Abstract: Aiming at the problems of destructive traditional fruit quality detection methods such as sampling assay, spectral information loss and incomplete feature extraction in the existing regression prediction models,a novel model and method based on near‑infrared spectroscopy analysis technology and a one‑dimensional Convolutional neural network (1D-CNN) topredict the soluble solids content of citrus fruit were proposed in this paper. A dataset has been established by gathering near‑infrared spectroscopy of citrus fruits and determining their soluble solid content. The network structure parameters have been optimized to enhance the performance of the model, including the depth of the network structure, the size and quantity of convolutional kernels, the presence or absence of batch normalization (BN) layers, pooling methods, the depth of fully connected layers, and Dropout values. The final 1D-CNN model comprises two convolutional layers, two BN layers, two maximum pooling layers, and two fully connected layers, with a Dropout value of 0.2. Traditional Partial least squares regression, principal component regression, and support vector machine regression prediction models have been established to compare their performance with the 1D-CNN model. The outcomes reveal that the 1D-CNN model exhibits significantly superior prediction accuracy and model stability compared to conventional algorithms. The Root‑mean‑square deviation on the verification set is 0.333 9, and the Coefficient of determination is 0.865 5. This demonstrates that the 1D-CNN model can carry out feature extraction of citrus near‑infrared spectral data, thereby permitting non‑destructive detection of soluble solid content in citrus. Consequently, this approach provides a better solution for non‑destructive detection of citrus based on near‑infrared spectral analysis technology.

Key words: citrus, near?infrared spectroscopy, convolutional neural network, soluble solids, non?destructive testing

摘要： 针对抽样化验等传统果实品质检测方法具有破坏性以及现有回归预测模型存在光谱信息损失和特征提取不够完备等问题，提出基于近红外光谱分析技术和一维卷积神经网络（1D-CNN）实现蜜柑果实可溶性固形物含量预测的模型和方法。采集蜜柑的近红外光谱和测定可溶性固形物含量建立数据集。并通过试验对比确定使模型性能最优的网络结构深度、卷积核尺寸和数量、有无批量归一化（BN）层、池化方式、全连接层深度和Dropout值等网络结构参数，形成包含2层卷积层、2层BN层，2层最大池化层和2层全连接层的一维卷积神经网络，并设置Dropout值为0.2。与偏最小二乘回归、主成分回归和支持向量机回归预测模型的性能对比试验表明：提出的1D-CNN模型预测精度和模型稳定性均优于传统回归预测算法，其验证集上的均方根误差为[0.333 9]，决定系数为[0.865 5]，能够实现对蜜柑近红外光谱数据特征的有效提取和对蜜柑可溶性固形物含量的无损检测。

关键词: 蜜柑, 近红外光谱, 卷积神经网络, 可溶性固形物, 无损检测