Geological Lithology Semantic Segmentation Based on Deep Learning Method

Remote sensing data has been widely used in geological researches. The geological researching tasks such as lithological and soil type mapping can be achieved easily by working on remote sensing images with classification approaches. Deep learning has made remarkable achievements in processing remote sensing data as it can usually indicate the better accuracy than traditional methods. This chapter proposed the deep learning models for semantic segmentation of remote sensing images for geological mapping. These models include Mobilenet-based U-Net, Mobilenet-based PSPNet and SegNet. The experiment was conducted on the randomly cropped datasets and partially overlapped cropped datasets made for the study site of Suiyang Town, China. The result indicates that the pixel segmentation accuracies of the three models are all higher than 95% on the randomly cropped dataset, and about 60% on the partially overlapped cropped dataset. The Mobilenet-based PSPNet indicates the best performance on the non-overlapped cropped dataset. The proposed models can successfully tackle the low accuracy issue of traditional methods and can achieve the more ideal performance for geological mapping tasks.