A Fish-Counting Method Using Fusion of Spatial Sensing and Temporal Information

Abstract

In modern aquaculture, accurate and efficient fish counting is crucial for the optimization of resource management and the enhancement of production profitability. Acoustic methods, known for their low energy consumption and extensive detection range, are widely utilized for underwater fish counting. However, traditional acoustic echo methods heavily rely on prior knowledge of fish schools and specific distribution models, leading to complexity and limited adaptability in practical applications. This paper introduces a fish-counting approach that integrates spatial sensing with temporal information. Initially, a spatial sensing matrix is constructed using ultrasonic Frequency-Modulated Continuous Wave (FMCW) technology, which facilitates the extraction of multidimensional features from fish echoes and reduces reliance on prior knowledge of fish schools. Subsequently, temporal information is extracted from echo signals using a Long Short-Term Memory (LSTM) network model, preventing missed detections caused by obstructions in single fish echoes during echo sessions. Finally, by fusing spatial and temporal feature information and employing a data-driven approach, we achieve fish counting while avoiding potential issues arising from improper selection of statistical distribution models. Tests on real fish datasets show that our proposed method consistently outperforms conventional statistical echo methods across all metrics, demonstrating its effectiveness in accurate fish counting.