HSI-NET: A Novel Machine Learning Framework for High-Resolution Spectral Data Interpretation

Abstract: High dimensionality, baseline drift, noise and
instrument-dependent variability make it difficult to
properly assess high-resolution spectral data, despite their
widespread usage in analytical, biological, and chemical
applications. Conventional machine learning techniques
and traditional chemometric methods frequently depend on
hand-crafted features or linear assumptions, which limit
their resilience and capacity for extension. This paper
provides a novel machine learning framework, HSI-NET
(High Spectral Interpretation Network), for interpreting
high-resolution spectral data that utilises deep
convolutional feature learning, a one-dimensional to twodimensional
transformation, and resilient spectral
representations. To stabilise spectral representations, the
preprocessing pipeline contains Savitzky-Golay smoothing,
intensity normalisation, and asymmetric least baseline
correction. Convolutional neural networks are used to
effectively learn spatial characteristics by adaptively
reconstructing one-dimensional spectra into twodimensional
representations. Within a single training and
assessment approach, the system handles both onedimensional
and two-dimensional CNN architectures. HSINET
framework consistently outperforms traditional
analytical, classical machine learning, and deep learning
baselines in terms of accuracy, precision, FDR and FNR.
The outcomes support the suggested procedure’s
robustness, portability, and capability for generalisation,
which render it suitable for a number of high-resolution
spectrum analysis applications.