Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Building Envelope Inversion by Data-driven Interpretation of Ground Penetrating Radar

Ahmed Nirjhar Alam, Wesley Reinhart, Rebecca Napolitano

TL;DR

This study tackles the challenge of diagnosing building envelopes with ground-penetrating radar by reframing GPR interpretation as two classification tasks: detecting vertical stud presence and classifying wall types. A data-driven framework combines baseline ML models with feature minimization and sparse neural networks using L0 regularization to produce accurate, interpretable inferences from complex, overlapping radar signals. SHAP analyses and wave-propagation mapping validate that the model-inferred features correspond to physically meaningful dielectric interfaces, enabling physically grounded interpretation and robust performance under limited data. The results establish a foundation for physically interpretable, data-efficient GPR inversion of wall assemblies and set the stage for extending to defect localization and broader wall configurations. The approach offers practical implications for non-destructive envelope diagnostics where interpretability and edge-computing feasibility are important.

Abstract

Ground-penetrating radar (GPR) combines depth resolution, non-destructive operation, and broad material sensitivity, yet it has seen limited use in diagnosing building envelopes. The compact geometry of wall assemblies, where reflections from closely spaced studs, sheathing, and cladding strongly overlap, has made systematic inversion difficult. Recent advances in data-driven interpretation provide an opportunity to revisit this challenge and assess whether machine learning can reliably extract structural information from such complex signals. Here, we develop a GPR-based inversion framework that decomposes wall diagnostics into classification tasks addressing vertical (stud presence) and lateral (wall-type) variations. Alongside model development, we implement multiple feature minimization strategies - including recursive elimination, agglomerative clustering, and L0-based sparsity - to promote fidelity and interpretability. Among these approaches, the L0-based sparse neural network (SparseNN) emerges as particularly effective: it exceeds Random Forest accuracy while relying on only a fraction of the input features, each linked to identifiable dielectric interfaces. SHAP analysis further confirms that the SparseNN learns reflection patterns consistent with physical layer boundaries. In summary, this framework establishes a foundation for physically interpretable and data-efficient inversion of wall assemblies using GPR radargrams. Although defect detection is not addressed here, the ability to reconstruct intact envelope structure and isolate features tied to key elements provides a necessary baseline for future inversion and anomaly-analysis tasks.

Building Envelope Inversion by Data-driven Interpretation of Ground Penetrating Radar

TL;DR

This study tackles the challenge of diagnosing building envelopes with ground-penetrating radar by reframing GPR interpretation as two classification tasks: detecting vertical stud presence and classifying wall types. A data-driven framework combines baseline ML models with feature minimization and sparse neural networks using L0 regularization to produce accurate, interpretable inferences from complex, overlapping radar signals. SHAP analyses and wave-propagation mapping validate that the model-inferred features correspond to physically meaningful dielectric interfaces, enabling physically grounded interpretation and robust performance under limited data. The results establish a foundation for physically interpretable, data-efficient GPR inversion of wall assemblies and set the stage for extending to defect localization and broader wall configurations. The approach offers practical implications for non-destructive envelope diagnostics where interpretability and edge-computing feasibility are important.

Abstract

Ground-penetrating radar (GPR) combines depth resolution, non-destructive operation, and broad material sensitivity, yet it has seen limited use in diagnosing building envelopes. The compact geometry of wall assemblies, where reflections from closely spaced studs, sheathing, and cladding strongly overlap, has made systematic inversion difficult. Recent advances in data-driven interpretation provide an opportunity to revisit this challenge and assess whether machine learning can reliably extract structural information from such complex signals. Here, we develop a GPR-based inversion framework that decomposes wall diagnostics into classification tasks addressing vertical (stud presence) and lateral (wall-type) variations. Alongside model development, we implement multiple feature minimization strategies - including recursive elimination, agglomerative clustering, and L0-based sparsity - to promote fidelity and interpretability. Among these approaches, the L0-based sparse neural network (SparseNN) emerges as particularly effective: it exceeds Random Forest accuracy while relying on only a fraction of the input features, each linked to identifiable dielectric interfaces. SHAP analysis further confirms that the SparseNN learns reflection patterns consistent with physical layer boundaries. In summary, this framework establishes a foundation for physically interpretable and data-efficient inversion of wall assemblies using GPR radargrams. Although defect detection is not addressed here, the ability to reconstruct intact envelope structure and isolate features tied to key elements provides a necessary baseline for future inversion and anomaly-analysis tasks.
Paper Structure (44 sections, 2 equations, 28 figures, 12 tables)

This paper contains 44 sections, 2 equations, 28 figures, 12 tables.

Table of Contents

  1. Abstract
  2. Introduction
  3. Data and Methods
  4. Data
  5. Preprocessing
  6. Dataset Challenges
  7. Unpredictable Far-field Variations
  8. Complexity
  9. Methods
  10. Results
  11. Stud Detection
  12. Information Minimization
  13. Sample Minimization
  14. Feature Minimization
  15. Sparse Neural Network
  16. ...and 29 more sections

Figures (28)

  • Figure 1: House floor plan with inset images of building envelope during construction.
  • Figure 1: Schematic of exterior wall with window wells that engulf portions of the exterior wall with rock and soil. As shown in the figure, exterior wall scan segments can have varying backgrounds. For instance, in this wall, H2's background is completely soil, whereas H1 and H3's background consists of a varying rock/soil mixture gradually transitioning to air (space) in the middle of the wall.
  • Figure 2: Each wall in the dataset is assigned an alphabetic label, and individual wall segments are indexed numerically; the resulting alphanumeric identifier uniquely specifies each GPR scan. Panel (a) shows exterior wall segments with partial rock background (H1 and H3) and soil background (H2). Panel (b) shows interior wall segments, including G1, as well as G2 and G3, which are located near electrical outlets. Panel (c) shows an exterior wall segment (I1). Panel (d) shows an interior wall segment (E2) containing an embedded column. Unless otherwise noted, wall segments I1 (exterior) and G3 (interior) are used in the training set for all classification and interpretability tasks considered in this study, with the exception of the wall-background classification task.
  • Figure 2: Tuning of the threshold fraction that defines stud. The red dotted line is the standard stud width of 1.5 inches. The cut-off fraction was chosen so that the mode of the stud width distribution coincided with 1.5 inches as closely as possible. This value of the cut-off fraction also yields a mean stud width close to 1.5 inches.
  • Figure 3: Insulation layers (and stud depth) are four inches for interior walls and six inches for exterior walls. Interior walls are drywalled on both sides. Exterior walls consist of drywall on the interior side (where the scans occur) and cement walls on the exterior. A vapor barrier separates the insulation and cement layers.
  • ...and 23 more figures