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

Area of Science:

• Forensic Science
• Spectroscopy
• Machine Learning

Background:

• Hyperspectral Imaging (HSI) is crucial for non-destructive forensic analysis, detecting subtle spectral variations.
• Beverage stains are common evidence at crime scenes, requiring reliable identification methods.

Purpose of the Study:

• To investigate the potential of HSI and deep learning for identifying nine types of beverage stains.
• To compare the performance of four deep learning architectures for stain classification.

Main Methods:

• Collected hyperspectral images of beverage stains (Papaya, Coffee, Pomegranate, Orange, Tea, Wine, Whisky, Rum, Brandy) in a controlled mock crime scene.
• Utilized ANOVA-based feature selection to reduce spectral redundancy, selecting 162 bands from 204 in the visible and near-infrared range.
• Trained and evaluated four deep learning models: Multi-Layer Perceptron (MLP), 1D-CNN, LSTM, and CNN-LSTM using standardized spectral data and five-fold cross-validation.

Main Results:

• The Multi-Layer Perceptron (MLP) model achieved the highest classification accuracy of 95.58%.
• All four deep learning architectures demonstrated effectiveness in classifying beverage stains based on spectral characteristics.
• Feature selection using ANOVA successfully reduced spectral redundancy while retaining discriminative information.

Conclusions:

• Hyperspectral imaging coupled with deep learning offers a promising non-destructive method for identifying beverage stains in forensic investigations.
• The MLP architecture demonstrated superior performance for this specific classification task.
• This approach has significant potential to enhance forensic crime scene analysis.