Maize biomass estimation by integrating spectral, structural, and textural features from unmanned aerial vehicle data

Abstract

The rapid and accurate estimation of maize aboveground biomass (AGB) and organ biomass at the field scale is crucial for monitoring crop growth and predicting yield. However, there is limited research on estimating crop organ biomass from unmanned aerial vehicle (UAV) remote sensing. This study used a multispectral (MS) camera and LiDAR sensor to acquire data at various maize growth stages across two experimental regions. The variations in maize organ biomass throughout the growing season were analyzed. Vegetation indices (VIs), canopy structure features (SFs), and texture features (TFs) were combined to create five different datasets and fed into two ensemble learning methods, i.e., Random Forest Regression (RFR) and XGBoost Regression (XGBR), to estimate maize AGB and organ biomass. The results indicated that: (i) Leaf and stalk biomass almost ceased to change after the tasseling stage. Stalk and ear biomass, compared to leaf biomass, are more strongly correlated with AGB. (ii) AGB estimation was improved by incorporating more indicators into the ensemble learning model, with the RFR model with all indicators achieving the best estimation accuracy (R² = 0.917, RMSE = 189.664 g/m², rRMSE = 21.2 %, MAE = 124.617 g/m²). (iii) Leaf and ear biomass estimation was comparable using models inputting all indicators or inputting VIs+TFs, suggesting that MS data were significant for leaf and ear biomass estimation, while SFs played an important role in stalk biomass estimation. This study accurately estimated organ-level maize biomass and AGB by combining different types of UAV remote sensing indicators and machine learning, which provides a valuable reference for organ biomass estimation of other crop types and related precision agriculture studies.