Hyperspectral Imaging Empowers Intelligent Identification of Trace Beverage Evidence in Forensics

Forensics Preface
Trace stains of beverages like coffee, juice, and alcohol are ubiquitous at crime scenes, making it extremely difficult to distinguish their types with the naked eye. Traditional detection methods mostly rely on sampling and chemical color-development processing, which can easily damage physical evidence and are detrimental to subsequent review. Nature published a study that constructed a complete analysis solution combining hyperspectral imaging, band selection, and deep learning. This solution conducted non-destructive identification experiments on nine categories of beverage stains, providing a new perspective for criminal investigation trace evidence detection. Integrating this study, CHNSpec discusses the landing value of hyperspectral technology in the forensic field.

I. Existing Difficulties in Traditional Physical Evidence Detection
1.Visual recognition has significant limitations. Coffee and tea, or whiskey and rum, are visually highly similar, making manual differentiation extremely difficult;

2.Chemical testing destroys the original form of the stains, preventing physical evidence from being repeatedly verified;

3.Ordinary spectrometer equipment can only perform single-point acquisition, rendering the detection of large-area traces time-consuming;

4.Interferences such as paper color differences and stain air-drying/oxidation hamper judgment, leaving a lack of standardized data support.

Hyperspectral imaging relies on its advantages of non-contact, full-field acquisition of chemical spectral fingerprints to make up for the above deficiencies, becoming a practical optical tool for trace evidence analysis.

II. Core Tasks Undertaken by Hyperspectral Cameras in the Experiment
This study utilized a 400–1000 nm visible-to-near-infrared hyperspectral imaging device, equipped with 204 continuous spectral bands, to fully collect the 3D spectral data cube (2D spatial coordinates + 1D spectral curve) of various stains. The entire acquisition process is standardized and reproducible.

1.Multi-Scenario Standardized Spectral Acquisition
The study simulated real crime scene conditions, uniformly controlling temperature, humidity, light source angle, and stain droplet volume to perform imaging on 9 categories of beverages (papaya juice, coffee, pomegranate juice, orange juice, tea, red wine, whiskey, rum, and brandy):

• Substrates covered white, pink, and brown absorbent paper towels, including both flat and folded forms;
• Acquisition was conducted at 6 time intervals from 0 to 5 hours to capture spectral changes caused by oxidation and evaporation;
• A 1.5×1.5 cm reflection-free ROI (Region of Interest) was designated for each stain, eliminating edge interference pixels to ensure data purity.

Before shooting, white board reflection calibration and dark current calibration were completed to uniformly convert data into standardized reflectance curves. This eliminated numerical deviations caused by equipment and illumination, allowing horizontal comparisons of spectral differences among different samples.

2.Capturing Invisible Chemical Spectral Characteristics
Every imaging pixel generates an exclusive spectral curve, recording reflection values at different wavelengths to unlock chemical differences unrecognizable to the human eye:

• Red wine and pomegranate juice are rich in anthocyanins, resulting in lower reflection values in the 400–700 nm visible light range;
• Coffee and tea polyphenols exhibit exclusive absorption characteristics in the blue-green bands;
• The distinguishing signals for whiskey and rum are concentrated in the 750–900 nm near-infrared range. It is precisely the fine data from hundreds of continuous bands that differentiate two types of beverage stains that look highly identical to the eye.

3.Constructing a Complete Standardized Evidence Dataset
The equipment collected over 120,000 raw spectral samples. After cleaning noise and duplicate pixels, more than 110,000 valid spectral samples were formed, with each sample corresponding to 204-dimensional spectral features, serving as the original data foundation for subsequent band selection and deep learning training. Sufficient, multi-variable spectral samples support the algorithm in learning stable spectral patterns of stains under different substrates and air-drying durations, reducing recognition fluctuations caused by the environment.

III. Supporting Algorithm Flow: Dimensionality Reduction + Multi-Model Comparison to Enhance Identification Stability
Raw hyperspectral data has a large number of bands, and redundancy exists between these bands. Direct modeling would increase the computational burden, so the study introduced a two-layer data optimization scheme.

1.ANOVA Band Selection to Retain Effective Discriminative Bands
The ANOVA F-test was employed to calculate the discriminative capacity of each band (inter-class variance / intra-class variance), ranking the 204 bands by their differentiation performance. Through 5-fold cross-validation testing of the recognition performance corresponding to different numbers of bands, 162 highly discriminative bands were ultimately selected for modeling, which reduced the scale of data computation while maintaining recognition performance.

2.Comparative Verification of Four Types of Deep Learning Models
Based on the selected 162-dimensional spectral data, four mainstream network models—MLP, 1D-CNN, LSTM, and CNN-LSTM hybrid network—were trained with unified hyperparameters. Precision, recall, F1-score, and overall accuracy were uniformly adopted for evaluation:

• MLP (Multilayer Perceptron) fully connected network: Displayed stable performance in fitting global spectral features, outperforming other models in experimental classification;
  1D-CNN: Excelled at capturing local spectral variations in adjacent bands;
• LSTM: Suited for sequential feature learning of continuous spectral bands;
• CNN-LSTM hybrid structure: Integrated local and sequential features, but due to the influence of data volume, its recognition performance was inferior to single networks.
• Concurrently, paired t-tests and McNemar's tests verified that the performance differences between models possessed statistical significance, ruling out fluctuation interference caused by random sampling.

3.Spectral Patterns Corroborating the Causes of Identification Confusion
The misclassification logic of the models can be intuitively seen from the spectral curves: the curves of coffee and tea overlap heavily in the 400–550 nm range, and whiskey and rum show only subtle differences in the near-infrared range. These two pairs of samples are prone to mutual misjudgment, which is consistent with the model's confusion matrix results, proving that spectral data can explain recognition errors and enhance the interpretability of identification results.

IV. Core Application Value of Hyperspectral Imaging in Forensic Evidence
1.Complete preservation of physical evidence: The entire process requires no reagents and involves no contact with stains. After detection, samples can be preserved for other identifications, conforming to judicial physical evidence specifications;

2.Enhanced investigation efficiency: A single imaging session collects the full spectrum of an entire stain area without the need for point-by-point sampling, making it suitable for rapid on-site screening;

3.Differentiation of visually similar traces: Hundreds of continuous bands capture subtle optical differences in organic matter and pigments, solving visual differentiation dilemmas;

4.Strong environmental adaptability: Experiments verified that the equipment can adapt to stains with different background colors, folds, and air-drying states, with future potential to expand to fabrics, credentials, skin, and other substrates;

5.Digital retention of data: Spectral values can be exported and archived, gradually building a trace spectral library to achieve intelligent sample comparison.

Product Recommendation: FigSpecFS-23 Imaging Hyperspectral Camera

• Image Resolution: 1920*1920
• Spectral Range: 400-1000nm
• Spectral Resolution (FWHM): 2.5nm
• Number of Spectral Channels: 1200

Disclaimer
This content is compiled and edited based on public academic literature under Nature (Literature Link: https://www.nature.com/articles/s41598-026-49928-8). It is intended solely for industry technical discussion and popular science learning, does not constitute any commercial commitment, and cannot serve as a basis for investment reference. The experimental data and conclusions listed in the text may be interfered with by multiple variables such as the test environment, sample individual differences, and model construction schemes. If applied in actual scenarios, relevant effects must be verified through independent testing combined with specific use cases.