I build machine-learning models that combine data from satellites and drones, atmospheric and ground-based sensors, and field measurements to detect, measure, and predict environmental conditions. What ties the work together is not the subject but what rides on the result: each measurement feeds a real decision, so it has to be right where it counts and honest about where it is not. The difficulty is that two sources rarely agree. Each was built for its own purpose, at its own resolution and revisit time, dense in some places and sparse in others, and everything downstream rests on first reconciling them into one measurement you can rely on.
The problem I keep returning to is what happens when a model leaves the ground it was trained on and has to answer for somewhere it has never seen: a different region, a later year, an ecosystem unlike anything in its training data. I believe a measurement earns trust by being checkable, not by the reputation of whoever produced it, so the data, the code, and the models should be there for anyone to run, verify, and improve.
This first drew me in as an undergraduate, mapping how severely California's largest wildfire had burned by comparing satellite images of the same ground before and after. The window for that measurement closes fast, since new growth on burned ground starts erasing the signal within weeks, so the comparison had to be caught days before the fire and a week after. At Yale it became the center of the work: I reconstructed two decades of methane over China's Yangtze River Delta from the single year when satellite methane, urban extent, and rice paddies were all observed together, training a segmentation network on that overlap and running it backward through the decades before the satellites existed. Alongside the thesis, a course project had me flying thermal and RGB sensors over a solar array, checking how closely the imagery tracked what the panels underneath were actually producing. At a peatland in Minnesota, I compared four forecasting methods against seasonal benchmarks on a single tower's record and projected the fitted relationships back into the decades before measurement began, where the water table the projection leans on sits nearly a third of a meter past anything the fit had seen. At flare sites across the Bakken and Permian, I trained a model to find oil-and-gas equipment from the air and measured how much methane each flare failed to burn.
Currently, at continental scale, I predict how much grass grows across US rangelands by combining process-based modeling with machine learning. A light use efficiency model converts sunlight into growth, learning how efficiently each site does that rather than assuming one rate everywhere, and a random forest trained on the model's residuals corrects the errors that show up predictably in conditions the physics handles poorly. Alongside it I build the counterfactual: it holds a site's efficiency at what it was before a management change and runs forward on the weather that actually happened, so the gap between the two tracks shows what the management did rather than what the year did.
The subject changes and the method changes, but it's the same problem underneath.
Published work
Duyst, K. M., & Smith, B. J. (2024). Methane Destruction Removal Efficiency Assessment in US Shale Basins Using UAV-Based TDLAS Measurements. SPE Annual Technical Conference and Exhibition, New Orleans, LA. https://doi.org/10.2118/221067-MS
Repositories
Peatland-Flux-Forecasting: Monthly methane and carbon dioxide flux at a Minnesota peatland tower, 2009 to 2024. The seasonal cycle here is predictable at every horizon tested, so the open question is the size of each season, which varies more than threefold between years and which nothing measured at the site predicted.
Yale-Masters-Thesis. The methane reconstruction above: data, model, and figures.
Carbon-Trading. Built for the Global Stocktake Climate Datathon, where it took an honorable mention: European emissions regressed against allocated allowances, projected forward under different rates of allowance decay.
UAV-PV-Assessment. Thermal and RGB orthomosaicking over Yale's West Campus solar array, checked against measured panel output.
Background
Yale University, M.E.Sc., School of the Environment: machine learning, remote sensing, and spatial statistics. UCLA, B.A. in Geography and Environmental Studies, with minors in Geospatial Information Systems and English.
mappingwithmatt.com · LinkedIn · matthew.duyst@gmail.com