Improvement method for tea leaf moisture content prediction using VIS-NIR spectrum based on transfer learning.

Abstract

Moisture significantly affects tea plants' growth and quality. Traditional methods of leaf moisture detection are usually destructive to samples, slow and labour-intensive. In this study, visible-near infrared (VIS-NIR) spectroscopy was used to detect the moisture content of tea leaves quickly and accurately in the spectral range of 500-870 nm. The experimental materials are "Longjing 43″, which are divided into two batches. The first batch consists of 135 tea samples collected in April 2022, and the second batch includes 349 tea samples collected in April 2024.The FD + SNV + CARS + ε-SVR model had the best prediction effect on the moisture content of tea leaf in 2024, with the prediction effects of R c , R p , RMSEC, RMSEP and RPD being 0.9676, 0.903, 0.0221, 0.04 and 2.3367, respectively. However, the prediction result R P of the constructed model applied to the 2022 data was only 0.138. In order to improve the generalisation of the model, this study proposes stacking ensemble learning and instance-based transfer learning. In particular, the transfer learning model only needed 55 transfer samples, and the R P was the highest at 0.851. Compared with the stacking ensemble, which required 60 samples, the R P was the highest at 0.85, which realised the use of fewer samples to achieve a better prediction effect. These studies not only confirmed the potential of VIS-NIR spectroscopy to assess the moisture content of tea leaves but also investigated the transfer optimisation of the model, which was helpful to improve the generalisation ability of the model.

Plant phenotyping relevance

VIS-NIR分光法と転移学習モデルにより茶葉の水分含量という植物形質を非破壊推定し、モデル性能と汎化を検証しているため、表現型取得・推定手法が中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.