Typical farming behaviors recognition in aquaculture using an improved VMamba approach

Accurate and real-time recognition of typical aquaculture farming behaviors, such as “Inspection”, “ApplyMedication”, and “Dead fish retrieval”, is essential for the development of intelligent aquatic product traceability systems. Existing manual recording methods are inefficient and lack reliability. To address this issue, a farming behavior recognition model FBR-Mamba (Farming Behavior Recognition Mamba) is proposed based on an improved VMamba, which can automatic identification of 3 representative behaviors from surveillance video. A motion feature extraction module, ME-Former, is introduced by integrating the STMEM and CGFormer to generate motion feature maps. Additionally, the IDConv module is incorporated into the original VMamba, allowing large-kernel convolution while preserving computational efficiency and parameter economy. An EMA module is further employed to enhance the extraction of informative motion features, improving recognition accuracy. Experimental results demonstrate that the proposed FBR-Mamba achieves a Top-1 accuracy of 97.28 %, representing a 2.53 % improvement over the baseline VMamba model. The proposed approach provides an effective solution for automatic behavior monitoring in aquaculture, contributing to intelligent traceability and farming management.