Automatic identification method for subaqueous gravity flow facies in core images using convolutional neural networks

Traditional methods for identifying subaqueous gravity flow facies in cores require extensive prior knowledge and a solid foundation in geosciences. These methods are often hindered by uncertainties related to human and environmental factors. This paper proposes an automatic identification method for subaqueous gravity flow facies in core images, which consists of assessing image validity and identifying gravity flow facies. The validity assessment, based on practical criteria, is essential for accurate identification. Experiments have demonstrated that this assessment can effectively identify blurred images, as well as those that are overexposed or underexposed, and can determine whether images represent slump, debris flow, or turbidite facies. For the identification process, transfer learning is employed using GoogLeNet, ResNet18, ShuffleNetV2, and MobileNetV3 as base models. These models are enhanced through the concept of ensemble learning, resulting in an accuracy of 88.8% on the test set. Based on these findings, a WeChat Mini Program, Core Image Gravity Flow Facies Identification V1.0, was developed and deployed.