The environment for seismic exploration is becoming increasingly harsh, and the collected seismic data is more complex. Conventional denoising methods lack flexibility in complex geological settings. Although deep learning-based approaches such as U-Net and its enhanced variant, MPUNet, offer significant benefits, they still face challenges such as insufficient feature extraction accuracy due to fixed convolutional kernel configurations. To address this issue, this study proposes the LNN-MPUNet model, which integrates a Lagrange Neural Network (LNN) with a Multi-stage Progressive U-Net (MPUNet). In this framework, LNN is embedded into the Channel Attention Block (CAB) of MPUNet. The model improves the extraction of effective seismic signal features via the integration of Lagrangian learning and adaptive convolutional kernel modification. Simultaneously, adding physical limitations improves multi-stage denoising’s stability. Experimental results on both synthetic and real seismic data demonstrate that, compared to time-frequency peak filtering (TFPF), conventional UNet, U-Net with residual dense blocks (RDBUNet), MIRNet, and MPUNet, LNN-MPUNet effectively suppresses noise while preserving useful signals under complex noise conditions.