Comparative Analysis of Deep Learning Architectures for Mineral Deposit Segmentation Using Multi-Modal Geophysical Data

This paper presents a unified framework for segmentation of mineral deposits from multimodal geophysical data using deep learning techniques. The approach integrates multiple types of geophysical maps such as geological, gravity, magnetic, topographic, and thermal data. The preprocessing stage includes image scaling, precise geographic alignment, and generation of deposit masks by removal of extraneous markings. In addition, data augmentation is performed to enhance training set diversity. Thirteen segmentation architectures are evaluated, including multiple variants of convolutional neural networks (with and without attention mechanisms), several U-Net variants (such as Attention U-Net, ResUNet, and UNet++), as well as other advanced segmentation models including SegNet, BiSeNet, DeepLabV3, PSPNet, and HRNet. The experimental results indicate that models with attention mechanisms and residual connections provide superior segmentation accuracy and robustness. This work demonstrates the potential of advanced deep learning methods for efficient mineral deposit exploration.