Lithological Classification Based on Large-Scale Pixel Neighborhood and VGGnet-Based Transfer Learning

Remote sensing technology can provide a powerful technical reference for different geological prospecting work. It is an effective method to quickly indicate the engineering geological situation of areas with limited investigation due to poor traffic conditions, which can provide guidance for the local engineering projects. In view of the problems such as insufficient samples, difficulty in intelligent interpretation, and difficulty in selecting appropriate scale in the current stage of intelligent extraction of rock mass, this study proposed a methodology that utilized large-scale pixel neighborhood data based on spectral and spatial characteristic information of different lithologies in remote sensing images and used VGG16 convolutional neural network to pre-train on ImageNet. With the trained model and the corresponding initialization parameters, the model was fine-tuned by using the constructed remote sensing data of the rock mass. The parameters of the network model were constantly adjusted, and the optimal rock mass classification model were successfully obtained. The experimental results showed that the overall classification accuracy in the 441 km2 study area reached 85%, which effectively improves the accuracy and efficiency of rock mass interpretation based on remote sensing.