Measuring salmon texture – developing methodology based on spectroscopy

Sammendrag

Texture in is a central parameter when evaluating quality of fish or fish muscle. From a consumer point of view this may be recognized by the soft and juicy mouthful of a salmon fillet, or by contrast, the compact and dry chew from an overcooked fillet. In view of production, product and fillet quality, texture will have an impact on product yield and production efficiency. Texture in salmon is affected by several factors, amongst others feed and growth rate during production, pre- and during harvesting handling regimes, and temperature and handling during post-harvest processing and storage. Too soft, or weak texture, in salmon and salmon fillets may cause customer complaints or financial reimbursement claims. From time to time there are also reports of large-scale quality issues related to poor texture. For a long time, the aquaculture industry has asked for tools suitable for assessing textural properties in large scale industrial production. Such tools would be of use and benefit for the seafood production industry, both in terms of improving product quality and consistency, but also in terms of production efficiency and product yield. The main goal of this project was to investigate if techniques based on spectroscopy (specifically hyperspectral imaging and Magnetic Resonance Imaging - MRI) can be applied as tools to measure and predict texture properties in salmon fillets. This project aims to determine if these technologies can non-invasively and in real time measure the texture properties of salmon fillets, which could potentially decrease market claims for texture defects and positively influence the aquaculture industry, including farmers, producers, and exporters of Norwegian salmon. Another objective was to evaluate whether hyperspectral measurements taken on the day of filleting could effectively predict the texture properties of salmon fillets after one week of storage on ice. We conducted a series of experiments to span a meaningful range of the natural variability in texture, which is a product of multiple biological and environmental factors. The experimental design involved quantifying the texture of salmon fillets through both sensory evaluation and instrumental texture measurements, as well as recording the corresponding spectroscopic data (hyperspectral images and MRI). In each experiment, all measurements were conducted on the day of filleting and again after 8 days of storage on ice. Instrumental and sensory analyses across seven experiments confirmed that the fillet texture is not homogeneous, with the Norwegian quality cut (NQC) area consistently firmer and more resistant than the loin. Texture generally softened during ice storage. Tensile and puncture test showed similar trends, though changes were more pronounced in specific measurements areas and groups. Sensory texture remained stable in some groups but softened in others, particularly under stress or delayed handling. The findings highlight the importance of standardized sampling and measurement locations, as well as the influence of pre- and post-mortem factors on fillet quality. Different models based on artificial neural networks were trained to predict both instrumental texture and sensory scores from the hyperspectral data. The analysis reveals a consistent pattern: models perform reasonably well when trained and validated on fillets drawn from the same set of experiments, but their accuracy drops when applied to data from independent experiments. This suggests that while the models can learn relationships between spectral signals and texture measurements, they struggle to make accurate predictions on unseen samples. The limited generalization observed in the models likely reflects the inherent complexity of texture as a quality attribute and the high variability in texture characteristics across experiments, which may contain texture features or ranges of variation not represented in the training data and therefore difficult for the model to handle. Although most models showed poor performance when applied to data from new experiments, some models for predicting instrumental texture measurements and sensory scores demonstrate moderate accuracy. These models may offer value for coarse texture classification (firm versus soft) but are may not deliver sufficiently reliable results across the full range of encounterable variation that would be required for quality grading at an industrial scale. In order to improve the predictive performance and generalization, future research could explore strategies such as large-scale data collection under industrial conditions, the use of alternative spectral ranges, or the integration of both spatial and spectral information for the predictions.