One-Dimensional Rock and Soil Characteristic Parameters Prediction Method Based on SRR

Acquiring precise geologic parameters for obstructed or complex geologic regions poses a difficult task in practical engineering. Current predictions depend on the expertise of engineers, leading to inadequate levels of precision. Therefore, in this study, geotechnical stratigraphic data were transformed into visualization images containing only red information corresponding to R values in RGB images. The generated visualization images were analyzed using a super-resolution convolutional neural network (SRCNN) for prediction and compared with linear interpolation-based prediction methods. Subsequently, a dataset containing 430,000 patches was generated using real geologic data from a specific project, and this dataset was used for SRCNN training to validate its prediction. The results showed that SRCNN yields a peak signal-to-noise ratio (PSNR) of 40.22 dB, exceeding the linear interpolation on the geologic map (39.93 dB). The SRCNN training was successful and outperformed the linear interpolation. The PSNR values of the SRCNN were higher (34.69 dB, 37.68 dB, 38.79 dB, 37.56 dB, and 44.99 dB) compared to linear interpolation (34.53 dB, 37.43 dB, 38.38 dB, 37.29 dB, and 44.31 dB). These findings confirmed the significant potential of the application of super-resolution reconstruction for predicting soil distribution, and this method is expected to yield more precise soil prediction results as the dataset grows.