Popular deep learning full waveform inversion (DL-FWI) methods are gaining better performance on the benchmark OpenFWI dataset, which provides \(\varvec{70 \times 70}\) velocity models corresponding to \(\varvec{700\,\textrm{m} \times 700\,\textrm{m}}\) profiles. In this paper, we create a new dataset named OpenFWI-140 by increasing the depth to \(\varvec{1400\,\textrm{m}}\) . To handle deep and complex strata, we propose a Fusion vision Transformer Enhanced network for Full Waveform Inversion (FuTE-FWI) with two techniques. First, we introduce a Transformer module between the CNN-based seismic feature extractor and the velocity model reconstructor. The multiheaded self-attention layers of the module can effectively capture the effective waves in complex seismic data. Second, we encode seismic data using residual network instead of directly dividing the data into flattened 2D patches. This technique complements the lack of obvious correlation between seismic data and velocity model. Our FuTE-FWI is compared with InversionNet and VelocityGAN, which use additional downsampling and upsampling modules to cope with the new dataset. Results show that: 1) FuTE-FWI has a significant improvement in terms of both pixel-wise evaluation metrics and structural similarity metrics, especially in deep-layered and complex dataset; and 2) FuTE-FWI exhibits better accuracy in capturing stratal information, including geologic interface and subsurface structure. This advantage becomes increasingly pronounced as the depth increases. The source code is available at https://github.com/palemoons/fute-fwi.
