VNIR-SWIR hyperspectral spectroscopy for nitrogen assessment in potato crops: Deep learning and regression models across field and laboratory conditions

Accurate evaluation of nitrogen (N) status is essential for optimizing fertilizer application and sustaining yield in potato production systems. This study investigated the potential of Visible–Near Infrared to Short-Wave Infrared (VNIR-SWIR) hyperspectral reflectance spectroscopy for in-season prediction of petiole nitrate N (PNN) under both field and laboratory conditions. Spectral data were collected from field canopies as well as from fresh and dried leaf tissues and analyzed using three predictive approaches: One-Dimensional Convolutional Neural Network (1D-CNN), Support Vector Regression (SVR), and Partial Least Squares Regression (PLSR). Variable Importance in Projection (VIP) analysis was applied to identify key wavelength regions associated with N prediction across different measurement conditions. Among the tested models, 1D-CNN achieved the highest accuracy for laboratory fresh leaves with a coefficient of determination (R²) of 0.90, root mean square error (RMSE) of 0.22 percent, and residual prediction deviation (RPD) of 3.18. SVR showed moderate accuracy, while PLSR exhibited the weakest predictive capability. Calibration transfer using Piecewise Direct Standardization (PDS) improved the performance of field-based models, increasing the R² from 0.77 to 0.82 for 1D-CNN and from 0.24 to 0.44 for SVR, confirming the effectiveness of spectral alignment for reducing environmental variability. Overall, laboratory-based fresh leaf spectroscopy provided more reliable N estimates than field and dried-tissue measurements, and the deep learning approach outperformed classical and machine learning regression models.