilization management, and yield estimation.