HYPERSPECTRAL IMAGE PROCESSING FOR MULTISPECTRAL SENSOR OPTIMIZATION AND FOLIAR NITROGEN ESTIMATION IN RICE

Monitoring foliar nitrogen is essential for nutritional management in agricultural systems, especially in rice cultivation, due to the strong relationship between nitrogen availability, productivity, and photosynthetic efficiency. In this study, a physiology-informed machine learning approach was proposed for predicting foliar nitrogen content using hyperspectral data and physiological variables. Field experiments were conducted in rice crops subjected to different nitrogen fertilization levels, in which hyperspectral reflectance measurements, SPAD readings, and laboratory analyses of foliar nitrogen were collected for model development and validation. The original spectral data, ranging from 400 to 1000 nm, were resampled into a standardized spectral grid containing 271 bands and used to simulate multispectral sensors with bandwidths ranging from 6 to 120 nm. Spectral indices were generated and submitted to a preprocessing pipeline composed of Savitzky-Golay smoothing, Standard Normal Variate (SNV) normalization, collinearity removal, and data standardization. The physiological variable SPAD was integrated into the model without spectral transformation, acting as a biological bridge between spectral signatures and foliar nitrogen content. The modeling strategy adopted a two-level physiological stacking approach, in which machine learning models were initially trained to predict SPAD values, which were subsequently incorporated as an additional explanatory variable for foliar nitrogen estimation. Random Forest, XGBoost, Linear Regression, and Ridge Regression models were evaluated using nested cross-validation. XGBoost achieved the best performance, with R² values above 0.97, RMSE values below 0.85, and MAE values close to 0.45. The proposed approach demonstrated high potential for the development of optimized multispectral sensors aimed at precision agriculture applications.