The effectiveness of ground penetrating radar (GPR) in identifying and categorizing diseases that occur underground beneath the surfaces of urban roads is investigated in this study. Both 2D and 3D forward modeling use simulation with the GprMax program to show the response characteristics of common cavity illnesses, which facilitates interpretation in practical situations. The cavity morphology classification accuracy is improved to 90.5% by using convolutional neural networks (CNNs), specifically transfer learning with AlexNet. This method outperforms existing approaches even with minimal data. Four primary types are identified from an analysis of 1965 subsurface cavity data: hollow bodies, empty bodies, loose bodies, and water-rich bodies. These categories are important for evaluating road risks such as voids and subsidence. However, it is still difficult to interpret picture features linked to cavity diseases accurately because of a variety of elements, such as anthropogenic, environmental, and geological influences. However, the accurate interpretation and recognition of image features related to cavity diseases remain challenging. Moreover, there are various factors involved in the formation of underground diseases and cavities, including geological and environmental factors, physical and chemical properties of the geotechnical materials, anthropogenic engineering activity and social population or commercial effects.