Automatic Lithology Classification Based on Deep Features Using Dual Polarization SAR Images

The lithology classification method based on pixel primitives, polarimetric synthetic aperture radar (SAR) data and traditional machine learning algorithm is easy to be affected by the inherent speckle noise, and the accuracy is not high. In order to reduce the effect of image noise, the neighborhood of large-scale pixels is considered as the primitive to characterize the spatial aggregation characteristics of surface geological units and the corresponding lithologic semantic information. Using GaoFen-3 dual polarization data, the polarization decomposition is carried out first, and a 3-channel color composite image is constructed as the input data of the subsequent model. Then, the deep convolutional neural network (DCNN) based migration learning method is used to extract the effective deep feature representation, so as to realize the automatic lithology classification under 5 m and 15 m spatial resolution conditions. The experiment results show that based on different resolution data and different DCNN algorithms, the total accuracy of automatic lithology classification is greater than 80%, and the highest accuracy is 91%. Generally, based on large-scale pixel neighborhood and DCNN migration learning method, high-precision lithology classification based on SAR data can be realized. The lithology remote sensing dataset based on dual polarization SAR created in this paper can also be used as the benchmark of lithology classification based on artificial intelligence. © 2022 China University of Geosciences. All rights reserved.