Evaluating The Impact Of Spectral Band Integration On 3D-CNN Performance For Mineral Identification

Abstract

Remote sensing plays a vital role in mineral exploration by offering cost-effective, large-scale, and non-invasive mapping solutions. However, many advanced mineral classification approaches rely heavily on hyperspectral data, which are costly, computationally demanding, and often lack global availability. This study evaluates the performance of extended spectral input combinations using ASTER multispectral imagery for mineral classification in skarn-type iron deposits of the Western Tianshan Mountains, China. While the earlier study demonstrated the potential of integrating selected band ratio (BR) images and principal component (PC) bands with ASTER data, the present research expands this by systematically assessing six additional configurations, including: (1) all 9 bands, (2) all 15 BR images, (3) a combined input of all PCs and BRs, (4) ASTER original bands with all PCs, (5) ASTER bands with all BRs, and (6) ASTER bands combined with both all PCs and all BRs. A 3D Convolutional Neural Network (3D-CNN) model was used to evaluate how spectral feature selection impacts mineral classification accuracy and model behavior. The results revealed significant variability in performance across configurations. Using all PCs alone achieved an overall accuracy (OA) of 89.31%, outperforming all BRs (OA = 84.73%), while the combined use of all PCs and BRs yielded the lowest accuracy (OA = 79.39%), likely due to spectral redundancy. In contrast, the best classification results were obtained using a targeted integration of ASTER bands, selected PCs, and selected BRs, achieving an OA of 98.64% and Kappa coefficient of 98.27%. The study also identified persistent misclassification challenges, with over 50% of skarn samples misidentified as wall rocks or intrusions and approximately 35% of intrusions misclassified as skarn or wall rock in certain setups. These issues were mitigated through optimized spectral integration, which enhanced class separability without the need for hyperspectral inputs. The findings demonstrate that a systematically optimized, sensor-flexible spectral strategy using ASTER data can deliver high classification performance while reducing the complexity, cost, and processing burden associated with hyperspectral approaches.