Improving 3D seismic resolution in the Loess Plateau: Leveraging 2D crooked-line gully survey through weak supervision

Shallow, porous loess layers on the Loess Plateau significantly attenuate seismic waves, particularly at higher frequencies, reducing seismic resolution. The plateau’s distinct, gully-crossed topography features deep gullies, often bare or with thinner loess due to consistent rainfall erosion. These gullies are ideal for 2D crooked-line seismic surveys, which achieve superior resolutions with minimal loess interference compared to traditional 3D data acquisition methods. To enhance 3D data resolution, we employ a novel approach using a cyclegenerative adversarial network (CycleGAN) under weak supervision. Initially, we employ CycleGAN to establish cycle mappings between low-resolution 3D data through conventional swath processing and high-resolution 2D crooked-line data. The CycleGAN then transfers the high-frequency features learned from the 2D data to enhance the resolution of the 3D data. To ensure effective resolution enhancement, this approach is supported by a carefully engineered bidirectional cycle structure, tailored loss functions, and targeted data preprocessing techniques, each specifically designed to navigate the complex resolution challenges posed by loess. Both synthetic and real data experiments demonstrate that our network effectively captures the high-resolution characteristics of the 2D data, significantly improving the fidelity and resolution of the 3D data from the Loess Plateau and marking a considerable advancement overzero-phase spiking deconvolution.