Non-destructive estimation of chlorophyll content in wasabi (Eutrema japonicum) leaves using spectral reflectance and deep learning models

Accurate, non-destructive estimation of chlorophyll content is essential for monitoring crop physiological status and supporting precision cultivation management. This study investigated the feasibility of combining leaf spectral reflectance with deep learning models to estimate chlorophyll content in hydroponically grown wasabi ( Eutrema japonicum ), while also deriving insights applicable to other leafy crops. A total of 179 leaf samples were collected under diverse nutrient conditions, including variations in pH, sulfur levels, and macronutrient composition across two growing seasons. Spectral data were preprocessed using second-order trend removal followed by a fractional-order derivative (FOD) transformation based on the Grünwald–Letnikov definition to enhance subtle spectral features. Three regression models—a one-dimensional convolutional neural network (1D-CNN), a Vision Transformer (ViT), and a Swin Transformer (SWIN)—were evaluated. The 1D-CNN achieved the highest accuracy when low-order fractional derivatives (0–0.4) were applied, highlighting its sensitivity to localized spectral variations, whereas SWIN performed best with minimally processed original spectra, and ViT showed relatively stable performance across preprocessing methods. These findings indicate that optimal preprocessing strategies depend on model architecture, providing practical guidance for selecting suitable combinations of spectral preprocessing and deep learning models when designing chlorophyll monitoring systems for wasabi and other crops. • Spectral reflectance enables non-destructive chlorophyll estimation in wasabi leaves. • Deep learning models were compared under different spectral preprocessing strategies. • Swin Transformer showed best accuracy using detrended spectra without derivatives. • Optimal preprocessing depends on model architecture and feature extraction mechanisms.