A hyperspectral imaging framework integrating band selection and deep learning for beverage stain classification in forensic analysis.

The technique of Hyperspectral Imaging (HSI) is significant in the field of non-destructive forensic crime scene investigation, as it allows for the identification of minor spectral changes over a broad range of wavelengths. In the present study, the spectral characteristics of nine beverage stains, such as Papaya, Coffee, Pomegranate, Orange, Tea, Wine, Whisky, Rum, and Brandy, were studied by simulating a controlled environment for a mock crime scene. The hyperspectral images were collected by an HSI system with 204 spectral bands in the visible and near-infrared (VNIR) range. To eliminate spectral redundancy, the ANOVA-based feature selection technique was implemented, which selected 162 spectral bands. These spectral characteristics were employed to train four architectures of deep learning for the classification of the beverage stains, which were implemented as Multi-Layer Perceptron (MLP), one-dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM), and CNN-LSTM. The training process was carried out using standardized spectral data with adaptive learning rates and early stopping for stable convergence. The strength of the models was evaluated through five-fold cross-validation on stratified data. The experimental results demonstrate that the MLP model attained the greatest classification accuracy of 95.58%. The results show that there is great potential for combining hyperspectral imaging with deep learning for non-destructive stain identification in forensics.