Environment-adaptive Acoustic–Optical sensing for fish abundance estimation in complex aquaculture waters

Reliable fish abundance estimation is essential for the operation and management of aquaculture systems, but remains challenging under turbidity, bubble interference, target occlusion, and acoustic reverberation. This study develops an Acoustic–Optical fusion sensing framework for robust fish monitoring in complex aquaculture waters. An environment-aware geometric fusion network, termed EAGF-Net, is proposed to integrate optical appearance information with sonar-derived geometric features. The network employs modality-specific feature extraction with an orthogonality constraint, sonar-guided spatial compensation, and adaptive gating to regulate modality contributions according to environmental conditions. Asymmetric cross-attention is further introduced to reduce semantic inconsistency and suppress unreliable acoustic or optical responses. For densely distributed fish, a keypoint-based topological association method uses fish-eye features and geometric constraints to separate overlapping individuals. The resulting observations are incorporated into Kalman-filter-based trajectory management to reduce missed and repeated counts. Fish abundance is then inferred by combining trajectory flux with the effective sensing volume determined by the Acoustic–Optical field of view and survey path. Experiments covering clear water, dense schooling, turbidity, and bubble disturbance show that the proposed fusion network achieves an F-measure of 0.924 and an average counting accuracy of 96.40%. Field-scale tests in eight aquaculture ponds estimate 1,939 fish against a reference count of 1,975, corresponding to a relative error of 1.80%. These results demonstrate the potential of environment-adaptive Acoustic–Optical sensing for non-invasive fish abundance monitoring in aquaculture engineering applications.