A cumulative learning method for pixel-level hyperspectral detection of aflatoxins on peanuts using convolutional neural network

Abstract

This study introduces a novel cumulative learning method to overcome the limitations of hyperspectral data in aflatoxin detection in peanuts. By aggregating spectral characteristics from remote sensing and near-infrared spectral images, the method enhances detection accuracy. We validate its effectiveness through comparative model analysis, utilizing superimposed images of similar materials to address data heterogeneity and low resolution. The results demonstrate that the cumulative learning model's performance is significantly improved, with all six methods achieving accuracies above 0.97, surpassing the original 1D-CNN and traditional transfer learning models. Additionally, compared to advanced semi-supervised models, the cumulative learning method exhibits superior performance, with accuracies exceeding 0.95. This approach not only reduces model complexity and data collection costs but also effectively enhances classification accuracy in peanut aflatoxin detection, thereby facilitating efficient online monitoring.