Search

Bivalve traceability is a major concern. It is of utmost importance to develop tools that allow providing important information to the consumer, not only on the origin of the product but also on its sustainability and safety, due to the harvest restrictions imposed by regulatory entities. This study evaluated the application of computer vision machine learning technologies for efficiently discriminating cockle harvesting origin based on shell geometric and morphometric analysis, improving the traceability methodologies in these organisms, and highlighting the potential of these low-cost techniques as a reliable traceability tool. Thirty Cerastoderma edule samples were collected along the five locations in Atlantic West and South Portuguese coast with individual images processed using lazysnapping segmentation, spectro-textural-morphological phenotype extraction, and feature selection through hybrid Principal Component Analysis and Neighborhood Component Analysis which resulted in R, a*, b*, entropy, and diameter. Three approaches of traceability models were developed and tested: pre-trained networks (EfficientNet-Bo, ResNet101, MobileNetV2, InceptionV3) with numerical inputs (Approach 1), image-based pre-trained networks (Approach 2), and hybrid deep neural networks of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) (Approach 3). Based on the test results, Approach 3 with GRU-LSTM-BiLSTM sequence exhibited the highest accuracy (96.91%) and sensitivity (96%) among the other thirteen machine learning models, hence, named as BivalveNet. Comparing the attained accuracy from the BivalveNet to other mollusc traceability studies, it was observed that an efficiency close to the attained using standard destructive, time-consuming, and expensive techniques, making BivalveNet a highly advantageous approach for common cockle geographical traceability studies, available for application to other bivalve species. • Bivalves are a valuable marine resource, prone to origin mislabelling and fraud. • A novel approach to depict bivalve harvesting locations developed with state-of-the-art deep learning. • Image feature pre-selection through hybrid PCA-NCA improved model accuracy by removing noise features. • Hybrid deep neural networks had higher predictive accuracy (96.91%) and performed better than pre-trained networks. • The BivalveNet tool provides high- throughput and efficiency at low cost for bivalve traceability. [ABSTRACT FROM AUTHOR]