Machine Learning-Based Band Selection for the Design of a Low-Cost VIS-NIR Microspectrometer for Smart Leaf Nutrient Monitoring

Band selection techniques are widely explored in spectroscopic analysis to enhance predictive performance. This study aims to inform the design of the spectral dispersion component of a low-cost smartphone microspectrometer by identifying the most informative bands within practical constraints on the number of bands and spectral range. The goal is to enable rapid, affordable, and accurate leaf Nitrogen (N) assessment as part of the DEMETRA project, which promotes methodological and technological innovation for smart nutrient estimation in agriculture. Three band selection methods were evaluated: Permutation Feature Importance (PFI), Recursive Feature Elimination (RFE), and Genetic Algorithms (GA), each integrated with machine learning regression models. Among these, GA proved most effective, outperforming models based on the full spectral range. The results highlight the potential of targeted band selection for cost-effective, in-field spectroscopic tools.