Abstract: Real-time estimation of fish biomass at the pond-wide scale is critical for precision aquaculture management. However, the low water transparency commonly observed in pond aquaculture environments severely limits the effectiveness of vision-based fish detection methods. To address this challenge, this study proposes a dynamic imaging sonar–based noise-robust fish biomass estimation method tailored for pond aquaculture scenarios. Dynamic sonar image data were collected using an imaging sonar system (Lowrance HDS-9) under typical pond conditions, and a benchmark dataset was constructed for model training, calibration, and validation. A frequency-dynamic convolution (FDConv) module was embedded into the backbone of the YOLOv11n architecture to enable multi-band sonar feature modeling and noise suppression. Fish counting was performed through an integrated pipeline consisting of detection, tracking, and region-based fusion, and the proposed method was further validated in real pond environments. In the model evaluation stage, the proposed method achieved favorable detection performance, with a precision (P) of 0.886, an F1-score of 0.891, and an mAP@0.5 of 0.897, representing improvements of 2.5%, 1.2%, and 1.8%, respectively, over the baseline model. In field experiments conducted at typical pond observation points, the frame-level counting deviation ranged from 10.13% to 17.65%, with an average deviation of 13.98%, while the overall counting deviation was 15.5%. The results demonstrate that the proposed method enables stable perception and reliable estimation of fish abundance in turbid pond waters, achieving effective pond-scale fish population statistics under visually degraded conditions. This study provides a practical solution for fish biomass monitoring in precision pond aquaculture management.