Abstract
<jats:p>Shrimp farming constitutes a strategic productive activity for Ecuador; however, monitoring of shrimp ponds still relies on manual observations, fragmented records, and qualitative criteria that hinder the early detection of visual conditions of operational interest. In this context, the present study proposes a prototype adaptive intelligent system to support the visual monitoring of shrimp ponds through computer vision. The proposal integrates image capture, visual information processing, organized record storage, and alert visualization in an administrative dashboard. The evaluation was conducted using public datasets employed as experimental approximations to three visual signals of interest: feed accumulation, surface changes, and foam presence. In this work, anomaly detection is understood as the preliminary identification of visual signals associated with possible anomalous conditions or conditions of operational interest. The results show high performance in feed accumulation, with mAP50 values above 0.98 for BOX and MASK; moreover, the bootstrap analysis revealed a statistically significant advantage of YOLOv11m-seg over YOLOv8m-seg for the shrimp class. In surface changes, YOLOv8n and YOLOv11n achieved mAP50 values of 0.884 and 0.885, respectively, with no statistically significant differences in F1-score. In foam presence, the models achieved mAP50 values of 0.767 and 0.758, respectively, and the bootstrap analysis confirmed non-significant differences between architectures, reflecting the visual complexity of this signal. Inference times of 3.9–21.2 ms/img support the use of the prototype for frequent monitoring. It is concluded that the proposed system effectively integrates computer vision, inference services, and alert visualization to support visual inspection, event prioritization, and operational management of shrimp ponds.</jats:p>