基于深度学习的西瓜可见/近红外光谱可溶性固形物预测模型研究

Prediction model research of SSC in watermelon based on deep learning and visible/near infrared spectroscopy

通过可见/近红外光谱技术结合K最近邻法回归、随机森林回归、深度卷积神经网络及带有残差块的深度卷积神经网络4种化学计量学方法对不同糖度的西瓜进行定量判别,并借鉴适用于图像处理的深度网络模块对可见/近红外光谱进行建模。结果表明,深度学习网络模块一维化在可见/近红外光谱数据处理中体现了巨大潜力,卷积神经网络CNN模型在预测集中Rp为0.8559,RMSEP为0.7781°Brix,加入Res-block后的改进卷积神经网络Res-CNN在预测集中Rp为0.8932,RMSEP为0.7104°Brix。

In this study,near infrared spectroscopy combined with chemometrics building models to was used to detect the soluble solid contents from watermelon,through four methods,K nearest neighbors regression,Random forest regression,convolution neural network,convolution neural network added residual-block.Some deep learning modules with image processing,coding modules in 1-d ways and apply in Visible/Near-infrared spectroscopy for modeling exploration was used.As a result,deep learning modules showed great potential in Visible/Near-infrared spectroscopy data processing,and CNN model got 0.8559 correlation coefficient and 0.7781°Brix RMSEP in prediction-set.The Res-CNN model achieved 0.8932 correlation coefficient and 0.7104°Brix RMSEP in prediction-set.The results of this study could provide a reference for the rapid and non-destructive model development of watermelon quality.