UAV-based rice aboveground biomass estimation using a random forest model with multi-organ feature selection
Jing Shi, Kaili Yang, Ningge Yuan, Yuanjin Li, Longfei Ma, Ya-Dong Liu, Shenghui Fang, Yi Peng, Renshan Zhu, Xianting Wu, Yan Gong
Abstract
Aboveground biomass (AGB) is important for monitoring crop growth and field management. Accurate estimation of AGB helps refine field strategies and advance precision agriculture. Remote sensing with Unmanned Aerial Vehicles (UAVs) has become an effective method for estimating key parameters of rice. This study involved four experiments conducted across varied locations and timeframes to collect field sampling data and UAV imagery. Feature extraction, including Vegetation Index (VI), textures, and canopy height, was performed. Key factors influencing biomass estimation across different rice organs were analyzed. Based on these insights, a Random Forest model was developed for AGB estimation. The VIS-Leaf factor-Spike factor-Stem factor (VIS-L-Sp-St) model proposed in this study outperformed traditional methods. The training set achieved an R 2 of 0.89 with a reduced RMSE of 191.30 g/m 2 , surpassing the traditional VIS model (R 2 =0.64, RMSE=363.53 g/m 2 ). Notably, in the validation set, the VIS-L-Sp-St model showed good transferability, with an R 2 of 0.85 and RMSE of 196.55 g/m 2 , outperforming MLR (R 2 =0.02, RMSE=5944.09 g/m 2 ), PLSR (R 2 =0.18, RMSE=934.27 g/m 2 ) methods, BP (R 2 =0.14, RMSE=581.61 g/m 2 ) method and SVM method((R 2 =0.45, RMSE=600.91 g/m 2 ). Sensitivity analysis showed that different rice organs respond differently to specific features. This insight improves feature selection efficiency and enhances AGB estimation accuracy. The organ-specific AGB estimation model highlights its potential to support precision agriculture and field management, contributing to advancements in agricultural research and application. • Model with organ-specific feature selection enhances AGB estimation accuracy, surpassing models with overall selection. • This study analyzes the sensitivity of rice organ-specific AGB to spectral and geometric features. • The model mitigates some accuracy loss caused by spike-leaf mixing and stem occlusion in late growth stages. • The proposed model outperforms models using traditional MLR, PLSR, BP, and SVM methods.