Advanced Sensor Technologies and AI Applications for Autonomous Site Characterization by Robot Platforms

<p dir="ltr">Robust high-throughput soil monitoring with autonomous robots demands coordinated advances in (i) the selection of modalities of observation based on the detection limit, dynamic range, and matrix interference; (ii) the rigorous benchmarking of portable instruments against regulatory laboratory methods to quantify accuracy, precision, and decision thresholds; and (iii) the integrated, data-driven sampling and modeling to automate in situ acquisition, fusion, and interpretation of in situ while reducing field time and cost. This dissertation addresses these needs through contaminant-specific studies, total petroleum hydrocarbons (TPH), asbestos, and heavy metals, united by a common field interface that includes a shared data schema, uncertainty reporting, and EPA-aligned stopping criteria. </p><p dir="ltr">For petroleum-impacted soils, this study evaluates two real-time TPH screening methods—thermal desorption–gas chromatography (TD–GC) and portable near-infrared (NIR) spectroscopy—along with their compatibility with robotic platforms. Both methods were compared with EPA Method 8015C (sonication and GC–FID) using soils from two crude oil impacted sites. TD–GC delivers superior analytical accuracy and robust cross-site calibration, making it the preferred option for definitive quantitative assays, although it requires more extensive engineering for robotic deployment. In contrast, NIR combines compact dimensions, low power consumption, and rapid preparation-free analysis, producing a screening-level precision of approximately 20-40% error. Although this error margin limits its use for regulatory quantification, the ease of integration and speed of NIR make it well suited for first-pass site triage, reserving TD-GC for high-precision follow-up analyzes. For heavy metals, a portable XRF validation-and-correction protocol benchmarked against ICP-MS provides bias-aware concentration estimates and uncertainty-mapped decisions appropriate for on-robot triage.</p><p dir="ltr">For rapid chrysotile screening in soils, a two-stage hybrid workflow couples polarized light microscopy (PLM) with deep learning instance segmentation and rotation pair optical confirmation. Stage one (Mask R-CNN, ResNet-50-FPN) operates at a strict 0.9 threshold to control false positives per field of view (precision 0.80, recall 0.55, F1 = 0.65). Stage two registers paired 0 ◦/90◦ views under a first-order red plate and applies a color-difference mask, raising the recall to 0.74 while maintaining precision at 0.82 (F1 = 0.78). A hierarchical decision model (FOV → slide → specimen) links per-FOV performance to sample-level predictions under multi-FOV/multislide aggre.g.ation (e.g., K = 20 FOV/slide, S = 3 slides/specimen), achieving approximately 99.2% positive predictive value and enabling near real-time auditable presence/absence decisions in complex matrices. </p><p dir="ltr">Finally, a Gaussian process adaptive sampling framework with EPA-compliant stopping criteria is introduced to ensure ≥ 75% accuracy/precision coverage and convergence before terminating sampling. Across 8,000 synthetic contamination fields that span extent and heterogeneity, adaptive sampling more easily achieves compliance coverage and accurately validates maximum concentrations for localized or highly heterogeneous plumes, while regular grids are more robust in mean square error and Earth Mover distance under broad contamination or high noise. These results delineate regimes for selecting adaptive policies versus grids in site characterization. </p><p dir="ltr">Collectively, these contributions establish reusable design patterns for robot-enabled environmental sensing and sampling: a decision framework for TPH sensor selection, a rapid and auditable asbestos detection pipeline, and a compliance-aware adaptive sampling strategy. The resulting workflows provide higher-resolution contaminant maps with fewer samples, shorter campaigns, and reduced operator burden, while remaining extensible to additional contaminants (e.g., metals, PFAS, microplastics) and platforms (human-led teams, aerial drones, hybrid operations), offering a scalable, regulatory-aligned paradigm for next-generation environmental site characterization.</p>