A method for monitoring moisture content in maize seeds using a deep temporal network based on hyperspectral features

Abstract

The moisture content of maize seeds is a key factor affecting seed quality, germination vigor, storage safety, and shelf life. Rapid, accurate, non-destructive detection can help prevent seed decay and mold growth, thereby reducing economic losses. This study employed hyperspectral imaging (400–1000 nm) to collect 237-band spectral data from 300 Haimai515 maize seeds, resulting in 71,100 pixel-level datasets aimed at achieving rapid, non-destructive, and accurate prediction of seed moisture content. Five machine learning algorithms-Ridge regression, Lasso regression, Support Vector Regression (SVR), CatBoost, and Partial Least Squares Regression (PLSR)-were evaluated for their predictive performance. Among the evaluated models, the one that combined Gaussian Window Smoothing (GWS) preprocessing with Gradient Boosting Decision Tree (GBDT)-based feature extraction for PLSR achieved the best performance, namely the GWS-GBDT-PLSR model, achieved the best performance with an R² of 0.953 and an RMSE of 1.557. To further improve prediction accuracy, a deep temporal learning model (GWS-LSTM) was developed using the same GWS preprocessing. This model achieved superior performance, with an R² of 0.978 and an RMSE of 1.461. The GWS-LSTM model improves accuracy while simplifying preprocessing and feature selection, providing an efficient, non-destructive moisture detection method.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.