Publication

Abstract : The perfect classification of corn-growing stages using hyperspectral data presents major challenges owing to the huge dimensionality of the spectral information and class imbalance among phenological phases. This study proposes a novel hybrid structure that combines spectral band optimization, wavelet feature extraction, ensemble oversampling-based data augmentation, and a trivial convolutional neural network (CNN) for robust corn data classification. Initially, the Dream optimization algorithm (DOA) was used to find the most optimal spectral bands from the hyperspectral data. Multiscale features were extracted from the selected bands using discrete wavelet transform. To address the class imbalance, the ensemble oversampling method was fused with pixel pair feature augmentation and ensemble training. Finally, an efficient CNN architecture was suitable for the improved feature set, and two-level majority voting was employed during interpretation to improve stability and accuracy. The experimental results validate that the proposed CNN method outperforms traditional architectures, such as ResNet-18, VGG-16, Imagenet-250, and GoogleNet, across multiple performance metrics, including accuracy, F1-score, and specificity, while also reducing the inference time. This framework provides an effective and scalable solution for precision agriculture applications involving corn-hyperspectral analysis.