Accurate estimation of fish biomass is essential for precision feeding, water quality regulation, and health management in pond aquaculture. However, high stocking density and low water transparency significantly degrade the performance of optical vision-based methods, posing a persistent challenge for reliable fish detection and counting.To address this issue, this study proposes a noise-robust fish biomass estimation framework based on dynamic imaging sonar for turbid pond environments. A dataset of sonar image sequences was collected in real aquaculture ponds using a Lowrance HDS-9 system. To improve feature representation under severe noise interference, a Frequency Dynamic Convolution (FDConv) module was integrated into the backbone of a YOLOv11n detector, enabling multi-frequency feature decomposition and effective noise suppression.Based on this enhanced detector, an end-to-end pipeline incorporating detection, multi-object tracking, and spatial region fusion was developed for robust fish counting and biomass estimation. Experimental results demonstrate strong detection performance, achieving a precision of 0.886, an F1-score of 0.891, and an mAP@0.5 of 0.897, outperforming the baseline by 2.5%, 1.2%, and 1.8%, respectively. Field evaluations across multiple pond locations show frame-level detection accuracy ranging from 82.35% to 89.87%, with an average of 86.02%. After multi-region fusion, the overall counting accuracy at the pond scale reached 84.5%.These results demonstrate that the proposed method enables stable perception and quantitative estimation of fish populations under low-visibility conditions, effectively addressing the “invisible fish” problem in turbid aquaculture systems. The framework provides a practical and deployable solution for real-time biomass monitoring, supporting the development of intelligent and precision aquaculture.