Research on potatoes defect classification based on hyperspectral imaging and convolutional neural networks

Abstract

Abstract Potato quality detection is a critical step that determines their market value. However, manual sorting suffers from low efficiency, high cost, and a high misjudgment rate. Therefore, the rapid and accurate classification of defective potatoes is of great economic significance for reducing industrial losses. In this study, a lightweight convolutional neural network (WavebandCNN) was constructed combined with hyperspectral imaging (HSI) technology to achieve rapid and accurate classification of four categories of potatoes: healthy, greening, skin damage, and dry rot. First, hyperspectral images of 400 potato samples were collected and calibrated, and regions of interest (ROI) were extracted to construct a spectral dataset. The performance of the lightweight convolutional neural network was evaluated using raw spectra and five preprocessed spectra, respectively, and compared with three traditional machine learning models: Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM). Meanwhile, the successive projections algorithm (SPA) was employed to select characteristic wavelengths for data dimensionality reduction. The results show that WavebandCNN achieved the highest classification accuracy of 93.11% with raw spectra, significantly outperforming all comparative models. After screening 20 characteristic wavelengths via SPA, the classification accuracy was improved to 95.98%, while the training time and data redundancy were greatly reduced. This study confirms that the combination of hyperspectral imaging technology and the WavebandCNN model enables accurate identification of potato defects, providing a new approach for the online detection and practical application of potato defects.

Plant phenotyping relevance

ジャガイモの欠損・病変状態をハイパースペクトル画像とCNNで直接分類する手法を構築・比較・評価しており、植物状態の取得・推定が研究の中心である。

Code and data availability

The supplied blocks describe a potato defect classification study using hyperspectral imaging and a WavebandCNN model, but contain no data availability statement, no public dataset or code deposit, and no author-provided URLs. The hyperspectral dataset, ROI-extracted spectral pixels, and model code are not stated as公开,

No evidence-backed public reproduction asset is currently recorded.