Determination of surface film thickness of heavy fuel oil using hyperspectral imaging and deep neural networks

Abstract

ABSTRACT The determination of surface oil film thickness is essential to safeguard the coastal water quality in major cities globally, particularly during the incidents of oil spills. The spilled oil film is typically very thin of the order of millimeters or less and thus the thickness quantification is very challenging. This study develops a laboratory approach for the thickness estimation using hyperspectral imaging combined with Deep Neural Networks for the image data analysis. Pool experiments were conducted in stagnant seawater with floating oil films of various thicknesses. Hyperspectral imaging was performed, and the images were augmented via a pixel extraction method. The data were then analyzed using two developed models of Dense Artificial Neural Network (DANN) and Convolutional Neural Network (CNN) to predict the thickness of the surface oil film. The results showed that both models managed to produce reasonably accurate predictions with a relatively high coefficient of determination of 0.87 and 0.95, respectively. Comparatively, the CNN model had overall better results by making use of the spatial information of surrounding pixels.