Advanced optical-based water stress detection in lettuce: a novel wavelength selection technique

Water stress is a critical factor limiting agricultural productivity, particularly in controlled environments such as greenhouses, where precision irrigation is essential. Accurate and timely detection of crop water status is vital to prevent yield losses and enhance resource use efficiency. Visible-near infrared spectral reflectance data can be acquired through non-destructive leaf-clip measurements and used to monitoring water stress in lettuce-grown plants. However, given the high dimensionality of hyperspectral data, which often includes redundant or non-informative variables that can lead to collinearity, several wavelength selection algorithms have been proposed to enhance classification performance. In this study, a new method for spectral features selection, called iGA-BOSS , is developed, specifically conceived for water stress detection. Its effectiveness was proven on lettuce plants under different water stress conditions induced through controlled irrigation deficits relative to optimal watering regimes. Spectral reflectance measurements were collected in vivo at two growth stages and analyzed using advanced machine learning techniques. For comparison, the reflectance data were processed using the most relevant algorithm available in the literature, and the iGA-BOSS achieved the best performance, with an accuracy of 0.95 and an F1 score of 0.95. The selected features were used to train Partial Least Squares Discriminant Analysis (PLS-DA) for binary classification of water stress. Notably, the proposed method outperforms models trained on the complete wavelength set and competing algorithms, offering the best trade-off between classification accuracy, spectral compactness, and computational efficiency. These results demonstrate the effectiveness of combining robust preprocessing, targeted wavelength selection, and supervised classification models for monitoring spectral water stress. The identified key wavelengths may inform the development of tailored multispectral sensors, enabling scalable and efficient irrigation management in precision agriculture.