Multi-Modal Deep Embedded Clustering (MM-DEC): A Novel Framework for Mineral Detection Using Hyperspectral Imagery

Identification of minerals through remote sensing imagery plays an important role in mineral exploration, especially when ground truth data or field work is not accessible. Hyperspectral imaging, which collects information from hundreds of narrow spectral bands, is a powerful tool that provides intricate surface information, but presents challenges due to its high-dimensionality and the complex structure of the data. To this end, we propose Multi-Modal Deep Embedded Clustering (MM-DEC) approach, an innovative unsupervised learning framework that integrates Convolutional Autoencoders(CAEs), Variational Autoencoders (VAEs), and Gray Level Co-occurrence Matrix (GLCM) based texture extraction that is able to exploit the spatial, spectral, and texture features of mineral in consideration We demonstrate the MM-DEC potential to identify hematite prospects in the mineralized Kiriburu area of Jharkhand, India using EO-1 Hyperion hyperspectral data. Preprocessing pipeline includes denoising using Machine Learning(ML) and statistical techniques, followed by major land cover classification based on spectral indices including Normalized Difference Vegetation Index (NDVI); Normalized Difference Water Index (NDWI) and Normalized Difference Soil Index (NDSI). By using the widely applied Pixel Purity Index (PPI), pure spectral pixels are extracted for refining and providing a high-confidence target point of hematite detection. Evaluation metrics based on cluster quality such as Silhouette Score, and Davies-Bouldin Index (DBI), show that MM-DEC is a competent model, as it is able to produce relatively compact and well-defined clusters as compared to conventional clustering methods. We achieved a similarity score of 98.69% in reference to the hematite spectral signatures. With accurate cluster labels, geographical coordinates, and spectral similarity scores, our method establishes a precedent in mineral extraction and resource planning based on remote sensing advancements.