Abstract:

Deep learning models have been widely applied in mineral prospectivity mapping (MPM) due to their powerful ability to extract features from data. However, supervised deep learning methods often face challenges such as insufficient training samples and class imbalance between positive and negative samples. The inherent rarity of mineralization events further compromises model robustness and generalization ability. To address these issues, this study employs three distinct data augmentation methods:1. Sliding Window Augmentation: This method uses known positive and negative samples as centers and performs multiple sliding operations to generate augmented samples; 2. Generative Adversarial Network (GAN) Augmentation: Generative models, specifically GANs, are utilized. The networks are trained on real samples, and augmentation is achieved using the trained generator; 3. Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) Augmentation: Similarly, the WGAN-GP framework is trained on real samples, and its trained generator is used for sample augmentation. These three data augmentation methods effectively expand the sample size while maximally preserving the geological significance of the samples. To validate the effectiveness of augmentation, this study employs the Fréchet Inception Distance (FID) between real and generated samples alongside evaluation using a Convolutional Neural Network (CNN). The results demonstrate that the CNN model trained on the WGAN-GP-augmented dataset exhibits superior generalization ability. Furthermore, the mineral prospectivity map for gold deposits generated for the Gannan area provides important insights for future mineral resource exploration efforts.

Key words: sample imbalance, data augmentation, convolutional neural network, quantitative prediction