Beyond linear: Hybrid neuron network for coherent noise suppression

Seismic data denoising is a crucial step in seismic exploration. Deep learning has emerged as a powerful tool for suppressing coherent noise in seismic data. Although most studies focus on designing complex network architectures, we investigate the benefits of neuronal diversity from a neuronal perspective. First, we introduce the construction of quadratic neurons and determine that they achieve a faster convergence rate than traditional linear neurons. Based on this, we develop a hybrid neuron (HN) block that integrates linear and quadratic neurons into a unified module, leveraging the strengths of both. Using the HN block, we build an HN network, HN-Net. We apply HN-Net to suppress coherent seismic noise and compare its performance with denoising convolutional neural network (DnCNN), U-Net, and f-k filtering on synthetic and field data. Experimental results indicate that HN-Net achieves competitive denoising performance with better generalization. Moreover, HN-Net reduces the number of parameters by approximately 30% compared with DnCNN and 80% compared with U-Net. Our findings highlight a novel approach to deep learning, demonstrating the potential of neuronal diversity in neural networks.