- Review
Reliable fish feeding assessment is essential for intelligent aquaculture, yet classification accuracy alone does not establish operational reliability. This structured critical review examines multimodal deep learning approaches, including selected unimodal deployment comparators, through five appraisal dimensions: label validity, data independence and adaptation, robustness to degraded or unavailable sensing, predictive reliability and calibration, and deployment and feeding-management evidence. A key concern is feature–label coupling, where motion-derived labels and model inputs share optical-flow representations, potentially limiting the biological interpretation of classification performance. Reviewed studies demonstrate advances in multimodal fusion, corruption testing, missing-modality handling, knowledge distillation, domain adaptation, and computational efficiency. However, inconsistent dataset provenance, evaluation protocols, and measurement boundaries constrain cross-study comparisons and generalization claims. Uncertainty-informed training does not establish calibrated inference-time confidence, while embedded inference benchmarks may exclude preprocessing and other operational costs. Research priorities include acquisition-independent evaluation, robustness testing under sensing failures and distribution shifts, confidence calibration, selective prediction, and end-to-end deployment assessment. Prospective feeding trials with predefined control policies and independently measured production outcomes are needed to establish operational benefits. By distinguishing demonstrated capabilities from insufficiently verified claims, this review provides an evidence-oriented foundation for developing dependable fish feeding-assessment systems.
19 November 2026





