Multimodal Mamba and PatchMLP-KAN for dissolved oxygen prediction in aquaculture

Clinical Proteomics · Springer Nature · 2026-10-30 · DOI: 10.1007/s10499-026-02720-x

Аннотация

Conventional dissolved oxygen (DO) forecasting mainly relies on water quality sensor sequences and has difficulty representing the biological responses of cultured fish. This study proposes a multimodal DO forecasting method that integrates fish behavior with water quality time series. Structured behavioral features are extracted from real aquaculture video and aligned with water quality observations by timestamp. A dual-branch encoder separately models behavioral changes and water quality dynamics. CombinedCrossMamba conditionally introduces information from one modality into the state update process of the other modality and controls the strength of cross-modal information injection through a gating mechanism. A PatchMLP-KAN prediction head is further constructed to jointly represent local temporal patterns and nonlinear relationships among variables. On the current single RAS dataset, the proposed method reduces RMSE by 14.29%, 23.20%, and 31.57% relative to S-Mamba, MulT, and TimeMixer, respectively, and increases R 2 by 0.0205, 0.0634, and 0.0736. Existing ablation results indicate that the performance gains mainly arise from cross-modal state interaction, hierarchical temporal modeling, and the patch-based prediction head design, whereas KAN provides an incremental improvement in nonlinear intervariable modeling. Additionally, the 5-, 10-, and 20-step forecasting results show relatively stable predictive performance across the three horizons. Overall, under the current dataset and experimental conditions, integrating fish behavior information with water quality time series improves DO forecasting performance; however, its cross-scenario generalizability and practical value require further validation under a broader range of aquaculture conditions.

Читать у издателя →

Похожие статьи