|
Hyperspectral imaging (HSI) systems provide non-destructive and high-resolution analysis in agriculture, healthcare, food safety, and industry-related problems. Despite its numerous advantages, HSI faces challenges due to large data generation, high computational power demands, and complexity in real-time monitoring due to redundancy in acquired data. There is an increasing demand to develop a less complex system for non-destructive analysis for various applications. On the same note, this study presents a potential solution for reducing the redundant information from the acquired data. The idea was investigated on HSI signals obtained from gongura (Hibiscus sabdariffa), amaranthus (Amaranthus viridis), and banana (Musa acuminata) leaves with two hundred four wavelengths and identifying the signature wavelengths to classify leaves. Twenty of the two hundred-four wavelengths were selected as signature wavelengths based on the importance scores obtained from the linear discriminant analysis (LDA). These signature wavelengths lie in the visible range of the electromagnetic spectrum. Afterwards, these twenty wavelengths were employed for further experimentation. The current study utilized extra tree classifiers (ETC), random forest (RF), and LDA classifiers for leaf classification. The ten-fold cross-validation findings indicated that LDA performed well among the other classifiers for full-range and signature twenty wavelengths. The promising results demonstrated the effectiveness of signature twenty wavelengths for the classification of leaves applications. Furthermore, these signature wavelengths also open the door to developing a low-complexity HSI system for future studies. Also, the robustness of the model can then be increased by utilizing large amounts of data.
|
