Mapping structural boundaries and lithological contacts in potential field data is fundamental to understanding crustal architecture in both continental and marine settings. Gravity and magnetic anomalies commonly exhibit lateral gradients where density or magnetization contrasts occur, providing key constraints on basin margins, rift systems, crustal structure, intrusive complexes, and fault networks (Telford et al., 1990; Blakely, 1995; Dwivedi et al., 2019; 2025). Consequently, boundary detection filter can detect the lateral location of the potential field sources (Dwivedi and Chamoli, 2022; Pham et al., 2025). Thus, numerous edge detection techniques have been developed to enhance and delineate such boundaries (Blakely and Simpson, 1986; Miller and Singh, 1994; Cooper, 2009; Nasuti and Nasuti, 2018; Sahoo et al., 2022; Pham et al., 2023; Ai et al., 2024; Alvandi et al., 2024; Alvandi et al., 2025; Deniz Toktay et al., 2025). One of the key advantages of these filters is that they do not require any prior knowledge about the source characteristics.

Derivative-based filters such as the Total Horizontal Derivative (THDR) enhance sharp lateral gradients and are widely applied to identify near-surface or steeply bounded sources (Cordell and Grauch, 1985; Cooper, 2009). The Analytic Signal Amplitude (ASA) is independent of magnetization direction and produces maxima over source edges (Nabighian, 1972; Roest et al., 1992), while the Tilt Angle (TDR) and its higher order derivatives like THDR_NTilt normalizes amplitude variations, aiding interpretation across regions of variable anomaly strength (Miller and Singh, 1994; Arisoy and Dikmen, 2013; Nasuti and Nasuti, 2018). Techniques such as the Theta map, Gompertz function (GF), improved logistic (IL) filter, TGAS and other normalized gradient operators aim to improve stability and reduce amplitude bias (Verduzco et al., 2004; Wijns et al., 2005; Alvandi and Ardestani, 2023; Pham et al., 2020; Pham et al., 2024). In addition, wavelet-based approaches address multi-scale decomposition, allowing structural information to be analyzed across depth-sensitive spatial scales (Mallat and Zhong, 1992; Daubechies, 1992; Moreau et al., 1997). Wavelet-based techniques have been successfully applied to boundary delineation in potential-field data (Dwivedi and Chamoli, 2021; Dwivedi et al., 2023), tsunami detection (Lockwood and Kanamori, 2006; Chamoli et al., 2010; Telesca et al., 2013) and different time series analysis (Chamoli et al., 2007; Yemets et al., 2025).

In this study, edge-detection filters are applied to a synthetic prismatic model to delineate subsurface structural boundaries. The results are systematically compared, and their performance is quantitatively evaluated to assess the robustness and reliability of each method under noisy conditions. The techniques are further applied to free-air gravity anomaly data from the Laxmi Basin, Arabian Sea, to test their robustness and delineate the tectonic boundaries.

A synthetic model (M) comprising multiple prismatic sources of varying dimensions and depths is constructed to replicate the key geological scenarios of interest. The model parameters are summarized in Table 1 and illustrated in Figure 1a. Magnetic data are generated on 8000 m × 8000 m grid with a sampling interval of 20 m. The extension of all the prisms vary laterally. In this configuration, prisms P2, P3 and P4 have strike azimuth of −45⁰ and 60⁰ and 45⁰. The prisms P4, P6, and P7 simulate relatively shallow sources and the combined presence of prisms P4 - P5 accounts for interference effects arising from closely spaced sources. In contrast, prism P1, P3, P8, and P9 represents an isolated deep seated sources. All prisms are assigned induced magnetization with an inclination of +90⁰ and declination of 0⁰. The total magnetic anomaly is computed for nine prismatic sources (P1-P9) positioned at different depths with the edge correspond to upper surface of the source (red dashed line) (Figure 1b). This synthetic configuration facilitates the evaluation of the methods under realistic conditions involving depth-dependent responses and mutual interference between nearby bodies.

I assess the robustness of THDR, TDR, ASA, THETA, THDR_NTilt, GF, IL, and WSE using a synthetic magnetic anomaly generated from the prismatic model shown in Figure 1a. The edges identified by these methods are presented in Figure 2. The comparison begins with a qualitative visual assessment, consistent with a standard practice in geological interpretation. Most edge filters perform well in delineating the shallow sources (P4, P6, and P7) and have limitations in resolving the edges due to oblique prisms P2 and P3. These methods show restrictions in resolving the edge of P5 due to interference with P4, except for THDR, TDR, THDR_NTilt, IL, and WSE. The boundary estimated for P4 using TDR appears significantly broader than the actual model boundary, indicating reduced spatial localization accuracy (Figure 2b). The ASA demonstrates limited capability in resolving the deeper sources (P1, P8, and P9) due to the decrease in anomaly amplitude with depth (Figure 2c). The edges of all sources identified from the THETA map appear broader compared to the actual boundaries (Figure 2d). The THDR_NTilt estimates edges (black dashed lines) that coincide with the actual boundaries for P4, P6, and P7 but also produces numerous spurious edges near the map boundaries (Figure 2e). The GF method shows limitations in delineating the edges of P5, P6, and P7 (Figure 2f). The IL method does not perform well in delineating the edges of P2 and P3 and generates additional false edges similar to THDR_NTilt (Figure 2g). The WSE method employs the Daubechies wavelet of order one (db1) at decomposition level 1. After calculating the mean, total and relative wavelet energy, it clearly delineates the boundaries of the prismatic sources, except for P2 and P3 (Figure 2h). The source boundaries are clearly identified by relatively high entropy values, reflecting behaviour similar to disorder in physical systems (Dwivedi et al., 2023).

To quantitatively assess performance, root mean square (RMS) error is computed along selected profiles (L1, and L2) that intersect the model source boundaries. In Figure 3, the true edges (vertical dashed lines) are compared with the edge signatures produced by these techniques. A noticeable lateral shift in the detected signatures is observed for the oblique prisms P2 and P3 across all methods along L1 (Figure 3a), and a similar shift is evident for P4 along L2 (Figure 3b). The RMS error for all the methods are shown in Figure 4. Among these methods, WSE and IL perform better for edge enhancement than other techniques.

In general, noise from various sources affects data quality during acquisition and influences the reliability of interpretation. Thus, the noise robustness of these techniques are evaluated by adding Gaussian noise with zero mean and a standard deviation of 2 nT to the total magnetic anomaly of the 3D synthetic model, following procedures commonly adopted in previous edge-detection studies (Arisoy and Dikmen, 2013; Dwivedi et al., 2023). Figure 5 illustrates the delineation of edges obtained using these methods after the addition of noise. The Theta map, THDR_NTilt, and IL filters perform poorly in detecting edges of lateral sources. In contrast, the THDR, TDR and ASA methods effectively identify the edges of sources P4, P6, and P7, but they are less successful in delineating the boundaries of deeper sources such as P1, P8, and P9. The GF filter also shows limitations, struggling to resolve the boundaries of P6 and P7, as well as those of the deeper sources. Visual inspection indicates that the WSE peaks clearly demarcates the upper surface edge of the P4, P6, and P7 and vice versa for deep sources P1, P8, P9. Overall, these results indicate that the sensitivity of edge-detection methods decreases with increasing source depth.

For further comparison among these methods, two profiles (L1 and L2) are extracted from the edge characterization shown in Figure 5. It is evident that the THETA and THDR_NTilt produce false edge peaks for P2, P3, P6, and P7 along profile L1. In contrast, the THDR, TDR, ASA, GF, IL, and WSE methods accurately delineate the edges of P6 and P7, although they generate multiple false peaks for the oblique prisms P2 and P3. Along profile L2, none of the methods are able to resolve the edges of P1, P5, and P8, with the exception of P4, whose boundaries are identified by THDR, TDR, ASA, GF, IL, and WSE (Figure 6). These observations highlight that while some methods are effective for certain sources, their performance is strongly dependent on source geometry and orientation. The calculated RMS errors for the edge detection methods, excluding THETA and THDR_NTilt are ∼75 m (THDR), ∼64 m (TDR), ∼46 m (ASA), ∼92 m (GF), ∼112 m (IL), and ∼41 m (WSE), with WSE exhibiting the lowest error.

The Arabian Sea records the breakup of India from the Seychelles microcontinent and the early stages of ocean basin formation during the Late Cretaceous-Paleogene (Krishna et al., 2006; Bhattacharya and Yatheesh, 2015). Within this region, the Laxmi Basin forms a deep seafloor depression bordered by the Laxmi Ridge (LR) to the west and the Indian continental margin to the east. This Basin contains the chain of the Raman (R), Panikkar (P) Seamount, and Wadia (W) Guyot in its axial region. These morphotectonic elements are clearly visible in the bathymetry map as prominent structural highs (Figure 7a), reflecting a history of rifting, extinct seafloor spreading, and later hotspot-influenced volcanism (Yatheesh, 2020; Savio et al., 2022). The bathymetry data is obtained from the GEBCO (General Bathymetric Chart of the Oceans) database using the GEBCO 2025 global grid (https://download.gebco.net/, last accessed: 09 November 2025). Likewise, the Free-Air Anomaly (FAA) data is extracted from the Bureau Gravimétrique International (BGI) using the WGM-2012 global gravity model (https://bgi.obs-mip.fr/data-products/outils/wgm2012 visualizationextraction/#/form-wgm2012-model, last accessed: 09 November 2025) to generate the FAA map of the study region (Figure 7b). The bathymetry patterns of the basin and surrounding highs correlate strongly with signatures in FAA data. Smooth deep basin regions correspond to lower gravity, while seamounts and volcanic ridges produce positive anomaly highs.

The edge detection methods are applied to evaluate their ability to delineate the structural boundaries of major seafloor features. The THDR, TDR, and ASA methods successfully delineate the boundaries of the Laxmi Ridge and other geological features (R, P, and W), showing good agreement with the trends observed in the bathymetry map. However, these methods also produce numerous spurious boundaries due to their sensitivity to high-amplitude responses (Figures 8a–c). In contrast, THETA and THDR_NTilt provide unsatisfactory results for delineating geological boundaries (Figures 8d,e). The GF filter identifies the boundary of the Laxmi Ridge, but the resulting boundary appears relatively broad compared to the bathymetry map and fails to resolve the R, P, and W structures (Figure 8f). The IL filter produces prominent structural edges accompanied by several artifacts, which hinder the reliable interpretation of structural boundaries (Figure 8g). The zones of maximum entropy response correspond closely to the boundaries of Laxmi Ridge, Raman and Panikkar Seamounts, Wadia Guyot, and other geological structures, as indicated by the black lines (Figure 8h). The identified boundaries by WSE method matches with the bathymetry map of the study region. This demonstrates that the WSE based edge extraction preserves the spatial continuity of geological structures, allowing sharp transitions to be mapped more reliably. The WSE approach produces a more coherent and geologically meaningful characterization of crustal boundaries and enhancing our ability to interpret basin architecture.

This study presents a comprehensive evaluation of several edge enhancement methods applied to gravity and magnetic data, using both a synthetic model and a real marine case study from the Laxmi Basin, Arabian Sea. The synthetic results show that most of the methods effectively highlight shallow sources but suffer from reduced localization accuracy, and sensitivity to interference. Quantitative RMS error analysis under noisy conditions indicates that WSE outperforms other methods, particularly for shallow prismatic sources, yielding significantly lower RMS errors and cleaner edge responses. Application to free air gravity data further demonstrates that the THDR, TDR, and ASA methods produce multiple geological edges owing to their sensitivity to high-amplitude responses; however, the THETA and THDR_NTilt methods fail to delineate boundaries effectively. The GF method identifies the boundary of the Laxmi Ridge, but the resulting boundary appears relatively broad when compared with the bathymetry map. The IL filter produces prominent edges accompanied by artifacts, which complicates reliable interpretation. In contrast, WSE reliably delineates major geological features, including the Laxmi Ridge, and seamount boundaries that correlate well with bathymetry map. The comparative framework presented here provides practical guidance for selecting an appropriate edge enhancement method in both synthetic studies and realistic scenarios.