Recently, the application of deep learning approaches has significantly contributed to the success of soil investigations, particularly in the categorization of soil images. This, in turn, facilitates informed decision-making in geotechnical engineering projects. This paper introduces an efficient soil classification method utilizing transfer learning models based on Convolutional Neural Networks (CNNs) to categorize soil images. Various pre-trained networks, including Alexnet, Vgg16, Googlenet, Resnet50, Mobilenetv2, and Squeezenet, are employed in this study. The highest accuracy, reaching 90.32%, is achieved by Googlenet, while Alexnet follows closely with an accuracy of 89.24%. Conversely, Squeezenet yields the lowest accuracy at 79.56%. This proposed framework has the potential for integration into Internet of Things (IoT) systems, leading to the development of an innovative, portable IoT-based device for soil type recognition at remote geo-sites, catering to the needs of geotechnical engineers. The results obtained highlight that the Googlenet-based framework outperforms others, offering a superior performance in soil classification. This framework serves as a valuable tool for implementing an intelligent soil image classification system, supporting real-time decision-making in geotechnical engineering applications.