Research

The study asks whether automated computer vision can credibly stand in for hand-tallied pollinator visitation surveys for two federally endangered coastal plant species — Ventura marsh milk-vetch (Astragalus pycnostachyus var. lanosissimus) and salt marsh bird's-beak (Chloropyron maritimum ssp. maritimum) — at the NCOS restoration site in Goleta, California.

Across nine weeks of the summer 2025 field season, two Panasonic Lumix cameras and a long-run "Doug" camera were deployed across four yarn-marked quadrants for three one-hour shifts daily — morning, mid-day, afternoon — with site order randomized and weather, wind speed, and on-site pollinator activity logged for every session. The result was more than 200 hours of raw video and a hand-built ground-truth log: timestamped time | quadrant | bumblebee count entries keyed to a numbered grid, against which every detection downstream is checked.

The training pipeline I built is the core of my contribution. Frames were extracted from raw video at 1–2 second intervals using OpenCV; a 165-frame subset (101 frames in the first pass, 64 in a second) was hand-annotated in Matplotlib, with every clearly visible Bombus getting a manually drawn bounding box and each annotation cross-checked against the ground-truth count log. Annotations were split 70 / 20 / 10 across train / validation / test.

A YOLOv8 detector (PyTorch · Ultralytics) was trained iteratively on this set — a 50-epoch baseline, a 150-epoch v2, and a 200-epoch v23 — with each iteration evaluated on withheld videos. The 150-epoch model reached mAP50 = 0.745, precision = 0.878, and recall = 0.648, with the precision–confidence curve plateauing at 1.00 above ~0.39 confidence; on a withheld test video the model identified bees at 100% detection rate at 1.000 confidence. A follow-on run is comparing this manually annotated dataset against a synthetic-overlay approach — high-resolution bumblebee cutouts composited onto vegetation stills with auto-generated bounding boxes — with a manuscript on the comparison in preparation under Dr. Katja Seltmann and Dr. Chris Evelyn at the Cheadle Center.

The thesis asks whether calcium carbonate from discarded bivalve shells — sourced through a partnership with Hendry's Boathouse, a local Santa Barbara seafood restaurant — can meaningfully counter ocean acidification when reintroduced into marine systems as a ground material. The framing draws on indigenous shell-stewardship practices from the Bodega Bay community alongside contemporary aquatic chemistry.

Ground bivalve shells were dosed into reef-tank treatments at UCSB and tracked over time against control tanks for pH, dissolved oxygen, conductivity, and salinity using YSI and pH meters provided by CCBER. A MANOVA evaluated the joint treatment effect across response variables.

Findings: ground shell additions did raise pH, supporting the alkalinity-enhancement hypothesis — but a measurable decrease in dissolved oxygen was observed alongside the pH rise, indicating a real biogeochemical trade-off rather than a clean win. The thesis ends by recommending future work on shell dispersal methods that maximize the alkalinity benefit while characterizing and mitigating the DO drawdown — particularly mechanistic and long-term studies.

A 141-page graduate-group investigation that synthesizes field data, environmental chemistry, and policy analysis into long-term remediation recommendations for a pollution-stressed river system. Group authorship is attributed inside the document; my contribution centered on the heavy-metals hydrochemistry chapters, the spatial figures, and carrying a consistent technical-writing voice across the report.

The Pilcomayo manuscript in preparation (Card 01) is my solo first-author extension of this group work — taking the heavy-metals analysis I led for the report and developing it into a standalone, peer-review-targeted paper.