Research
Three Hawaiʻi reefs designed for continuous monitoring, with Indigenous research partners and scientific staff structured as co-investigators on every dataset.
The work begins with diagnosis.
Most reef interventions go after symptoms. Coral gets transplanted back into water that is still polluted. Sunscreen gets banned before anyone measures whether the herbicide running off a nearby golf course is the larger stressor. The work happens in earnest. What is usually missing is the continuous, source-attributable, enforcement-grade evidence that says which of those pollutants is actually killing the reef in front of you.
Pillar 01 is designed to close that gap. Each dataset is structured to feed one of four downstream decisions, and each measurement is structured to double as a labeled training example for the AI models in Pillar 02.
Identify the specific discharge, watershed, and transport pathway behind each pollutant at each reef. Without source attribution, remediation is permanent symptom management.
Coral stress shows up in fluorometry data and combined sensor signals 5 to 14 days before visible bleaching. That is enough lead time to pre-deploy filter arrays while the intervention still has a chance to matter.
Continuous telemetry and year-over-year photogrammetry are designed to measure how much the filters and the biorefinery improve water quality and coral cover at each reef.
Every measurement is structured to label training data for AI models meant to compress the five-year Hawaiʻi validation into a one-year deployment standard at the next reef.
Synoptic coverage of what ground sensors cannot see.
Six sondes can characterize three reef zones, but a 74-hectare invasive algae bloom, a submarine groundwater plume kilometers from any sensor, and thermal stress at the basin scale will all slip past that grid. Satellite imagery picks up what the ground network is too sparse to see.
Acquisition is phase-gated so that capital follows validation. The program is structured to begin on free public imagery, graduate to commercial constellations once biorefinery revenue is paying for them, and acquire owned UAV hyperspectral hardware only after Phase 2 has proven its value.
FAI for invasive algae detection. SDB for shallow-water bathymetry. CIre as a chlorophyll and submarine groundwater discharge proxy. NOAA DHW for bleaching stress, with the published threshold of 8°C-weeks corresponding to greater than 50% coral mortality.
Tying pollution back to its source.
A filter array at the reef face is reactive remediation. The durable fix is to close the discharge upstream, which requires evidence strong enough to support regulatory action against the party responsible. Attribution is the bridge between measurement and policy.
The method combines four independent lines of evidence. A source assignment that lines up across all four can survive cross-examination by a regulatory agency or a court.
“Hawaiʻi's 88,000 cesspools discharge approximately 5.1 million kg of nitrogen per year statewide.”
Street et al. (2008) measured DIN fluxes from submarine groundwater of 140 to 180 mmol/m²/day across Kāneʻohe Bay transects, roughly four times the surface water input. The dominant transport pathway is subsurface, a fact most stormwater-focused interventions overlook.
Continuous water-quality telemetry across three reef systems.
The sensor network is designed to do three things. First, it establishes a per-reef baseline so that anomalies (storm-pulse nutrient spikes, oxybenzone peaks at peak tourist hours, oxygen sags in stagnant water) are detectable in real time. Second, it cross-validates satellite indices against in-situ measurements, which is what allows free Sentinel and MODIS data to substitute for expensive commercial constellations. Third, it produces the high-temporal- resolution training data the AI models need in order to learn from a small number of reefs and predict on a much larger set.
YSI EXO3 with titanium housing rated to IP68. Anti-fouling automated wiper for long-deployment integrity. Fourteen parameters logged simultaneously.
DO, pH, temperature, conductivity, turbidity, chlorophyll-a, phycocyanin, CDOM, ORP, depth, and total dissolved gas.
Verizon LTE primary with Iridium satellite backup. Data processed under the QARTOD four-level QC scheme before publication.
BlueROV2 Heavy with on-board AI inference.
Sondes measure water chemistry. Satellites cover broad geographic extent. Neither captures the photosynthetic state of an individual coral colony, which is the level at which bleaching actually begins. The robot fills that gap: PAM fluorometry measures coral stress at the colony scale, and stereo photogrammetry produces a centimeter-resolution 3D record of coral cover that can be differenced year over year to quantify gain or loss.
On-board inference is what makes everything else actionable. Real-time benthic segmentation at 5 fps is designed to let stress indicators trigger filter pre-deployment during the dive itself, before the response window has closed. The published 5 to 14 day pre-bleaching window only matters if the analysis runs at the reef rather than at the desktop weeks later.
Survey runs are designed to follow a back-and-forth lawn-mower pattern at 1.5 m altitude and 0.5 m/s, with 80% overlap between image tiles for clean photogrammetry. Stress signals from PAM fluorometry, DHW, turbidity, dissolved oxygen, and chlorophyll-a are fused to feed the early-warning model in Pillar 02.
Analytical chemistry built and operated in-house.
Source attribution that holds up in regulatory or legal proceedings requires an unbroken chain of custody and full control over the analytical method. Outsourcing the work to commercial labs adds weeks of turnaround and obscures the QC process. Both effects break the tight loop from sensor anomaly to lab confirmation to enforcement action.
The lab will be operated by an analytical chemist with prior enforcement laboratory experience. External NELAP-accredited laboratories will be used only for enforcement submissions that require third-party authentication. LIMS is SENAITE, an open-source platform to be configured and maintained by the in-house engineering team.
Nutrients via ion chromatography and heavy metals via inductively coupled plasma mass spectrometry, with column switching for combined runs.
PAHs, total petroleum hydrocarbons, organochlorine pesticides, and PFAS screening.
Composition checks for incoming algae feedstock and process monitoring across the biorefinery.
Acid digestion of solid biological and sediment matrices ahead of metals quantification.
Total and dissolved organic carbon measurements supporting source attribution and biorefinery process water characterization.
Rapid metals screening at intake, before samples enter the chain-of-custody pipeline.
Knowledge held in partnership.
Hawaiian reefs were managed sustainably for centuries before colonial-era land use changes. Modern datasets do not contain the pre-degradation baseline, the seasonal current patterns, or the species distributions you need in order to set a meaningful restoration target. That information lives with kūpuna and Indigenous practitioners. A restoration goal anchored to the past twenty years of degraded reef merely re-creates the damaged state.
Indigenous ecological knowledge is structured to enter the AI models as spatial priors, set the timing of robot surveys, and validate model predictions before they reach the public dataset. The framework draws from the CARE Principles for Indigenous Data Governance and from Etuaptmumk, the Two-Eyed Seeing approach to integrating Indigenous and scientific knowledge.
Relationship-building precedes any data collection so that community-controlled data sovereignty agreements are established before fieldwork begins.
Seasonal current patterns, historical reef baselines, and species distributions are designed to be encoded as quantified prior probabilities, giving Indigenous observation real numerical weight in every model output.
Robot survey timing is structured to follow Indigenous calendars governing reef use. The cultural calendar comes first; the engineering schedule fits around it.
Predictions and CV training taxonomies are structured for review by Indigenous co-investigators before models are released to the public dataset.