Monitoring leaf nitrogen concentration for different leaf layers based on multiangle remote sensing and fusion models

Timely and accurate monitoring of changes in leaf nitrogen concentration (LNC) across different leaf layers of crops is crucial for early management of nitrogen fertilization. Here, we collected LNC data from different leaf layers of winter wheat and hyperspectral data from 13 view zenith angles (VZAs) and then employed competitive adaptive reweighted sampling (CARS) to screen sensitive original reflectance (OR), vegetation indexes (VIs), and wavelet features (WFs). Three machine learning algorithms (MLs)—namely partial least squares regression (PLSR), back propagation neural network (BPNN), and extreme learning machine (ELM)—were used to invert vertical distribution of LNCs. The results demonstrated that ±60°, ±50°, and 0° were suitable for inverting upper layer LNC; ±40°, ±30°, and 0° were suitable for inverting middle layer LNC; and ±20°, ±10°, and 0° were suitable for inverting bottom layer LNC. The accuracy of MLs for the vertical distribution of LNCs consistently ranked as PLSR > BPNN > ELM. The combination of multiple data modalities achieved higher accuracy than using a unimodal dataset; the OR+VI+WF_PLSR integrated model yielded the highest accuracy in the upper, middle, and bottom canopy layers (with R² values of 0.93, 0.88, and 0.86, respectively). In particular, for bottom layer LNC, the accuracy of OR+VI+WF_PLSR was 72% higher than that of OR_ELM. In conclusion, combining multiangle hyperspectral data with MLs improves the prediction accuracy and stability of LNC vertical distribution assessment and provides a technical basis for rapid large-scale monitoring of bottom layer LNC.