Reservoir-enhanced Segment Anything Model for Subsurface Diagnosis

TL;DR

Res-SAM addresses the challenge of subsurface anomaly detection in GPR data under limited labeled samples by fusing visual segmentation with dual-directional wave-dynamics analysis. It builds a normal-dynamics feature bank from non-target patches using a Dual-Directional Echo State Network (2D-ESN) and uses Segment Anything Model (SAM) to propose candidate regions, which are refined by local 2D-ESN analysis to produce complete anomaly boundaries and category labels. Real-world experiments on cement and asphalt GPR data show high AUROC and F1 scores across prompt configurations and demonstrate superior anomaly delineation and categorization compared to baselines, while requiring minimal non-target data and little offline training. The approach offers a scalable, resource-efficient solution for rapid urban subsurface safety monitoring with reduced manual effort and computational cost.

Abstract

Urban roads and infrastructure, vital to city operations, face growing threats from subsurface anomalies like cracks and cavities. Ground Penetrating Radar (GPR) effectively visualizes underground conditions employing electromagnetic (EM) waves; however, accurate anomaly detection via GPR remains challenging due to limited labeled data, varying subsurface conditions, and indistinct target boundaries. Although visually image-like, GPR data fundamentally represent EM waves, with variations within and between waves critical for identifying anomalies. Addressing these, we propose the Reservoir-enhanced Segment Anything Model (Res-SAM), an innovative framework exploiting both visual discernibility and wave-changing properties of GPR data. Res-SAM initially identifies apparent candidate anomaly regions given minimal prompts, and further refines them by analyzing anomaly-induced changing information within and between EM waves in local GPR data, enabling precise and complete anomaly region extraction and category determination. Real-world experiments demonstrate that Res-SAM achieves high detection accuracy (>85%) and outperforms state-of-the-art. Notably, Res-SAM requires only minimal accessible non-target data, avoids intensive training, and incorporates simple human interaction to enhance reliability. Our research provides a scalable, resource-efficient solution for rapid subsurface anomaly detection across diverse environments, improving urban safety monitoring while reducing manual effort and computational cost.

Figures (10)

• Figure 1: (a) GPR typically consists of a host and an antenna. The antenna emits EM waves into the ground, receiving reflections from subsurface anomalies. These signals are transmitted to the host for processing and visualized as B-scan images. (b) Left side shows a B-scan with horizontal and vertical axes representing detection position and depth-related time windows. Right side shows a waveform that correlates gray values with wave amplitude variations. In this paper, "GPR data" specifically refers to GPR B-scan data.
• Figure 2: Res-SAM framework includes two phases: Feature Collection and Anomaly Detection. In the Feature Collection phase (#1), normal data patches are extracted from non-target GPR data and fitted using 2D-ESN, with the derived dynamic features collected into a feature bank. In the Anomaly Detection phase (#2), SAM initially identifies candidate anomaly regions. Local patches around each point in the region are fitted similarly using 2D-ESN, deriving features that are compared with the feature bank, with anomaly-associated patches identified and further merged into the final anomaly region.
• Figure 3: Distribution of the collected GPR data frames: type and amount.
• Figure 4: The candidate and final anomaly regions identified via Res-SAM. Each column represents a different type of anomaly (Cavity, Crack, Loose soil, Manhole, and Pipe). The first row shows the original GPR data. The blue dotted boxes highlight the candidate regions determined by SAM. The heatmaps overlaying the images show the anomaly likelihood examined based on the discriminability of local dynamic features, transitioning from blue (low) to yellow (high). The red boxes outline the precisely determined final anomaly regions, obviously more accurately encapsulate the target anomalies. Green and red dots represent positive and negative click prompts, respectively.
• Figure 5: Each row shows an anomaly and the corresponding anomaly regions (the red box) obtained by Res-SAM and three other high-performing methods, supported with two sets of click prompts.