Unlocking Geological Secrets: A Neural Network Approach Comparing Microwave and Multispectral Imaging

Technical Summary: Unlocking Geological Secrets: A Neural Network Approach Comparing Microwave and Multispectral Imaging

Problem Statement
Geological analysis and mineral detection increasingly rely on remote sensing technologies, yet gaps remain in the comparative evaluation of microwave imaging versus multispectral imaging (MSI) under unified testing protocols. While microwave technology is recognized for its ability to penetrate materials and detect internal structural changes, and MSI is established for characterizing surface physical and chemical properties, there is limited integration of Artificial Intelligence (AI) to classify geological materials based on these distinct data types. Furthermore, traditional analytical methods often struggle to extract clear patterns from the complex, non-linear, and overlapping response curves generated by diverse stone samples. This study addresses the need to evaluate the efficacy of both modalities in detecting geological frequencies and structural features across a diverse set of stone samples.

Methodology
The research employed a comparative experimental framework involving twenty distinct stone samples sourced from diverse global geological formations. The methodology integrated two primary data acquisition techniques:

1. Microwave Imaging: Utilizing an X-band network analyzer, the study measured Scattering Parameters (S-parameters), specifically reflection ($S_{11}$) and transmission ($S_{21}$), across the frequency range. This approach captured the dielectric properties and internal structural variations of the stones.
2. Multispectral Imaging (MSI): Samples were imaged using a system operating in the visible spectral range (400 nm to 700 nm) with 450 discrete spectral bands. This generated a three-dimensional data cube capturing surface reflectance spectra.

To address the complexity and non-linearity of the raw data, particularly the overlapping S-parameter curves, the researchers employed Artificial Neural Networks (ANNs) for classification. The workflow included:

• Data Preprocessing: Noise reduction, outlier elimination, and missing value imputation.
• Model Training: The ANN was trained to associate S-parameter data (and MSI spectral data) with specific stone types. The study iterated through training configurations, adjusting parameters such as input size, hidden units, and epoch counts (up to 497 epochs) to optimize convergence.
• Performance Evaluation: Models were assessed using accuracy metrics, loss progression, and confusion matrices to determine classification stability and feature separability.

Key Results
The study yielded distinct performance profiles for the two imaging modalities when processed through ANNs:

• Microwave Imaging: The ANN trained on X-band S-parameter data demonstrated superior robustness. After optimization, the model achieved approximately 95–100% accuracy with stable loss behavior (stabilizing around 0.2). The confusion matrix showed a perfect diagonal structure, indicating that the microwave signals provided highly discriminative features capable of distinguishing all twenty stone types with minimal misclassification. The model exhibited rapid convergence and high stability even with initial training fluctuations.
• Multispectral Imaging: The MSI-based ANN initially displayed unstable learning behavior, with low accuracy (<15%) and high, fluctuating loss values. Even after optimization, while the MSI model eventually reached comparable accuracy levels (~95%), it required significantly more complex tuning and a larger number of iterations to converge. The confusion matrices revealed higher susceptibility to misclassification among spectrally similar classes, suggesting that surface reflectance features were less distinct than internal dielectric features for this specific dataset.

Key Contributions

• Unified Comparative Analysis: The paper provides a direct comparison of microwave and multispectral imaging for stone classification under identical experimental conditions and using the same AI classification framework.
• ANN Integration: It demonstrates the necessity of using ANNs to interpret complex, non-linear scattering parameter data where traditional analytical methods fail to identify clear correlations.
• Dataset Expansion: The study utilizes a diverse dataset of twenty globally sourced stone types, enhancing the understanding of variability in mineralogical composition and dielectric behavior.
• Performance Benchmarking: The research establishes that while both methods can achieve high accuracy, microwave imaging offers superior stability, faster convergence, and better feature separability for internal rock characterization.

Significance and Claims
The authors conclude that microwave imaging is preferable for precise and stable geological analyses, particularly when internal structural features and dielectric properties are critical. The method's ability to penetrate materials allows for the detection of features that surface-based multispectral imaging cannot capture. Conversely, the paper acknowledges that multispectral imaging remains a viable option for rapid assessments where moderate accuracy is sufficient, despite its lower stability and higher sensitivity to surface conditions (e.g., polish, dirt) and the need for extensive preprocessing.

The study modestly claims that its findings highlight the specific advantages of integrating microwave sensing with neural networks for rock characterization, filling a gap in the literature regarding the comparative performance of these technologies. It does not claim to replace existing methods but rather to offer a more robust alternative for scenarios requiring high-precision internal geological analysis. The authors note that while the microwave model achieved near-perfect classification on the tested dataset, further validation with independent datasets is necessary to confirm generalization capabilities.