ICAT-net: a lightweight neural network with optimized coordinate attention and transformer mechanisms for earthquake detection and phase picking

Seismic signal detection is a crucial technology for enhancing the efficiency of earthquake early warning systems. However, existing deep learning-based seismic signal detection models often face limitations in resource-constrained seismic monitoring engineering environments due to the high computational resource demands of the models. To address this issue, this study employs spatial-depth convolution techniques in the downsampling process of seismic signal sequences, effectively minimizing the loss of fine-grained feature information. Concurrently, we leverage the coordinate attention module to enhance the model’s ability to recognize spatial features in seismic signal sequences. To reduce computational costs, we map the keys and values in the transformer architecture to a lower-dimensional subspace, significantly decreasing the demand for computational resources. By employing concatenation operations between the encoder and decoder, the model retains rich contextual information and progressively restores the spatial resolution of the signal during the decoding process. Based on these models, we propose the integration of coordinate attention and transformer network (ICAT-net), an efficient multi-task network designed to simultaneously handle various tasks, including seismic sequence recognition and phase picking. ICAT-net integrates local feature relationships with long-range dependency processing capabilities to meet the requirements of multi-task learning. Experimental results demonstrate that ICAT-net requires only 4.743168G of floating-point operations (FLOPs) and has a parameter count of 0.260755M, while performing excellently in tasks such as seismic waveform detection (DET), P-wave phase picking (Ppk), and S-wave phase picking (Spk). These advantages render ICAT-net particularly suitable for deployment in resource-constrained environments, providing valuable solutions for earthquake monitoring and disaster risk assessment.