1. INTRODUCTION Hyperspectral data, inceptive reflectance spectra unique to minerals and rocks, making it a powerful tool not only for mineral identification and processing but also for a wide range of applications such as environmental monitoring, agriculture, and medical imaging (Okada et al., 2024), (Kumar et al., 2024a), (Shen et al., 2015), (Zhong & Zhang, 2012). Machine learning methods utilizing this data have demonstrated high accuracy in mineral species identification and classification (Kim et al., 2022). However, its high dimensionality and complexity of hyperspectral data analysis present significant challenges for effective utilization (Bioucas-Dias et al., 2013), (Li et al., 2019). Previous studies have developed Python-based scripts to analyze hyperspectral data, finding applications in mining and civil engineering (Laura et al., 2022), (Windrim et al., 2023). While these scripts have proven effective, they typically require command-line operations and programming skills, making them difficult for non-specialists to use. Furthermore, challenges in data visualization and module transitions pose additional barriers to accessibility for civil engineers and geological technicians. The importance of special user interface design is also emphasized in other fields, such as GIS-related fields (Goodchild, 2011). These challenges underscore the need for more intuitive and accessible solutions to facilitate broader adoption among non-expert users. In addition to usability concerns, the traditional machine learning techniques applied to hyperspectral image classification face intrinsic limitations related to feature selection, spatial information, and scalability. Selecting a feature can be associated with overfitting risk, biases selection, lack of generalizability and computational cost (Cheng, 2024). The spatial dependence can be ignored, and spatial heterogeneity as well as data representation problems can arise restraining the spatial information aspect in machine learning models (Nikparvar & Thill, 2021). These models have high computational costs, aspects such as data storage and processing constrain, distributed computing complexity and the compromise between accuracy and efficiency are challenges when handling large datasets (Saberi & Lilasathapornkit, 2024). To address the challenges in traditional machine learning models, deep learning models were created. The feature selection is embedded in the deep learning architecture models which learns automatically the importance of each feature using specialized layers (Figueroa Barraza et al., 2021). It improves accuracy and efficiency and reduces the risk of overfitting. The use of parallel processing and introduction of specialized hardware such as GPUs and TPUSs were introduced in deep learning to overcome the scalability constraints (Alzubaidi et al., 2021). CNN for instance evaluates the relationship between neighborhood points recognizing the spatial pattern, aspect that machine learning lacks (Son et al., 2021). Recent studies demonstrate that AI applied to hyperspectral data has significantly improved geological exploration tasks, including mineral identification and environmental impact assessments (Kumar et al., 2024b). Common methods for mineral or rock identification are timeconsuming, energy-intensive, and often costly due to the specialized instruments required for data processing, and they do not always guarantee accuracy (Long et al., 2022), (Zhang et al., 2023).
RkJQdWJsaXNoZXIy MTM0Mzk2