Enhancing maize above-ground biomass estimation through multispectral, digital and LiDAR fusion on UAV platforms

Accurate estimation of maize Above-Ground Biomass (AGB) is critical for precision field management and accelerated programs. Existing vegetation index (VI)-based maize AGB estimation methods using unmanned aerial platforms show limited generalizability due to insufficient integration of environmental and management factors. To address this limitation, plant density (PD) and spacing between two plants (STP) were extracted from digital imagery, while slope (SL) was calculated using LiDAR data. A plant structure factor (PS) was developed using PD and STP to quantify planting characteristics. The influence of SL on soil water dynamics, nutrient transport, and dry matter accumulation patterns was systematically analyzed. Finally, a machine learning (ML) model for maize AGB estimation with multimodal data fusion was constructed using Partial Least Squares (PLSR), Support Vector Machine (SVM), and Random Forest (RF). Results indicated that SL significantly affected soil moisture distribution and nutrient transport across all growth stages (V6, VT, R4). Areas with gentler slopes showed nutrient-moisture accumulation, resulting in strong negative correlations between slope and AGB (V6: −0.62; VT: −0.69; R4: −0.69, p < 0.001). PS exhibited a significant positive correlation with maize AGB in the three growth stages (V6: 0.58; VT: 0.59; R4: 0.63, p < 0.001). Among all validated models, RF achieved superior performance in both VI-based and multimodal (VI+PS, VI+SL, VI+PS+SL) configurations, outperforming SVM and PLSR. Incorporating either PS or SL as additional input features improved AGB estimation accuracy compared to VI-only models. No significant accuracy difference was observed between the best-performing VI+PS-RF and VI+SL-RF models, confirming the validity of both PS and SL as input features. Fusion of the three-modal data enabled the VI+PS+SL-RF model to achieve the highest estimation accuracies for V6, VT, and R4 stages, with validated R2 values of 0.93, 0.90, and 0.92, and RMSE values of 84.37, 615.05 and 1326.82 kg/ha, respectively. Spatial validation across three regions revealed that model robustness is influenced by environmental uniformity; the model achieved higher accuracy in a stable environment (FQ: R²= 0.85, RMSE=3687.54 kg/ha) than in a heterogeneous one (YZ: R²= 0.56, RMSE= 4325.16 kg/ha). This study underscores the importance of multimodal data fusion in maize AGB estimation via remote sensing inversion. We pioneer the validation of a “phenotypic feedback (VI) + management (PS) + environment (SL)” triadic integration framework, addressing the critical gap in existing models that neglect ecological interactions between planting structure and terrain. The synergistic integration of PS and SL improved model robustness by 15.8 % (R²) and 10.9 % (RMSE) compared to VI-only models, demonstrating their unique ecological value in regulating soil moisture distribution (SL: correlation up to −0.69) and optimizing light competition (PS: correlation up to 0.63). The results offer technical support and opportunities for enhanced field management and growth monitoring in smart agriculture.