Intelligent identification of carbonate reservoir cavities based on VQ-VAE feature enhancement and multi-attribute soft label generation

Abstract Accurate and intelligent identification of subsurface cavities in carbonate reservoirs remains challenging due to the absence of reliable labeled data and sensitive seismic attributes (SAs). To overcome these limitations, this paper proposes an intelligent framework that integrates multi-attribute label generation, vector quantized variational autoencoder (VQ-VAE) based feature enhancement, and TransUNet for carbonate cavity identification. Multiple SAs, including root mean square amplitude, envelope, sweetness, and reflection intensity, are combined with geometric constraints to construct weakly supervised labels with confidence masks, which effectively reduce annotation dependence and constrain uncertain regions. The VQ-VAE further generates a new class of data-driven sensitive attributes by mapping seismic data into a latent codebook space, where typical local reflection patterns of cavities are adaptively learned. Guided by soft labels, the enhanced feature volumes and the original seismic data, are then fed into a TransUNet network for cavity prediction. We apply the proposed method to both a 3D physical simulation model and a field carbonate reservoir dataset. Comparative tests demonstrate that the selected VQ-VAE features enhance the expression of cavity reflections, suppress strong reflection interference, thereby improve the depiction accuracy of cavity structures. Furthermore, the identified cavities from field data show high consistency with impedance inversion results, as well as drilling leakage and overflow evidence, confirming the geological reliability and robustness of the approach.