Multi-scale estimation of vegetation nitrogen concentration using a physically informed neural network

This study presents a Physically Informed Neural Network (PINN) designed for estimating nitrogen concentration (N%) in various vegetation scales, a critical aspect of sustainable land management. The PINN effectively integrates spectral data with Radiative Transfer Model (RTM) features to establish cross-scale relationships, addressing the complexities inherent in physical processes. Leveraging a comprehensive spectroscopy dataset and pinpointing critical spectral bands linked to chlorophyll and protein content, our model outperforms traditional methods in N% estimation, achieving an overall R2 of 0.71 and an RMSE of 0.42 (%N) on validation data collected at multiple scales. This innovative approach reveals new insights into spectral data and N% relationships and facilitates the development of accurate and transferable models applicable to various vegetation types.