抄録
This study presents a practical approach for monitoring pond bottom conditions in kuruma shrimp (Marsupenaeus japonicus) aquaculture by combining an autonomous boat, an underwater camera, and deep learning-based image classification. Underwater videos were collected during autonomous navigation in a shrimp pond located in Amakusa, Japan. From the videos, 128×128 pixel image patches were manually extracted and labeled into four categories based on visual features: (1) gravel, (2) residual feed pellets, (3) white fungal-like matter, and (4) sludge-like sediment. A convolutional neural network (CNN) was developed and trained using these labeled patches. The trained model was then applied to classify pond bottom conditions by segmenting video frames into non-overlapping 128×128 patches and assigning class labels to each. The resulting spatial distribution maps revealed that residual feed tended to accumulate in low-flow regions, while sludge-like sediment was often found behind aerators. These spatial patterns indicated localized organic matter accumulation and environmental degradation, providing useful information for pond management. The proposed approach offers a low-cost and efficient tool for visualizing bottom conditions and supporting decision-making in aquaculture operations.
https://ror.org/01c697313 
