Rapid and nondestructive prediction of chlorophyll fluorescence parameter Fv/Fm in Juncao leaves based on Vis-NIR spectroscopy and deep transfer learning.

The maximum quantum yield of photosystem II photochemistry (Fv/Fm) is an important chlorophyll fluorescence parameter for evaluating plant photosynthetic activity and stress responses. Reliable Fv/Fm measurement using pulse amplitude modulated fluorometry generally requires approximately 30 min of dark adaptation, which reduces measurement throughput when large numbers of samples need to be evaluated. In this study, a rapid and nondestructive method was developed for predicting leaf Fv/Fm in Juncao seedlings by integrating visible-near infrared (Vis-NIR) spectroscopy, deep learning, and transfer learning. Vis-NIR spectra (400-1075 nm), SPAD, and Fv/Fm data were collected under different temperature treatments. Raw and preprocessed spectra were reconstructed into two-dimensional pseudo-images and used as inputs to an improved 18-layer residual network incorporating a convolutional block attention module (ResNet18-CBAM). The model was compared with conventional machine learning methods, and SHapley Additive exPlanations (SHAP) were used for interpretation. Transfer learning was performed using Jujuncao as the source domain and king grass as the target domain. The results showed that the ResNet18-CBAM model based on pseudo-images reconstructed directly from the raw spectra achieved the best prediction performance, with a prediction set coefficient of determination (RP2) of 0.9703, a root mean square error of prediction (RMSEP) of 0.0242, and a residual predictive deviation (RPD) of 5.8269. SHAP analysis revealed that the most influential wavelengths were mainly associated with photosynthetic pigment absorption, chlorophyll fluorescence emission, lipid related responses, and leaf water status. For the target domain with 120 samples, progressive fine-tuning using candidate samples selected by the Kennard-Stone algorithm showed that the cumulative subset containing 36 samples achieved the highest validation performance under the current sampling protocol, yielding a validation coefficient of determination (RV2) of 0.8154, representing an improvement of 19.06 percentage points relative to the model without target domain fine-tuning. These results demonstrate that integrating Vis-NIR spectroscopy with deep transfer learning enabled rapid and nondestructive Fv/Fm prediction across Juncao species, offering an effective approach for rapid assessment and high-throughput phenotyping of photosynthetic activity in Juncao and other bioenergy crops.