# Non-Destructive Freshness Assessment of Atlantic Salmon (Salmo salar) via Hyperspectral Imaging and an SPA-Enhanced Transformer Framework

## TL;DR

This paper introduces a non-destructive method using hyperspectral imaging and a Transformer model to assess the freshness of Atlantic salmon efficiently and accurately.

## Contribution

The novel framework combines hyperspectral imaging with a Transformer deep learning model enhanced by the Successive Projections Algorithm for freshness assessment.

## Key findings

- The Transformer model with Savitzky-Golay smoothing achieved high accuracy (R² ≥ 0.97) for predicting freshness indicators.
- Using 30 characteristic wavelengths selected by SPA maintained high precision while improving computational efficiency sixfold.
- The study confirms the effectiveness of attention-based deep learning in hyperspectral data analysis for food freshness.

## Abstract

Monitoring the freshness of Salmo salar within cold chain logistics is paramount for ensuring food safety. However, conventional physicochemical and microbiological assays are impeded by inherent limitations, including destructiveness and significant time latency, rendering them inadequate for the real-time, non-invasive inspection demands of modern industry. Here, we present a novel detection framework synergizing hyperspectral imaging (400–1000 nm) with the Transformer deep learning architecture. Through a rigorous comparative analysis of twelve preprocessing protocols and four feature wavelength selection algorithms (Lasso, Genetic Algorithm, Successive Projections Algorithm, and Random Frog), prediction models for Total Volatile Basic Nitrogen (TVB-N) and Total Viable Count (TVC) were established. Furthermore, the capacity of the Transformer to capture long-range spectral dependencies was systematically investigated. Experimental results demonstrate that the model integrating Savitzky-Golay (SG) smoothing with the Transformer yielded optimal performance across the full spectrum, achieving determination coefficients (R2) of 0.9716 and 0.9721 for the Prediction Sets of TVB-N and TVC, respectively. Following the extraction of 30 characteristic wavelengths via the Successive Projections Algorithm (SPA), the streamlined model retained exceptional predictive precision (R2 ≥ 0.95) while enhancing computational efficiency by a factor of approximately six. This study validates the superiority of attention-mechanism-based deep learning algorithms in hyperspectral data analysis. These findings provide a theoretical foundation and technical underpinning for the development of cost-effective, high-efficiency portable multispectral sensors, thereby facilitating the intelligent transformation of the aquatic product supply chain.