A data-driven approach for risk assessment and material identification of buried objects using microwave measurements and neural networks Authors: Gülçin Yarımay, İbrahim Halil Bayat, Semih Doğu, İbrahim Akduman Published: 2025 ABSTRACT Buried object detection plays a critical role in a wide range of applications, including defense, security screening, and subsurface sensing. Traditional microwave-based techniques often depend on handcrafted features or manual interpretation of measurement data, which limits their robustness and accuracy in complex or cluttered environments. To address these challenges, this study presents a deep neural network model trained on experimental multi-frequency microwave S-parameter data for the automated detection and classification of buried objects. The proposed framework processes 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}21 S-parameter matrices to capture the electromagnetic response of the subsurface medium. It classifies each sample into three distinct categories: safe, plastic-based threat, and metal-based threat. The model achieves an accuracy of 0.9647, precision of 0.9677, recall of 0.9647, and an F1 score of 0.9635, demonstrating consistent performance across all evaluation metrics. By learning directly from real, experimentally measured data, the network effectively models the complex relationship between microwave responses and object characteristics, eliminating the need for simulation-based data or handcrafted features. These findings represent a substantial step toward the realization of practical, data-driven microwave detection systems, offering improved reliability, scalability, and operational potential for future security and defense applications. Introduction The detection of buried objects has long been a critical aspect of both military operations and humanitarian efforts, such as minefield clearance1. While the ability to detect subsurface anomalies is a fundamental requirement, it is not sufficient on its own to ensure operational success or safety. Equally important is the capacity to accurately classify and assess the potential threat posed by the detected objects. This is particularly crucial when dealing with metallic components, which are often indicative of explosive devices or other hazardous materials2. Effective classification enables the differentiation between dangerous objects and harmless trash, thereby reducing false alarms and optimizing resource allocation. Moreover, precise identification of threat levels contributes significantly to safeguarding personnel, enhancing situational awareness, and improving the overall efficiency and reliability of minefield clearance and subsurface surveillance missions. Considering buried objects, Ground Penetrating Radar (GPR) is the most commonly employed technique for its detection3. One of the primary advantages of GPR is the ability to non-invasively detect subsurface anomalies, making it suitable for a variety of military and civilian applications. GPR generates graphical representations, also known as radargrams, of reflected electromagnetic signals that require expert interpretation. This allows for the inference of object shape, depth, and composition, particularly when data is collected over a grid and reconstructed into three-dimensional models. However, a notable limitation of GPR is its reduced performance in high-conductivity soils where signal attenuation can significantly affect data quality and interpretability4. Microwave systems represent a valuable tool for extracting information about the objects in unreachable places due to their capability to penetrate a range of non-metallic and dielectric materials5. These systems function within the microwave spectrum, transmitting electromagnetic waves that interact with underground anomalies and return signals that can reveal an object’s location, size, and material properties6. A major strength of microwave-based techniques lies in their high spatial resolution, allowing for more accurate identification and positioning compared to some conventional subsurface sensing methods. Moreover, the non-contact nature of microwave systems enhances their safety and usability, particularly in dangerous or hard-to-access areas. Consequently, microwave-based systems have gained widespread adoption across both academic research and industrial applications7–11. Microwave systems for buried object detection are limited by signal attenuation in soils with high moisture content or electrical conductivity, which reduces penetration depth and detection accuracy. Furthermore, complex subsurface environments can cause scattering and multipath effects, complicating the interpretation of returned signals. Additionally, differentiating between materials with similar dielectric properties remains a significant challenge, potentially leading to ambiguous identification of buried objects. To address the inherent challenges associated with microwave-based buried object detection, we propose a data-driven artificial intelligence approach. By leveraging advanced machine learning algorithms, this approach can enhance the interpretation of complex microwave measurement data, improve detection accuracy, and reduce false alarm rates. The AI framework is capable of learning intricate patterns and subtle variations within the electromagnetic signals, thereby overcoming limitations posed by conventional signal processing techniques and enabling more reliable identification of buried objects in diverse and challenging subsurface conditions. Among the research on artificial intelligence–driven detection and enhancement methods, Chiu et al. make a notable contribution to the electromagnetic inverse scattering problem for buried conductors through the use of a deep convolutional neural network (CNN)12. Conventional iterative approaches to this problem typically require repeated Green’s function calculations and are therefore computationally intensive, sensitive to noise, and difficult to scale for real-time applications. In contrast, the CNN-based framework developed by Chiu et al. directly reconstructs conductor geometries from scattered electromagnetic fields, effectively bypassing these computational bottlenecks while maintaining strong reconstruction accuracy. Numerical experiments demonstrate that the proposed method achieves root mean square errors generally below 18%, even for irregular and complex conductor shapes, outperforming established architectures such as U-Net. The ability to robustly recover diverse geometries underscores the model’s broader applicability across domains where electromagnetic imaging is essential, including nondestructive industrial testing, underground pipeline monitoring, biomedical imaging, and military detection. By integrating data-driven modeling with principles of electromagnetic theory, this work illustrates the potential of deep learning to overcome the limitations of traditional inverse scattering methods and provides a foundation for future advances in efficient, high-resolution subsurface imaging. A complementary line of research emphasizes improving the speed and computational efficiency of buried object characterization by minimizing reliance on extensive preprocessing stages13. In this work, the authors proposed a time–frequency regression model (FRM) capable of directly analyzing raw ground-penetrating radar (GPR) signals. This design substantially accelerates the interpretation workflow, thereby reducing detection latency and improving the responsiveness of subsurface sensing systems. Rather than relying on conventional intermediate steps such as background removal, filtering, or handcrafted feature extraction, the FRM employs data-driven surrogate modeling to estimate object parameters in a more direct and computationally lightweight manner. This streamlined pipeline not only improves processing speed but also highlights the potential of regression-based surrogate models to enhance the practical deployment of GPR-based detection technologies in real-time scenarios. Despite these advantages, the study also presents certain limitations that restrict its broader applicability. The proposed model was validated exclusively on a single cylindrical perfect electric conductor (PEC), a relatively simplified test case compared to the variability encountered in real-world conditions. Such a narrow evaluation framework limits confidence in the method’s ability to generalize across diverse subsurface environments, where buried targets may differ significantly in terms of geometry, material heterogeneity, and electromagnetic scattering characteristics. Addressing these constraints through more comprehensive validation on diverse object classes and realistic soil conditions would be a critical next step to establish the robustness and scalability of this approach for practical field applications. A recent line of investigation has advanced the classification of buried objects by introducing a second-order deep learning framework that extends the conventional convolutional neural network (CNN) paradigm14. In this approach, ground-penetrating radar (GPR) images are first processed to extract hyperbola-shaped thumbnails, which then serve as inputs to a baseline CNN. Unlike standard models that rely solely on first-order feature representations, this method computes covariance matrices from intermediate convolutional layer outputs, thereby capturing second-order statistical dependencies that provide richer structural information about the feature maps. These covariance descriptors are subsequently processed through specialized layers designed to handle symmetric positive definite (SPD) matrices, enabling the network to model complex spatiotemporal patterns and discriminate subtle characteristics of buried objects more effectively than traditional CNN-based pipelines. Empirical evaluation on a large and heterogeneous dataset demonstrated that the second-order model consistently outperformed both shallow GPR-specific neural networks and widely adopted CNN architectures from computer vision, particularly under conditions where training samples were scarce or mislabeled. The robustness of the model was further validated across variable environmental conditions, including differing soil properties and weather modes, underscoring its potential for real-world deployment. Beyond subsurface detection, the authors argue that this architectural innovation has broader applicability, with possible extensions to radar-based remote sensing and biomedical signal analysis domains such as electroencephalography (EEG), where noisy inputs and limited annotated data are common challenges. Despite these promising findings, the study acknowledges that the increased architectural complexity and associated computational demands may hinder large-scale or resource-constrained implementation, highlighting the need for future research on optimization strategies to balance accuracy with efficiency. Generative adversarial networks (GANs) are applied to solve the inverse scattering problem for buried dielectric objects, addressing challenges faced by traditional iterative methods such as nonlinearity, high computational cost, and incomplete data in15. The GAN framework consists of a generator that produces realistic images and a discriminator that distinguishes between real and generated images, refining the results through an adversarial process. However, GANs are known to be unstable to train, which may affect convergence and reconstruction quality16. Compared to the U-Net, GAN demonstrates superior image reconstruction accuracy and reliability for objects with varying dielectric properties and shapes. This approach represents a novel application of deep learning techniques for enhanced imaging of buried objects. Recent advancements in ground-penetrating radar (GPR) and microwave imaging have increasingly leveraged deep learning techniques to enhance subsurface object detection, clutter suppression, and data inversion accuracy. Ni et al. (2020) introduced a robust autoencoder (RAE)-based framework for clutter suppression in GPR B-scan images, demonstrating that decomposing radar data into low-rank and sparse components enables effective separation of background clutter and target responses17. Their method outperformed classical approaches such as mean subtraction, SVD, RPCA, and MCA on both simulated and real GPR data, yielding cleaner and more interpretable subsurface imagery. Building on this, Ni et al. (2022) applied conditional generative adversarial networks (cGANs) to the same problem, developing a data-driven approach that maps cluttered B-scan images to clutter-free counterparts18. By augmenting simulated clutter-free data with real non-target measurements, their model improved generalization to field data and achieved superior clutter suppression and computational efficiency compared to state-of-the-art low-rank and sparse decomposition methods. While these studies primarily focused on improving signal clarity, Wang et al. (2025) extended deep learning applications toward quantitative inversion of subsurface parameters by proposing a semisupervised neural architecture for reconstructing the relative permittivity of buried defects19. Their model integrated a SAR feature fusion branch and multidimensional attention module to enhance the representation of defect boundaries, achieving strong generalization to new defect types and outperforming existing inversion methods even under limited annotated data conditions. In parallel, Cheng et al. (2024) advanced the interpretability of GPR data through a signal processing approach based on the high-order synchrosqueezing transform (HSST), which provided superior time–frequency resolution and precise reconstruction of geological structures in both simulated and real mining environments20. Collectively, these studies underscore the rapid evolution of deep learning and advanced signal processing techniques in GPR research from clutter suppression and feature extraction to data inversion and structural interpretation. However, most of these approaches rely heavily on simulated datasets or specific geological assumptions, leaving a gap in experimentally validated, data-driven microwave detection systems for real buried object classification. Addressing this gap, the present study proposes a deep neural network trained exclusively on experimental multi-frequency S-parameter data, aiming to achieve robust and automated classification of buried threats under realistic measurement conditions. In this study, a neural network model was developed and trained using real experimental data provided by Mechanical and Chemical Industry, ensuring practical relevance and robustness in performance. The input features for the model consisted of scattering parameters (S-parameters), which capture the frequency response characteristics of the system and provide rich information for identifying buried objects. The model is designed to classify detected objects into three distinct categories: safe (no threat), metal-based threat, and plastic-based threat, enabling more precise risk assessment for buried threats. By leveraging these experimentally obtained S-parameters, the neural network learns complex patterns associated with subsurface anomalies while maintaining computational efficiency. This approach demonstrates the potential for efficient and accurate buried object detection and classification using data-driven models grounded in real-world measurement data. It is worth noting that our earlier study21 laid the foundation for this line of research by integrating a lightweight neural network model with a microwave-based detection system. In that work, 16 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 16 S-parameter measurements were employed, which were then transformed into a compact 256-dimensional feature vector, allowing for efficient binary classification of objects into threat and non-threat categories. The framework achieved remarkably high accuracy and demonstrated strong potential for defense and security applications. However, that study was limited in several respects: the dataset comprised only 583 samples in total, of which 48 represented threat cases, and the model operated under a single-frequency setting with a relatively simple network structure. Building upon these findings, the present study introduces several key advancements that substantially extend the prior work. First, the input dimensionality has been expanded by employing 21 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 21 S-parameters, enabling a more detailed representation of the electromagnetic responses. Second, the dataset has been dramatically scaled up to 20,405 samples, including 1680 threat cases, and importantly, the data used in this work are entirely distinct from those employed in21. This increased volume and diversity of measurements not only enhances statistical robustness but also ensures improved generalization across varied subsurface conditions. Furthermore, unlike the single-frequency setting of the earlier study, the current work leverages multi-frequency data, thereby incorporating additional spectral complexity and necessitating a more sophisticated neural network architecture. Finally, the classification task has been extended beyond a binary framework to a three-class setting, distinguishing between safe (no threat), metal-based threat, and plastic-based threat. This refined categorization offers a practical advantage by differentiating the nature of threats, allowing for tailored precautionary measures depending on whether the detected object is metallic or plastic. Collectively, these advancements illustrate the progression from our earlier work toward a more robust and data-rich solution for buried object detection and risk assessment. Data acquisition The dataset employed in this study was collected using a fully automated microwave measurement system capable of transmitting, receiving, and recording scattering parameters (S-parameters) while simultaneously repositioning the antenna array with high precision. The antennas were designed to operate with a reflection coefficient below –10 dB across the measurement band, ensuring minimal mismatch losses and stable radiation characteristics. Unlike studies relying solely on simulations, the dataset in this work is entirely experimental, thereby capturing the inherent complexities of real-world subsurface electromagnetic interactions. Measurements were conducted over a wide frequency range of 600 MHz to 4 GHz, sampled in 100 MHz increments, resulting in a total of 35 discrete frequency points. This multi-frequency acquisition allowed the model to capture both the penetration capability of lower-frequency signals and the higher-resolution information provided at higher frequencies. The antenna array, configured to measure 21-port reflection and transmission coefficients, was positioned at a fixed height of 30 cm above the ground surface. The testbed was installed inside an anechoic chamber to suppress unwanted reflections, while the measurement pool was filled with gravelly sand. The chamber temperature was controlled and maintained at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$20^\circ$$\end{document}C during the entire experimental campaign to ensure repeatability and environmental stability. The experimental setup is illustrated in Fig.1.Fig. 1Illustration of the experimental measurement setup for buried object detection. Illustration of the experimental measurement setup for buried object detection. The scattered S-parameters characterizing the buried object problem can be expressed as22,23:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} S_\text {a,b}^\text {sca} = \int \limits _\Omega \mathbf {g\left( m_\text {a},r'\right) }, \mathbf {\chi \left( r'\right) }, \mathbf {e\left( n_\text {b},r'\right) }, d\Omega \mathbf {(r')}, \quad \mathbf {r'}\in \Omega ;\quad \mathbf {m_\text {a},n_\text {b}}\in \Gamma \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_\text {a,b}^\text {sca}$$\end{document} denotes the scattered field coefficients measured between transmitting antenna \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_b$$\end{document} and receiving antenna \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m_a$$\end{document}. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {\chi \left( r'\right) }$$\end{document} represents the object function, defined as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathbf {\chi \left( r'\right) } = \mathbf {k\left( r'\right) } - k_0, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {k\left( r'\right) }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_0$$\end{document} are the wavenumbers of the soil at point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r'$$\end{document} and the background medium, respectively. The Green’s function, which links the excitation from the transmitting antenna to the scattered field at the receiving antenna, is expressed as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathbf {g\left( m_\text {a},r'\right) } = \frac{iZ_0}{2\omega \mu }, \mathbf {e^\text {inc}\left( m_\text {a},r'\right) }, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_0$$\end{document} is the input impedance of the antenna, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega$$\end{document} is the angular frequency, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu$$\end{document} is the permeability, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {e^\text {inc}}$$\end{document} denotes the normalized incident field. The total S-parameters recorded in the experiments can thus be written as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} S_\text {a,b}^\text {tot} = S_\text {a,b}^\text {inc} + S_\text {a,b}^\text {sca}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_\text {a,b}^\text {inc}$$\end{document} is the directly measured incident coefficient without object interaction. These formulations provide the physical foundation for the experimental measurements utilized in this work. To guarantee representativeness, a variety of objects were intentionally buried in the sand pool and grouped into three primary categories: (1) Safe (No Threat): consisting of background materials such as wood chips, glass, plastics, as well as measurements from empty-ground regions with no buried items, (2) Metallic Threats: including metallic fragments and containers, and (3) Non-metallic Threats: hazardous powders stored within non-metallic containers. This classification ensured that the dataset encompassed both common background conditions (predominantly safe) and the less frequent but high-risk hazardous cases. In total, the dataset comprised 20,405 samples, distributed as 18,605 safe, 840 metallic threats, and 840 non-metallic threats. The distribution is presented in Table1, illustrating the realistic imbalance between safe and hazardous samples, consistent with real-world operational environments where threats are far less frequent than non-threatening situations. The experimental setup and an example of a buried object are shown in Fig. 2.Table 1Composition of the experimental dataset used in this study.ClassNumber of SamplesSafe (No threat)18,605Metallic Threat840Plastic Threat840 Table 1Composition of the experimental dataset used in this study.ClassNumber of SamplesSafe (No threat)18,605Metallic Threat840Plastic Threat840 Composition of the experimental dataset used in this study. This extensive and experimentally collected dataset, spanning a broad frequency range and diverse object categories, provides the basis for training and validating the proposed neural network model. The multi-frequency S-parameters allow the network to exploit both penetration depth and high-resolution features, enabling robust and reliable buried object classification in realistic scenarios. Method Artificial neural networks (ANNs) are powerful computational models inspired by the structure and functionality of the human brain, specifically designed to identify, represent, and generalize complex patterns and relationships within data24. Over the past decades, ANNs have become a cornerstone of modern machine learning due to their remarkable adaptability and scalability across various problem domains. They have achieved state-of-the-art performance in numerous applications, such as image processing and computer vision25, natural language understanding and generation26, as well as predictive modeling and analytics in diverse scientific and industrial contexts27. The flexibility of neural networks stems from their ability to model nonlinear mappings from input to output spaces, enabling them to handle both structured and unstructured data effectively.Fig. 2Experimental setup showing the sandbed (left) and a representative buried sample (right). Experimental setup showing the sandbed (left) and a representative buried sample (right). The neural network model employed in this study was implemented using the TensorFlow Keras API, following a sequential architecture specifically designed for a multi-class classification task involving three distinct output categories. The input layer is structured to accept feature vectors of dimensionality 441, which represent a flattened form of structured S-parameters (21 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 21). This transformation ensures that the network can directly process the essential features extracted from raw microwave data. The illustration of the network is given in Fig. 3.Fig. 3Illustration of the network. Illustration of the network. An important strength of this architecture lies in the use of dense (fully connected) layers. Dense layers operate by computing weighted sums of their inputs, followed by the application of non-linear activation functions. Such functions are essential for enabling deep networks to capture and approximate complex mappings between inputs and outputs. The ReLU function, defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f(x) = \max (0, x)$$\end{document}, is widely used because of its computational efficiency and ability to mitigate the vanishing gradient problem in deeper networks. Another common activation function is the Sigmoid function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma (x) = 1 / (1 + e^{-x})$$\end{document}, which squashes outputs into the range (0, 1) and is particularly useful in classification tasks. By introducing non-linearity, these activation functions empower deep neural networks to learn hierarchical feature representations that progressively capture low-level patterns in earlier layers and higher-level abstractions in later layers28. The model begins with an input dense layer containing 1024 neurons, activated by the Rectified Linear Unit (ReLU) function. The large number of neurons in this initial layer allows the network to capture a broad spectrum of complex, non-linear interactions present within the input data, ensuring that subtle variations in the electromagnetic responses are preserved in the early stages of processing. Following this, the network incorporates a series of hidden layers with gradually decreasing neuron counts, forming a hierarchical compression pathway. Specifically, layers with 512, 256, 128, 64, and 32 neurons are employed in sequence, each utilizing the ReLU activation function. This structured reduction in dimensionality serves multiple purposes: it mitigates the risk of overfitting by preventing excessive capacity at deeper stages, it forces the network to progressively abstract and refine the learned features, and it encourages the emergence of higher-level representations that are more robust to noise and variability in the data. The ReLU activation at each stage introduces the non-linearity required to model the complex, non-Gaussian nature of electromagnetic scattering phenomena, which would otherwise be intractable using linear transformations alone. An important characteristic of this multi-layered architecture is its capacity to balance expressive power with computational efficiency. The deeper layers with fewer neurons act as information bottlenecks, encouraging the network to filter out redundant or irrelevant patterns while preserving salient features necessary for discrimination. This design enables the model to generalize effectively to unseen data by focusing on essential structural properties of the S-parameters rather than memorizing spurious details. Additionally, the hierarchical feature extraction process supports multi-frequency input analysis, allowing the network to integrate information across different frequency bands and thereby handle more complex scenarios compared to single-frequency models. The final layer of the network is a dense output layer consisting of three neurons, each corresponding to one of the target classes: safe (no threat), metal-based threat, and plastic-based threat. A softmax activation function is employed at this stage, which transforms the raw outputs into normalized probability distributions. This probabilistic interpretation provides not only the most likely class label but also a measure of confidence for each prediction, an essential feature for risk-sensitive applications such as buried object detection. The explicit separation of metallic and non-metallic threats further enhances the practical utility of the system, as it enables tailored precautionary measures to be taken depending on the nature of the identified risk. Taken together, this architecture illustrates a deliberate progression from high-dimensional input modeling to compact and semantically meaningful feature representations. By leveraging a deep, gradually narrowing structure with ReLU activations and a probabilistic softmax output, the network achieves a balance between complexity, interpretability, and computational efficiency. This design is particularly well-suited for real-world defense and security applications, where large-scale data must be processed accurately and efficiently. The model is compiled using the Adam optimizer, an adaptive learning rate optimization algorithm that combines the advantages of momentum and RMSProp, thereby achieving both stability and efficiency in training deep neural networks. In this study, the learning rate is set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.001$$\end{document}, which represents a commonly adopted default value that provides a balance between convergence speed and training stability. For the loss function, sparse categorical cross-entropy is employed, as it is particularly well-suited for multi-class classification tasks. The sparse categorical cross-entropy loss for a dataset with N samples, C classes, ground-truth labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i \in {1, \dots , C}$$\end{document}, and predicted probability distributions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_i$$\end{document}, is mathematically defined as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} L = -\frac{1}{N} \sum _{i=1}^{N} \log \big ( \hat{y}_{i,{y_i}} \big ) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{i, y_i}$$\end{document} denotes the predicted probability assigned by the model to the true class \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i$$\end{document} for the i-th sample. This formulation ensures that the model learns to maximize the likelihood of the correct class, penalizing incorrect predictions more strongly when the confidence in the wrong class is high. The epoch for the training process has been set to 100. Model performance is assessed using four complementary evaluation metrics: accuracy, precision, recall, and the F1 score, providing a comprehensive understanding of the network’s classification capability across different aspects. Accuracy (Acc) quantifies the overall correctness of the model by measuring the proportion of correctly predicted samples relative to the total number of samples. Formally, it is defined as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} \end{aligned}$$\end{document}where TP and TN denote the true positives and true negatives, respectively, and FP and FN represent false positives and false negatives. While accuracy provides an intuitive measure of overall correctness, it can be misleading in the presence of class imbalance. To address this, precision and recall are employed. Precision (P) indicates the proportion of correctly predicted positive instances out of all instances predicted as positive:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Precision} = \frac{TP}{TP + FP} \end{aligned}$$\end{document}Recall (R), also known as sensitivity, measures the proportion of correctly predicted positive instances relative to all actual positive instances:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Recall} = \frac{TP}{TP + FN} \end{aligned}$$\end{document}Finally, the F1 score provides a harmonic mean of precision and recall, balancing the trade-off between these two metrics and offering a single scalar value that reflects the model’s ability to correctly identify positive samples while minimizing false positives:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F1 = 2 \cdot \frac{\text {Precision} \cdot \text {Recall}}{\text {Precision} + \text {Recall}} \end{aligned}$$\end{document}By considering all four metrics together, we can obtain a more nuanced assessment of model performance, particularly in scenarios involving multiple classes or imbalanced datasets. Accuracy captures the overall correctness, precision emphasizes the reliability of positive predictions, recall evaluates the model’s sensitivity to actual positive instances, and the F1 score provides a balanced measure that accounts for both false positives and false negatives. This multi-metric evaluation ensures that the network’s performance is rigorously quantified and practically interpretable for buried object detection and classification tasks. Results The performance of the proposed neural network model was evaluated using two complementary validation strategies: a k-fold cross-validation with k = 5 and a conventional train–test split approach. These methods were jointly employed to ensure a comprehensive and reliable assessment of the model’s generalization capability. The k-fold cross-validation procedure provided an averaged estimate of performance across multiple randomized partitions of the dataset, thereby reducing dependence on any single split and mitigating the effects of overfitting. The choice of k = 5 was made as a practical compromise between computational efficiency and statistical reliability. As discussed by Kohavi (1995), increasing k typically decreases the variance of the validation estimate but increases computational cost and potential bias, whereas smaller k values can lead to unstable results due to higher variance. Given the moderate size and complexity of the dataset, k = 5 was found to offer a suitable balance between these competing factors. In parallel, a conventional train–test split was employed to simulate a realistic deployment training once and evaluating on an unseen subset to facilitate direct comparison with previous studies that use single split validation. The combination of these two evaluation strategies thus provided both robustness and practical relevance in characterizing the performance of the proposed model29. Table 2 summarizes the results in terms of four standard classification metrics: accuracy, precision, recall, and F1 score.Table 2Performance metrics for the proposed network using different validation strategies.ValidationAccuracyPrecisionRecallF1-ScoreK-Fold (k=5)0.95610.95240.95610.9502Train-Test Split0.96470.96770.96470.9635 Table 2Performance metrics for the proposed network using different validation strategies.ValidationAccuracyPrecisionRecallF1-ScoreK-Fold (k=5)0.95610.95240.95610.9502Train-Test Split0.96470.96770.96470.9635 Performance metrics for the proposed network using different validation strategies. The results indicate that the model achieves consistently high performance under both evaluation protocols. The train-test split approach produced slightly higher scores across all metrics, with an accuracy of 0.9647, precision of 0.9677, recall of 0.9647, and an F1 score of 0.9635. The k-fold cross-validation approach, while marginally lower, still demonstrated robust performance with an accuracy of 0.9561, highlighting the model’s stability across varying training and testing subsets. An additional consideration during model development was the effect of dropout regularization. Preliminary experiments incorporating dropout layers revealed a significant deterioration in performance, particularly for the recall metric. Specifically, the recall dropped to as low as 0.6413 when dropout was applied, suggesting that the network became less sensitive to positive instances of buried threats. Consequently, dropout was omitted in the final architecture to preserve the model’s discriminative capability and maintain robust detection of all classes. The confusion matrix for the train–test split validation is presented in Fig. 4. This matrix provides a detailed overview of the classification outcomes for each class: Safe (No Threat), Metal-based Threat, and Plastic-based Threat. The network demonstrates outstanding performance in correctly identifying safe samples, with 5573 true negatives out of 5618 total safe instances. For metallic threats, 192 out of 252 samples were correctly classified, and for plastic-based threats, 141 out of 252 samples were accurately identified.Table 3Comparison of classification performance between the proposed MLP and the CNN baseline.ModelAccuracyPrecisionRecallF1-ScoreProposed (MLP)0.96470.96770.96470.9635CNN Baseline0.91770.84210.91770.8783 Table 3Comparison of classification performance between the proposed MLP and the CNN baseline.ModelAccuracyPrecisionRecallF1-ScoreProposed (MLP)0.96470.96770.96470.9635CNN Baseline0.91770.84210.91770.8783 Comparison of classification performance between the proposed MLP and the CNN baseline. Fig. 4Train-test split confusion matrix illustrating the classification performance across the three target classes. Train-test split confusion matrix illustrating the classification performance across the three target classes. To further assess the effectiveness of the proposed architecture, its performance was compared with that of a convolutional neural network (CNN) baseline trained under identical conditions, as well as a generative adversarial network (GAN) based classifier. The comparative results, summarized in Table 3, clearly demonstrate the superiority of the MLP-based approach. While the CNN achieved an accuracy of 0.9177 and an F1-score of 0.8783, the GAN classifier reached a simulated accuracy of only 0.3371 despite adversarial training over 50 epochs (with generator loss decreasing from 0.761 to 0.218 and discriminator loss increasing from 0.920 to 1.623). The proposed model attained substantially higher values across all evaluation metrics. This improvement reflects the ability of the MLP to capture complex nonlinear correlations within the multi-frequency S-parameter data while maintaining computational simplicity. The few misclassifications observed occur primarily between the two threat categories, where plastic-based threats were predicted as metallic and vice versa. This behavior is consistent with the quantitative error distribution summarized in Fig. 5, which shows that misclassified samples are sparsely scattered across the frequency range of 0.6–4.0 GHz without exhibiting any systematic or frequency-dependent trend. While minor variations are observed at isolated frequencies, most notably around 0.6 GHz, 1.1 GHz, and 2.0–3.7GHz, the distribution of errors does not indicate a consistent spectral bias or region-specific degradation in performance. The network maintains stable discriminative capability across the entire operational bandwidth, suggesting that its decision-making mechanism relies on global spectral–spatial correlations rather than frequency-localized cues. This finding is particularly significant given that environmental parameters such as soil moisture, temperature, and burial depth were carefully controlled throughout the experiments, thereby isolating the misclassifications as consequences of intrinsic scattering similarities between materials rather than external influences. In particular, plastic-based threats with relatively high dielectric constants or embedded conductive components can produce scattering patterns partially resembling those of low-conductivity metals, leading to confusions between these two categories. Nevertheless, the total number of misclassifications remains very small relative to the dataset size, underscoring the robustness and generalization capability of the proposed network across multiple frequencies and material classes. The absence of a discernible correlation between frequency and misclassification numbers further confirms that the model effectively captures broadband electromagnetic characteristics, enabling consistent detection performance over the full operational spectrum.Fig. 5Misclassified Sample Counts Across Frequencies by Class. Misclassified Sample Counts Across Frequencies by Class. Overall, the results underscore the effectiveness of the proposed deep neural network for multi-class buried object classification. The combination of a multi-layered architecture, multi-frequency input data, and a well-structured training process without dropout regularization has led to high classification accuracy, reliable precision, and strong recall performance. These outcomes highlight the practical applicability of the approach in real-world detection scenarios, where accurate discrimination between safe and threat objects is critical for operational safety and risk mitigation. Discussion The findings presented in this study demonstrate the robustness and reliability of the proposed neural network in the classification of buried threats across different materials and frequencies. The consistent performance observed under both k-fold cross-validation (k=5) and train-test split validations underscores the generalization ability of the model and highlights its potential suitability for real-world deployment. In particular, the marginal differences between the two evaluation strategies, where the train-test split achieved slightly higher values than the k-fold method, suggest that the model is not overfitted to specific subsets of the data and can maintain stable performance across diverse training and testing partitions. This observation is critical in practical applications, where models are often required to handle highly variable and previously unseen operational conditions. A notable aspect of the results is the high accuracy in identifying safe instances compared to the classification of threat categories. The confusion matrix reveals that the vast majority of safe samples are correctly recognized, with only a negligible proportion misclassified. This outcome can be partly attributed to the inherent class distribution in the dataset. Specifically, the number of safe samples significantly outweighs that of threat samples. While such an imbalance might initially appear to bias the model, it is, in fact, reflective of realistic operational environments. In everyday scenarios, encounters with threats are naturally rarer than encounters with non-threatening conditions. This distribution mirrors the actual occurrence rate of buried explosives or hazardous materials, which are expected to constitute only a small fraction of all scanned regions in field operations. Consequently, the predominance of safe samples in the dataset does not introduce an unrealistic skew but rather aligns with the practical expectations of surveillance and detection tasks. Nevertheless, despite the relatively smaller sample size of threat categories, the model demonstrates considerable capability in detecting both metallic and plastic-based threats. Although a portion of the misclassifications occurs between these two categories, the model’s ability to correctly identify a substantial number of threat instances underscores its sensitivity to critical features that differentiate hazardous materials from non-hazardous conditions. From an operational standpoint, this is of particular importance, as the misclassification of safe instances as threats (false positives) generally incurs less risk compared to the opposite case of threats being mislabeled as safe (false negatives). The high recall rates observed, especially in the train-test split evaluation, indicate that the network is effective in minimizing false negatives, thereby reducing the likelihood of overlooking potential dangers. An additional point of discussion pertains to the role of dropout regularization. While dropout is commonly employed to mitigate overfitting in deep learning models, our experiments revealed that its inclusion substantially degraded performance, particularly recall. This suggests that the dropout layers reduced the network’s capacity to preserve essential discriminative patterns required for detecting low-frequency threat instances. The decision to remove dropout in the final architecture thus appears justified, as it enabled the network to maintain high sensitivity toward minority classes without compromising overall accuracy. Taken together, these results indicate that the proposed architecture, informed by multi-frequency input and a carefully balanced design, effectively addresses the dual challenge of high-class imbalance and the need for precise detection of rare but critical threat cases. By achieving strong accuracy, precision, and recall simultaneously, the network demonstrates a meaningful advancement in the application of deep learning for buried object detection. Furthermore, the observed performance stability across validation schemes enhances confidence in the model’s scalability to larger datasets. In operational contexts where safety and risk mitigation are paramount, such as military or humanitarian demining, the ability to reliably distinguish between safe and hazardous objects, even under challenging conditions, positions this approach as a valuable tool for future deployment and further research. Conclusion This study presents a deep neural network framework trained on experimental multi-frequency microwave S-parameter data for the automated detection and classification of buried objects. By directly utilizing 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}21 S-parameter matrices, the proposed approach effectively captures the complex electromagnetic interactions between the buried object and the surrounding medium, enabling accurate and data-driven decision-making without reliance on handcrafted or simulation-derived features. The model demonstrates robust performance, achieving an accuracy of 0.9647, precision of 0.9677, recall of 0.9647, and an F1 score of 0.9635, confirming its strong capability to distinguish between safe, plastic-based threat, and metal-based threat cases. A key innovation of this work lies in the integration of deep learning with real, experimentally obtained microwave data, bridging the gap between theoretical developments and practical field applications. Unlike many previous studies that depend primarily on simulated datasets or simplified assumptions, this research emphasizes experimental validation, ensuring that the trained model reflects the complexities and variabilities encountered in realistic detection environments. The multi-frequency design further enhances the discriminative power of the system, allowing the neural network to exploit spectral diversity for more accurate material characterization. The findings highlight the potential of the proposed neural network framework as a scalable and reliable tool for subsurface threat detection, particularly in real-world defense and security scenarios where class imbalance and environmental uncertainty are common. Moreover, this study establishes a foundation for the future development of real-time, portable, and adaptive microwave detection systems, integrating advanced signal processing and machine learning techniques. Future work will aim to expand the dataset by incorporating diverse soil types, temperature conditions, and environmental variations to further assess model generalization.