Dual-branch feature-enhanced neural network for apple SSC estimation from hyperspectral imaging

Abstract

Rapid and accurate assessment of apple soluble solids content (SSC) is essential for breeding research and enhancing marketing efficiency. Hyperspectral data provides rich spectral-spatial information for internal quality assessment, while conventional machine learning methods rely on manual feature engineering and linear assumptions, limiting their ability to capture complex spectral characteristics. Although deep learning techniques offer improved representation learning, existing architectures often fail to model global spectral dependencies and lack customized design for hyperspectral data properties. To address these limitations, we propose a dual-branch feature-enhanced network (DBFENet) for rapid, non-destructive estimation of SSC in apples using hyperspectral imaging. DBFENet integrated reconstruction learning and regression tasks within a complementary framework. The reconstruction branch employs a self-supervised autoencoder network to learn features that preserve essential, generalizable spectral information by accurately reconstructing the original input. The regression branch incorporates a 2D attention mechanism to capture long-range spectral dependencies beyond local patterns. This dual-branch design enables more robust and generalized feature extraction from high-dimensional spectral data. Comprehensive experiments demonstrate that DBFENet significantly outperforms six state-of-the-art methods, including PLSR, RR, SVR, 1D-CNN, MLP, and ResNet18-1D, achieving an Rp of 0.9437 and MSE of 0.2485. The results validate DBFENet as an effective tool for non-destructive SSC evaluation, providing a significant advancement in hyperspectral data analysis for agricultural product quality monitoring.

Plant phenotyping relevance

リンゴ果実のSSCという植物器官形質を対象に、ハイパースペクトル画像から非破壊推定するニューラルネットワークを開発し、既存手法と比較検証しており、表現型取得・推定法が中心である。

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.