Granitoid Mapping with Convolutional Neural Network and ASTER VNIR-SWIR Data: A Case Study of the Western Junggar Orogen

The Western Junggar Orogen (Xinjiang) is featured by widespread granite intrusions and vast Au-Cu-Mo resource, making it an ideal site to study granitoids and their metallogenic link. Here, we utilized spectral information, remote sensing imaging and statistics, and textural features to select band combinations from ASTER VNIR-SWIR data. These combinations serve as the input layers for convolutional neural networks (AlexNet, VGG16 and GoogLeNet), which are used for remote sensing identification of granitoid lithology, and the results were compared with the Landsat 8 data. We suggest that the AlexNet model can best identify granitoid subtypes in the Western Junggar, with the 9B+T1 band combination being the most accurate (best weighted F1 score: 91.98%; kappa coefficient: 0.84). Landsat 8 images performed poorly, possibly because they have only two SWIR bands. The best lithological mapping results have identified Cu-Au ore-related diorite in the Karamay III intrusion, I-type granite in the Hongshan intrusion related to quartz vein-type and magmatic-hydrothermal gold ores, as well as A-type granite in the Akbasito and Karamay I and II intrusions. Our findings offer detailed spatial distribution characteristics of granite subtypes and provide remote sensing exploration methods for studying polymetallic ore belts in the Central Asian Orogenic Belt.