Non-destructive prediction of total sugar content in tobacco based on a hyperspectral imaging and autoencoder network

Total sugar content is a vital quality indicator for flue-cured tobacco. Conventional chemical analysis is destructive, while deep learning often overfits on limited agricultural datasets. To resolve this bias-variance trade-off, we propose an applied hybrid strategy integrating spectral selection, an autoencoder (AE), and regression. Diverging from direct dimensionality reduction, our approach first employs correlation-based screening as a hard-attention mechanism to prune spectral noise, followed by an autoencoder (AE) to embed nonlinear manifolds into a compact latent space. A conceptual shift is implemented by benchmarking this latent representation against multiple regression heads, specifically using a linear regression (LR) bottleneck to encourage the AE to learn linearly separable features. Compared to PLSR and SVR benchmarks, the AE-LR configuration achieved superior performance (R²: 0.901, RMSE: 2.389, MAE: 1.873, RPD: 2.303, and RPQI: 3.251), demonstrating excellent predictive accuracy and robust generalization capabilities. Ultimately, this hybrid strategy provides an effective and practical approach to managing data complexity in small-sample quantitative analysis by necessitating the learning of robust, disentangled representations.