Comparative analysis of hyperspectral and near-infrared spectroscopy for bloodstain deposition time estimation

Bloodstains are a prevalent and critical type of forensic evidence at crime scenes. Accurate determination of bloodstain age is essential for crime resolution, and non-destructive spectral methods are instrumental in this process. While extensive research has established the practicality of hyperspectral imaging (HSI) in specific forensic contexts, limited studies have explored near-infrared (NIR) spectroscopy. Owing to its superior penetration capabilities and high sensitivity, NIR holds promise in addressing certain limitations of HSI. This study aims to assess the applicability of NIR spectroscopy for bloodstain age estimation in forensic contexts and to compare its efficacy with HSI.

Bloodstains were aged on various substrates over a 60 day period, with periodic analyses conducted using both spectral methods. Chemometric analysis of the spectral data was performed following SNV preprocessing and application of different regression algorithms. First, linear regression analysis was utilized to determine the effect of material on bloodstain deposition. Under the premise of distinguishing materials, partial least squares (PLS) regression was employed to extract eight latent variables from HSI and NIR spectral data for regression prediction. However, the prediction performance was suboptimal. To address this, polynomial features were introduced into the PLS regression algorithm to capture the nonlinear relationships in the spectral data, and the improved model significantly enhanced the prediction performance. Furthermore, PLS polynomial regression was applied to predict homologous data, and the results also demonstrated favorable performance. Finally, to optimize the prediction accuracy of multimodal data, a multilayer perceptron (MLP) was introduced for regression prediction through multimodal data fusion, further improving the overall performance of the model. Finally, predictive performance was evaluated across models, emphasizing their specific strengths. For homologous data fusion, comparable root mean square errors of prediction (RMSEP) were achieved for HSI and NIR spectra, at 8.35 and 8.15 days, respectively. Similar RMSEP values were observed in multimodal data fusion, and the accuracy of both low-level and intermediate-level fusion methods was evaluated.

HSI and NIR spectroscopy each provide unique advantages in bloodstain detection. Data fusion of these methods helps mitigate external influences, enhancing the approach's general applicability. This integrated method facilitates rapid estimation of bloodstain age at crime scenes, aiding in crime timeline determination and presenting valuable potential for forensic applications.