Estimating chlorophyll content in tea leaves using spectral reflectance and deep learning methods

Abstract

Accurate estimation of chlorophyll content in tea leaves is essential for evaluating plant health, managing fertilization, and optimizing harvest timing in precision agriculture. This study investigates the use of hyperspectral reflectance data (400–850 nm, 5 nm intervals; 91 bands) to estimate chlorophyll content in tea leaves (Camellia sinensis) using three deep learning models: a one-dimensional convolutional neural network (1D–CNN) tailored for spectral regression, a vision transformer (ViT) adapted for one-dimensional inputs, and a self-supervised learning (SSL) model with regression. The key innovation of this study is the introduction of a self-supervised learning framework specifically adapted for spectral data, in which an autoencoder is first trained on unlabeled spectra to learn compact and noise-tolerant representations. These pretrained features are then used in a downstream regression task to predict chlorophyll content, allowing effective use of limited labeled data. To our knowledge, this is the first application of SSL in chlorophyll estimation using high–resolution leaf–level spectral measurements. Among the three models, the SSL approach achieved the highest accuracy, with a root mean square error (RMSE) of 3.33 μg/cm², outperforming both the 1D–CNN (5.05 μg/cm²) and ViT (4.28 μg/cm²). These findings demonstrate that SSL is particularly effective for capturing subtle spectral patterns and improving prediction performance, especially when labeled data are scarce. This study highlights the potential of combining hyperspectral sensing with advanced representation learning to non–destructively monitor chlorophyll dynamics in tea cultivation, supporting more sustainable and data–driven agricultural practices.