*2.1. Network Architecture*

The single-task network used in this paper is a FCRN from Wu et al. [23], and the multi-task network is a hard parameter sharing network developed from the single-task FCRN. The multi-task network structure is shown in Figure 1. In order to better capture the low-frequency characteristics of seismic data, the first convolution layer of FCRN has 16 kernels of size 299 × 1. After the first convolution layer, three residual blocks are stacked, and each residual block is composed of two convolution layers. The first layer and the second layer, respectively, have 16 convolution kernels of size 299 × 1 and 3 × 1. To enable the network to complete multi-task inversion, two output channels are set after the residual block. Each output channel contains two one-dimensional convolution layers, and each one-dimensional convolution layer has a kernel of 3 × 1. The convolution step size is 1, and zero-padding is used to all convolution layers to ensure the same input and output sizes. Rectified linear unit (ReLU) and batch normalization (BN) are introduced into the network to accelerate network training and convergence.

**Figure 1.** Architecture of the multi-task FCRN.
