The Pacific white shrimp (Litopenaeus vannamei) is a globally prominent aquaculture crustacean adaptable to diverse habitats. Identifying common features for predicting production methods, spoilage indicators, and freshness in shrimp is crucial for enhancing quality and consumer satisfaction. However, the influence of environmental factors during growth on aquatic product quality remains unclear.
This study aimed to develop a deep learning model for predicting aquaculture conditions directly from raw Pacific white shrimp. Raw shrimp samples and corresponding aquaculture data (type, salinity, pH, sampling month) were collected from seven farms across three Taiwanese counties. Compact deep learning models with 6 to 12 hidden layers and less than 350k parameters were built and trained using custom Python code based on TensorFlow-Keras libraries. ReLU activation function, Adam optimizer, and mean squared error were employed for training the regression models.
Multispectral imaging (MSS), a non-invasive technique, was used to capture reflection spectra (26 wavelengths between 369490 nm) from 970 raw shrimp samples. All four tissue regions (eye, head, trunk, tail) and their combinations exhibited similar Mean Absolute Percentage Error (MAPE) during model validation, ranging from 0.7% to 12.2%. Subsequently, the best models were tested on data from two new shrimp farms, achieving a MAPE of 21.8±11.5% (best) for the eye-trunk combination and 35.1±17.0% (worst) for the head-eye combination.
Our compact deep learning models effectively predicted aquaculture conditions with acceptable error using raw shrimp data. These findings suggest a potential link between aquaculture conditions and post-harvest product characteristics like shelf life and quality, which is valuable for both producers and consumers. The developed models hold promise for optimizing white shrimp aquaculture practices and promoting sustainable aquatic food production