Random-Drop Data Augmentation of Deep Convolutional Neural Network for Mineral Prospectivity Mapping
- PDF / 3,013,825 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 4 Downloads / 192 Views
Original Paper
Random-Drop Data Augmentation of Deep Convolutional Neural Network for Mineral Prospectivity Mapping Tong Li,1 Renguang Zuo,1,2 Yihui Xiong,1 and Yong Peng1 Received 2 August 2020; accepted 29 August 2020
Convolutional neural network (CNN) has demonstrated promising performance in classification and prediction in various fields. In this study, a CNN is used for mineral prospectivity mapping (MPM) in the southwestern Fujian Province, China. Two limitations of applying CNNs in MPM are addressed: insufficient labeled samples and difficulty of applying CNNs to geological prospecting big data for MPM, which are characterized by massive size, multiple sources, multiple types, multi-temporality, multiple scales, non-stationarity, and heterogeneity. The random-drop data augmentation method, which repeatedly takes dropouts from data, is adopted in this study for generating sufficient training samples. Various experiments are conducted to determine a suitable CNN architecture for MPM. The mapped areas obtained by the constructed CNN are strongly spatially correlated with the locations of known mineralization, and most of the known Fe polymetallic deposits are located in areas with high probabilities. Our findings indicate that such a random-drop data augmentation method is suitable and effective for constructing training datasets to predict the locations of rare geological events. Additionally, CNN appears as a promising tool for integrating multisource geoscience data, thereby supporting further mineral exploration. KEY WORDS: Mineral prospectivity mapping, Convolutional neural network, Data augmentation, Geological prospecting big data.
INTRODUCTION The aim of mineral prospectivity mapping (MPM) is to quantify the probability of the presence of mineral deposits in a given area for facilitation of mineral exploration. Since the 1980s, various methods have been adopted or proposed for MPM; these can be divided into knowledge- and data-driven methods. Knowledge-driven MPM methods, such as fuzzy logic (An et al. 1991), are propitious to less explored geologically permissive regions, where no or very few mineral deposits are discovered (Car1
State Key Laboratory of Geological Processes and Mineral Resources, China University of Geosciences, Wuhan 430074, China. 2 To whom correspondence should be addressed; e-mail: [email protected]
ranza and Laborte 2015a). Data-driven methods, such as the weight of evidence (Bonham-Carter 1989; Agterberg et al. 1990; Cheng and Agterberg 1999; Carranza 2004; Porwal et al. 2006a; Zhang et al. 2016), and logistic regression (Agterberg and Bonham-Carter 1999; Chen et al. 2011; Zhang et al. 2018a, b; Xiong and Zuo 2018), have also garnered significant attention. Among the data-driven methods, machine learning algorithms such as support vector machines (Zuo and Carranza 2011; Abedi et al. 2012; Guerra Prado et al. 2020; Xiong and Zuo 2020), random forests (Carranza and Laborte 2015a, b, 2016; McKay and Harris 2016; Wang et al. 2020), neural networks (Brown et al. 2012; Xiong and
Data Loading...
