The application of machine learning in mineral exploration has gained considerable momentum over the past decade; however, its effectiveness is often constrained by the need for high-quality data, particularly in regions where comprehensive datasets are unavailable or difficult to acquire. The recent expansion of the space race and advances in earth-observation technologies have generated a wealth of remotely sensed datasets capable of capturing regional conditions at ever increasing spatial, spectral, and temporal resolutions. While high-resolution airborne multispectral datasets provide limited spectral detail and airborne hyperspectral data remain cost-prohibitive, open-source spaceborne datasets offer significant advantages despite their coarser resolution. Sentinel constellations, the Landsat series, EMIT hyperspectral imagery, and ASTER are among the most widely used spaceborne datasets, providing medium- to high-resolution imagery at commercially viable price points.
Although conventional machine-learning models such as random forests and support vector machines are widely used in remote-sensing applications for mineral exploration, the complex, noisy, and high-dimensional nature of these datasets often requires more sophisticated techniques to produce actionable results. This is where deep learning becomes indispensable. Deep-learning models—particularly convolutional neural networks (CNNs)— are uniquely suited to handle large-scale, multidimensional remote-sensing data. Specialized architectures such as UNET (U-shaped encoder-decoder network), LSTMs (Long Short-Term Memory networks), and GANs (Generative Adversarial Networks) facilitate automated feature extraction for mineral mapping, object detection to identify mine workings, and advanced time-series analysis, among other applications.
At SRK, we have developed an innovative variation of the UNET architecture— SCAR-UNET (Spatial Channel Attention Optimized Recurrent UNET Model). This advanced recurrent UNET is specifically designed to capture both spatial and channel features, enabling accurate identification of mine workings on a global scale. The training dataset for this model was assembled by sourcing, validating, and integrating information from the Global Mining Footprint dataset (https://doi.org/10.1038/s43247-023- 00805-6) along with corresponding Sentinel-2 satellite imagery.
The integration of spaceborne data with advanced deep-learning models such as SCAR-UNET represents a transformative leap in mineral exploration. Harnessing these tools enables faster, more accurate, and cost-effective exploration, unlocking new opportunities in regions once considered inaccessible.
