Abstract: Aiming at the problems that satellite images are difficult to obtain geological outcrop data with large slopes, and traditional classification methods cannot effectively use image information leading to geological outcrop section lithology classification accuracy being relatively low, this research obtains high-precision field geological outcrop data with large slope based on UAV remote sensing technology and proposes a multi-scale hybrid feature network model. The results show that the combination of UAV and close photogrammetry is feasible in collecting geological outcrop data. The multi-scale hybrid feature network model can effectively extract the spectral features and spatial features from the multi-spectral images of UAV and realize the high-precision lithology classification of geological outcrops with large slopes. Taking an outcrop in Yuntaishan geopark as an example, the overall classification accuracy of the proposed model can reach 89.91%, and the Kappa coefficient can reach 0.85. The general classification accuracy is nearly 15% higher than traditional machine learning algorithms SVM and MLC, about 10% higher than Inception V3 and ResNet18, and 1.5% higher than Hybrid CNN.

Key words: UAV, geological outcrop, deep learning, lithology classification