Explainability of deep convolutional neural networks when it comes to NIR spectral data: A case study of starch content estimation in potato tubers

Explainable AI is gaining popularity as a way to understand the decision-making processes of neural networks and gain insight into their predictions. In this paper, Integrated Gradients (IG) was applied to assess the relevance of spectral features used by deep convolutional neural networks in predicting the starch content of potatoes. For this purpose, spectral information of 7651 tubers of 12 potato varieties was acquired using a NIR spectrometer in the spectral range of 940–1650 nm. This was followed by a reference measurement of starch content. Three one-dimensional deep convolutional neural networks i.e. VGG-19, InceptionV3, and SpectraNet-32 were developed using the Keras API. The deep networks outperformed traditional models in the starch content prediction, with SpectraNet-32 achieving the highest prediction accuracy (R² = 0.84, RMSE = 1.41%, RPD = 2.46, and rRMSE = 9.88%). Further analysis of the neural networks by IG indicated that the predictions were generated based on starch relevant spectral bands. The results of this study demonstrated that the deep convolutional neural networks not only could accurately predict starch content in potatoes, but also provided certainty in the predictions.