MDBTabNet: An Enhanced TabNet with Feature Enhancement and Biological Consistency for Intelligent Abalone Plumpness Classification

Abalone is a high-value mollusk whose aquaculture relies on industrial production. Quality assessment and selective breeding are important for industry efficiency, with plumpness as a key indicator. However, available public datasets are not designed for plumpness, and conventional TabNet models have limited capacity for geometric feature extraction and nonlinear interaction modeling. We propose a two-stage strategy: first, we integrate SAM-extracted morphological indices with field measurements to construct a multi-source dataset; second, we develop MDBTabNet, which combines morphological feature enhancement (MFE) and dynamic feature interaction (DFI). It uses an explicit geometric computation layer to fuse morphological and raw features, and integrates multi-head self-attention with local convolutions to jointly model global dependencies and local interactions, improving representation of complex nonlinear relationships. Then, a biological consistency (BC) auxiliary regression branch predicts wet weight through multi-task learning, which serves as mild regularization. Experiments indicate that MDBTabNet outperforms baseline models in accuracy, F1, and AUC, and shows the lowest cross-validation standard deviation in F1. The results suggest that shell morphology combined with sex carries moderate but non-negligible predictive signal, supporting feasible preliminary screening based on phenotypic and routine biological attributes in abalone aquaculture.