Combining machine learning with a mechanistic model to estimate maize nitrogen content from UAV-acquired hyperspectral imagery

Crop nitrogen (N) content reflects crop nutrient status and is an important trait in crop management. Over the decades, non-destructive N estimation has greatly benefited from remote sensing and data-intensive computational approaches. However, previous studies mostly focused on the estimation accuracy under a specific environment; few of them considered estimation robustness across varying growth conditions. As climate change intensifies, crops are facing more unexpected stresses. It is critical to improve N estimation under changing environments with better model generalizability. Thus, we proposed a novel hybrid method with merits of both mechanistic and machine learning models and integrating in-situ data and simulated data for an improved model training. The in-situ data were the canopy reflectance extracted from hyperspectral images collected by an Unmanned Aerial Vehicle (UAV) and destructively sampled plant N content;the simulated data referred to the canopy reflectance simulated by a mechanistic model, the PROSAIL-PRO. The performance of the hybrid method was compared with one of the most popular machine learning models (i.e., Gaussian Process Regression, GPR) across three study sites. Results showed that the hybrid method outperformed the GPR by reducing RRMSE up to 6.84% on canopy nitrogen content (CNC) estimation. It also achieved more stable performances across varying soil water and N availabilities. Altogether, we demonstrated an approach to estimate CNC under diversesoil and environmentalconditions from remotely sensed spectral data with better accuracy and generalizability. It leverages the robustness of mechanistic models and the computational efficiency of machine learning models and has great potential to be transferred to other crops and many common crop traits.