Terahertz Spectroscopy Combined With Deep Learning for Predicting the Depth and Duration of Underground Sand Pollution by Crude Oil

Crude oil spills near oilfields, a continuous threat to the surrounding soil security and ecological environment, are attracting increased environmental monitoring and protection. There is an urgent need to improve the existing methods for identifying pollution location and quantifying pollution amount. Recently, deep learning (DL) methods have shown strong ability to extract information in many fields. As the environmental impact of oil exploration rises, development of new artificial intelligence (AI) prediction models of oil spills becomes all the more important. Here, shallow sands in a desert oilfield of northwestern China are used as the contaminated object. We exploit the sensitivity of terahertz (THz) waves to polar molecules and the data processing capabilities of DL to develop a set of hybrid experimental calculations. Using a hybrid experimental-calculation approach, namely, THz spectroscopy in combination with a DL method, we propose a panel of prediction analysis models that provide detailed curves of pollution depth and duration. The technology can not only be used to analyze polluted sand particles but also be used to monitor crude oil spills in real time. After extensive training, the convolutional neural network (CNN) was able to detect whether sand samples were contaminated, with the
$F1$
score reaching 97.90%, and the AUC value was 0.99, indicating high consistency and robustness. The pollution probabilities were analyzed via subsequent pollution simulation experiments, revealing that they predict essential features comparable with experimental results. In the future, this method will be used to predict oil spill pollution in more complex natural environments and from other types of crude oil in order to achieve a wide range of applications.