Interspecies Prediction of Nitrogen Content in Processed Plant Samples Using Spectroscopic Modeling and Transfer Learning

ABSTRACT Employing machine learning models on preprocessed samples is an effective alternative for leaf nitrogen quantification, reducing analysis time and improving fertilizer efficiency. This study evaluates predictive performance and transfer learning of models for nitrogen (N) concentration across different plant species, along with visual analysis of spectral patterns. A spectral dataset was developed using pre‐processed samples from crops (coffee, pear, sugarcane, bean, and maize), forage (Brachiaria), and ornamental plants (e.g., Gypsophila). Leaf samples were collected from field‐grown plants, oven‐dried at 60°C with forced air circulation, and ground to 2.0 mm. Spectra were measured with a FieldSpec spectroradiometer (350–2500 nm). Visual analysis compared plants of distinct photosynthetic cycles (C3 and C4) and among species of the same cycle. Nitrogen quantification was performed using Partial Least Squares Regression (PLSR) and Random Forest (RF). Transfer learning was assessed in three ways: (i) between species; (ii) temporal stability; (iii) evaluation with an independent dataset comprising multiple agricultural species. Results showed C3 plants had lower reflectance in the 400–670 nm bands and higher N levels compared to C4 crops. Regardless of crop type or photosynthetic cycle, characteristic absorption features were detected at 530 nm and 615 nm, absent in fresh samples. PLSR achieved superior performance ( R 2 = 0.95, RMSE = 2.16 g kg −1 , MAPE = 10.70%) compared to RF ( R 2 = 0.88, RMSE = 3.4 g kg −1 , MAPE = 13.48%). Edaphoclimatic and physiological conditions influenced transfer learning, highlighting the potential and limitations of applying spectral models across species and environments.