Matt Duyst

I engineered the team's shared covariate extraction tooling, a Python client that queries five authoritative geospatial sources through one interface and consolidates what were previously fragmented workflows into a single maintained codebase. I benchmark biomass and structural vegetation architectures spanning deep learning, ensemble methods, and process-based ecological models, with outputs supporting quantification under the Regenerative Standard, VM0042, and CAR-SEP, and extend them into counterfactual baselines that reconstruct expected productivity under business-as-usual management, a required step under each methodology. Ranch-level stratification is designed around SEP v2, where clustering geophysically similar ground directly reduces the sampling uncertainty that drives credit deductions. This work is under active development and unpublished; publications are forthcoming.

As one of three scientists on the Science team, reporting to the co-founder and CTO, I led a quantitative assessment of flare destruction removal efficiency across the Bakken and Permian basins. The result covered 42 flare systems across 71 observations, including 27 revisits, and was published as first author and presented at SPE ATCE 2024 (SPE-221067-MS), contributing to enhanced leak detection and repair programs in two major shale basins. I trained a fine-tuned YOLOv8 detector to classify wellheads, compressors, storage tanks, and flare stacks from aerial imagery for autonomous site characterization, managing the full cycle from annotation to held-out validation, and built the client-facing and internal visualization suites in Python and GIS platforms that communicated spatial emissions distributions across survey campaigns.

In the Lee Lab I investigated the relationship between urbanization, rice cultivation, and methane across China's Yangtze River Delta, home to more than 150 million people: multi-decadal land cover classified from 30 m Landsat in Google Earth Engine, with urban extent from the GAIA dataset and paddy rice from the Phenology and Pixel-Based Paddy Mapping algorithm, set against Sentinel-5P TROPOMI column methane and tested for spatial clustering with Getis-Ord Gi* and Global Moran's I. The thesis proposed a DeepLabV3+ network to extend the emissions record before the satellite era. A 2026 reproduction of the analysis does not reach the thesis conclusion: no land-cover model outperforms a spatial null at 0.25 degrees, and both the original thesis and the reproduction are published together with an errata.

The 2022 work applied feature importance and seasonal decomposition to isolate the environmental drivers of flux and built an ensemble of LightGBM, SARIMAX, Prophet, and XGBoost over OLS regression, reported at the time to explain 87 percent of variance with forecasts extended to 52 weeks. The project was rebuilt from source in 2026 as a monthly study of the full 2009 to 2024 record, forecasting one to twelve months ahead against seasonal benchmarks and projecting the fitted relationships back two decades. The rebuilt finding is narrower and more useful: the seasonal cycle repeats reliably, the size of each season varies without trend, and none of the recorded covariates predict that variation once seasonality is accounted for.

The coursework covered the full geospatial stack: supervised and unsupervised image classification, spectral index analysis, SAR interpretation, DEM construction, spatial interpolation and autocorrelation, SQL spatial querying, cartographic design, and predictive environmental modeling. The undergraduate thesis measured surface vegetation change across the Mendocino Complex, at the time the largest wildfire in California history, from Landsat 8 pre- and post-fire scenes in Google Earth Engine using NDVI and NBR, and was presented at the 2019 Los Angeles Geospatial Summit, hosted by the USC Spatial Sciences Institute.