Field evaluation of artificial intelligence early warning for bird depredation at carp ponds

Abstract Piscivorous birds such as the great cormorant ( Phalacrocorax carbo ) can cause substantial losses in pond aquaculture, yet continuous on‐site vigilance is difficult under labor shortages. We evaluated whether an artificial intelligence‐based early‐warning system could be feasibly integrated into bird‐depredation management at a commercial common carp ( Cyprinus carpio ) farm under labor‐limited field conditions. Our primary focus was operational feasibility: linking existing surveillance cameras, screen‐captured regions of interest, cloud‐based multimodal inference, and Discord/email notifications into a workflow that could support targeted pond checks. As secondary evaluations, we quantified preliminary classification performance using 100 field screenshots and compared the multimodal generative model with an untuned COCO‐pretrained You Only Look Once version 11 (YOLO11) detector. Four Internet Protocol (IP) cameras continuously monitored the ponds, and a Python application captured predefined regions of interest from the camera‐viewer screen and sent the screenshots to the multimodal generative model via the OpenAI application programming interface (API) for binary classification of bird presence or absence. When birds were detected, the system automatically notified the farmer via Discord or email with the image and the model's textual rationale. During the main early‐morning monitoring period, producer‐reported observations suggested that alerts helped shift monitoring away from continuous video watching toward targeted pond checks. In a preliminary evaluation using 100 field screenshots (50 positive/50 negative), the multimodal generative model (o4‐mini) achieved an accuracy of 0.71, precision of 0.96, and recall of 0.44, while the YOLO11 detector achieved an accuracy of 0.52, precision of 1.00, and recall of 0.04 under low‐light conditions. The average inference cost was 0.002 USD per query, corresponding to 18.6 USD per month at 300 inferences per day. We discuss operational implications of high‐precision/low‐recall behavior and propose year‐round, log‐based monitoring to support parameter tuning and future migration to dedicated edge‐deployed detectors.