Spatial classification model of port facilities and energy reserve prediction based on deep learning for port management―A case study of Ningbo

Abstract

Port facilities and energy storage capacity significantly affect maritime logistics efficiency and supply chain security, necessitating accurate and timely port facility information. However, unavailable real-time open port data complicate effective quantitative evaluations of port development along the 21st-Century Maritime Silk Road. This research addresses these issues by combining deep learning with remote sensing, using image data from key ports. A method was proposed to classify port facilities and energy reserve information spatially. A multi-classification framework using U-Net semantic segmentation was developed to segment key facilities in remote sensing data sets. The You Only Look Once v8 (YOLOv8) model was applied to locate oil tanks within Ningbo Port. The actual roof area of oil tanks was then extracted using a deep learning model, facilitating statistical analysis and comparative studies with other major oil and gas ports. Additionally, the real-time remote sensing image index calculated oil tank heights based on shadow lengths for capacity measurement of floating roof tanks. Experimental results showed a pixel accuracy of 90% and an intersection over union of 84% for oil tank region extraction, with an oil tank recognition model achieving a mean average precision of 98.9%. Compared with traditional Hough transform methods, the average absolute error, average relative error, and standard deviation for tank roof area calculations were reduced by 229.18 m², 4.6%, and 81%, respectively. This framework effectively determined the number of oil tanks in various ports, enabling real-time reserve detection and providing a data foundation for energy port management and resilience research.