Intelligent chlorophyll estimation by attention-integrated deep learning and dual-modal fusion in tencha drying using snapshot multispectral camera.

Abstract

Background: Chlorophyll content during the drying process of tencha, as the precursor of matcha before grinding, is bound up with sensory evaluation of the final product. This study employed a snapshot multispectral technology in conjunction with chemometrics to assess the chlorophyll content of drying samples. Multispectral images consisting of 25 bands ranging from 660 to 924 nm were collected. Reflectance data from the spectra were fused with 10 grayscale texture features to develop regression prediction models. Subsequently, convolutional neural network integrated with the attention mechanism and Squeeze-and-Excitation ResNet18 (SE-Res18) models were built and compared.

Results: The results demonstrated that this fusion approach significantly improved prediction accuracy over using a single data source. The SE-Res18 model with min-max normalization processing achieved a correlation coefficient of 0.9814 for the training set and 0.9337 for the testing set; the relative percent deviation value was 2.79.

Conclusion: The results underscore the viability of leveraging snapshot multispectral technology coupled with spectral-image dual-modality fusion for precise chlorophyll content monitoring. This innovative approach provides a rapid, non-invasive solution for assessing chlorophyll levels in tencha, offering enhanced quality control throughout the drying process with emerging technological capabilities. © 2025 Society of Chemical Industry.