Automatic filtering and classification model construction of multi-beam data based on improved convolutional neural network

In this paper, an automatic filtering and classification model of multi-beam data based on improved convolutional neural network (CNN) is proposed. Aiming at the problems of noise pollution, multi-modal feature coupling and poor adaptability to complex terrain in Multibeam Echo Sounder (MBES) data, a multi-scale feature extraction module is designed, which combines multi-scale hole convolution and channel-space mixed attention mechanism to enhance the adaptability to complex terrain and extract multi-scale terrain features. At the same time, a dynamic threshold generation sub-network and an edge perception loss function based on residual connection are constructed to realize adaptive filtering and protect terrain details. Furthermore, a joint filter-classification dual-task learning framework is proposed, which achieves collaborative optimization of both tasks through shared feature layers and a dynamic weight adjustment mechanism. Experimental results show that the proposed model outperforms traditional and deep learning baseline methods in filtering tasks, achieving a lower mean squared error (MSE = 1.2×10−3), higher structural similarity (SSIM = 0.927), and better edge preservation index (EPI = 0.89). In classification tasks, the model achieves an average accuracy of 91.4%, a class-weighted F1 score of 0.893, a seamount recall rate of 87.6%, and a shipwreck false alarm rate as low as 4.2%, demonstrating significant advantages in recognizing and classifying complex terrains. Ablation studies further validate the effective contribution of each module to overall performance, confirming the model’s efficiency and accuracy in multibeam data processing.