An Application Case of CHNSpec FS-13 Hyperspectral Camera in Non-destructive Detection of Amino Acids in Live Fish

A study published in "Food Research International" utilized visible/near-infrared hyperspectral imaging technology to achieve non-destructive prediction of muscle amino acid content in live common carp. This study was jointly completed by Shanghai Ocean University, Chinese Academy of Fishery Sciences, and other units. The FS-13 hyperspectral camera (FigSpec FS-13) provided by CHNSpec Technology was used as the core detection equipment. Xiajun Qi, an engineer from CHNSpec Technology, deeply participated in the research, providing a new technical path for the real-time evaluation of the nutritional quality of live fish.

I. Research Background and Detection Requirements
The amino acid composition of fish meat is an important indicator to measure its nutritional value and commercial value. Although traditional detection methods (such as high-performance liquid chromatography) are accurate, they are destructive—the fish cannot be sold further or used for selective breeding after detection. For application scenarios that require maintaining the live status of fish, such as precision feeding, nutritional grading, and parental selection, the industry has long lacked a rapid, non-destructive, and online detection tool.

The starting point of this study lies in: can fish scales serve as a "window" for spectral signals? Can near-infrared light penetrate fish scales and skin, carrying chemical composition information from the muscle back to the detector? If feasible, it will fundamentally solve the problem of live fish nutrition detection.

II. Experimental Protocol and Core Equipment
The research team collected two populations of common carp from different years and different weight ranges, totaling 481 live fish. For each fish, it was first briefly anesthetized using MS222 anesthetic, and the surface of the scales in the dorsal fin region was gently dried with absorbent paper. Then, the CHNSpec Technology FS-13 hyperspectral camera (spectral range 400-1000 nm, spectral resolution 2.5 nm) was used to acquire hyperspectral images of the dorsal fin region of the scales. The region of interest for each sample covered 200×200 pixels, with each pixel containing spectral information across 300 bands.

Subsequently, sampling was performed at the corresponding dorsal muscle site, and the actual content of 17 amino acids was determined by high-performance liquid chromatography for modeling and validation.

III. Model Construction and Prediction Effects
The researchers compared five models: Partial Least Squares Regression (PLSR), Least Squares Support Vector Machine (LS-SVM), Extreme Learning Machine (ELM), Random Forest (RF), and Backpropagation Artificial Neural Network (BP-ANN). Modeling was conducted using full-band spectral signals (400-1000 nm), and the R² values of different models on the training and prediction sets were generally higher than 0.95.

Among them, the BP-ANN model showed relatively stable prediction effects for most amino acids. In the independent validation set (181 fish from different years and different farming environments), the validation R² values of the BP-ANN model all exceeded 0.777. The validation R² for the three highest-content amino acids—glutamic acid, aspartic acid, and lysine—reached 0.848, 0.858, and 0.858, respectively. The study also found that after replacing the full bands with characteristic wavelengths (selected by the CARS algorithm), the improvement in prediction accuracy was limited (average R² increased by about 0.013), indicating that amino acid-related spectral information is widely distributed.

IV. Key Factors Affecting Accuracy
The study systematically evaluated the impact of six factors on prediction accuracy, and the results showed that: sample population heterogeneity was the most significant factor affecting accuracy. When the model was applied to independent populations from different years and weights, the average R² decreased by about 0.182. This may be related to differences in the distribution of amino acid content between the two populations (e.g., the median of most amino acids in the first population was significantly higher than that in the second population). Despite this, the BP-ANN model still maintained acceptable accuracy (R² > 0.777) in heterogeneous populations.

In contrast, the model type, amino acid type, wavelength selection method, fish body weight, and body length had less impact on accuracy (average R² variation less than 0.103). For example, after dividing the fish into upper, middle, and lower groups according to body weight, the average difference in R² for the BP-ANN model was only 0.076 (when using characteristic wavelengths). This indicates that the spectral signal is mainly driven by the biochemical composition of the muscle, rather than simple physical size scattering effects.

In terms of characteristic wavelengths, the CARS algorithm selected sensitive bands for glutamic acid and lysine concentrated in 516-584 nm, 707-738 nm, 828-834 nm, and 939-1032 nm. These regions are associated with the overtones and combination frequencies of C-H bonds, O-H bonds, and N-H bonds, validating the feasibility of near-infrared light interacting with amino acid molecules in the muscle after penetrating the scales.

V. Spatial Distribution and Application Value
Utilizing the spectral information of each pixel from the FS-13 hyperspectral camera, the research team mapped the heatmap distribution of the total amino acid content throughout the live fish body. The results showed that: the total amino acid content in the muscle of the lower jaw, pectoral fin, and abdomen was relatively high, while that in the dorsal fin region and tail was relatively low. This distribution matches the functional differences in muscle fiber types (red muscle and white muscle) across different parts—the pectoral fin and abdomen are dominated by slow-twitch oxidative red muscle, where protein metabolism is more active. This heatmap can provide a visual reference for consumers to select parts with high nutritional value.

The CHNSpec FS-13 hyperspectral camera paired with deep learning algorithms successfully broke through the technical bottleneck of non-destructive detection of amino acids in live aquatic products, providing a lightweight, practical detection tool for precision aquaculture and high-quality aquatic product screening. In the future, with the continuous improvement of the model database and the development of portable equipment, this solution can be further promoted to a variety of freshwater and marine fish species, helping the aquatic industry upgrade toward intelligence, standardization, and nutritional visualization.

Product Recommendation: FigSpecFS-13 Hyperspectral Camera (Line Scan)

• Spectral Range: 400-1000nm
• Spectral Resolution: 2.5nm
• Spectral Bands: 1200
• Spatial Pixels: 1920