Effect of germ orientation during Vis-NIR hyperspectral imaging for the detection of fungal contamination in maize kernel using PLS-DA, ANN and 1D-CNN modelling

Fungal contamination of maize during pre and post-harvest is rampant and omnipresent. Hyperspectral imaging (HSI) is a popular non-invasive technique for detection of fungal contamination in maize kernels and secures acceptance for online implementation. Designed experiments were conducted to identify the effectiveness of models (Partial Least Square- Discriminant Analysis-(PLS-DA); Artificial Neural Network (ANN); and 1D-Convolutional neural network (CNN) for predicting fungal contamination of maize kernels in relation to orientation of germ with respect to the lens of HSI (398–1003 nm) camera. Pixel wise dataset was collected for two groups of maize kernels such as sterile (G1) and Aspergillus niger contaminated (G2) and each having 100 maize kernels imaged in germ-up (GU), germ-down (GD) and germ-randomly (GR) positioned; all the three models were trained with these datasets. It was observed that in terms of “error-rate” the prediction capability was best for GU (1.31, 1D-CNN) followed by GR (1.65, ANN) and GD (1.95, ANN). In case of a mismatch of trained and testing dataset for the various models, it was observed that GR exhibited the lowest average error rates of 5.71, 4.94 and 3.15 for PLS-DA, ANN and 1D-CNN, respectively. From the results obtained in the present study confirmed that HSI along with suitable classification technique with proper germ orientation can be used to separate the infected grains. •Fungal contamination in maize kernels was studied by hyperspectral imaging.•The kernels were imaged with germ-up, germ-down and in mixed groups.•Maize kernel germ orientation influenced the spectral signatures.•Comparative performance of classification models (PLSDA, ANN, ID-CNN) is reported.•1D-CNN model based on germ-up dataset demonstrated the best classification results.