TerraTrace

What it does

30% of potential harvests are lost because we lack a global directory of crops. Agriculture drives deforestation, but opaque supply chains hide bad actors. TerraTrace uses satellite pictures & AI to map every crop and expose deforestation in real time.

Your inspiration

EUDR bars commodities linked to deforestation from entering the EU. Small firms lacking legal resources face $11B in lost exports from compliance burdens. Meanwhile, the same food corporations that drove deforestation up by 14% in just 4 years will hire lawyers to make themselves look like they're following the rules, even as they continue to cut down forests. As small farms exit, harvests shrink. Within ten years, food production could fall by 280 M tons, enough to feed 150M people for a year. Scarcer supply will push prices up, making food less affordable. We need to solve deforestation, but not at the cost of worsening global hunger.

How it works

When a user enters coordinates, TerraTrace retrieves the corresponding Sentinel-2 satellite imagery and extracts the NDVI (a score for how green and healthy plants are) values for each date. Those values form a time-series “signature curve” for the plot. This curve serves as the land's activity tracker throughout the year. As part of our R&D, we assembled a catalog of over 50 time-series “signatures,” each tied to a known crop species. By comparing the user’s curve against this catalog, we identify what most likely is plotted at the given area. For the U.S., we compare this decision with the official information given by the Crop Data Layer Dataset. We use mathematical & ML models to measure how vegetation curves deviate from baseline patterns, indicating deforestation. To determine if it's human-caused, we track wildfire activity in the region. Finally, we leverage large language models to translate technical findings into clear, user-friendly explanations.

Design process

Our initial plan, deploying computer‑vision models to flag deforestation, fell short. The algorithms treated all plots as “green,” so an old‑growth forest looked no different from a commercial plantation. If the system can’t tell the forest from the crops, it can’t even start detecting deforestation. To better understand the vegetation (i.e. how quickly plants rise, thrive and thin out), we started analyzing multispectral satellite data. We compared more than 100 satellite-based plant indicators and kept only the ones that clearly tracked real plant growth and decline. Through testing, we cut the list to ten signals, including but not limited to NDVI, EVI, key red-edge and SWIR bands, that give the best view of plant cover and health. We graphed time-series “growth curves” for each of the ten shortlisted indices across multiple crops and locations. Only NDVI produced a curve whose shape shifted with crop type yet remained consistent for the same crop, regardless of the location. Since we couldn’t find an existing NDVI dataset, we built one. We sliced the planet into 500-m squares, calculated NDVI for every tile, and stored the results in our own database. By caching the data, users don’t have to wait for live real-time calculations, and hence, can get the results immediately.

How it is different

First, accuracy: While working with farmers in Seattle, we found that most forest maps mistakenly classify orchards as natural forest, simply because they look alike. TerraTrace avoids this by analyzing each field’s 9-day NDVI time-series. This is important because if we misidentify the plot at the start, every later deforestation and compliance calculation will be wrong. Second, timeliness: Annual tools like CropScape are not only limited to the United States, they also update just once a year, so they miss mid-season clearing & replanting and provide no detailed timeline of land-use change. TerraTrace updates every 9 days worldwide, giving a clear week-by-week record of how and when the land use changes. Third, cost & compute: Many new AI tools run transformers on full image stacks and need GPUs. TerraTrace uses a lightweight 3rd-order polynomial and a 50 kB reference-curve library, so it runs fast on everyday laptops or phones and keeps costs low for users.

Future plans

We co-designed TerraTrace with Seattle growers to ensure field accuracy and will extend that model worldwide. With our prototype complete, we're now shifting to global rollout, targeting both farmers and regulatory officials (policy makers and border agents). We’re partnering with nonprofits to offer TerraTrace subsidized to farmers. With 6 years of climate activism experience, I’m leveraging my expertise to launch pilot programs with UN agencies that integrate TerraTrace’s nine-day NDVI feeds into border-inspection systems so officers at ports of entry can cross-check importer claims against real-time crop maps and deforestation alerts.

Awards

TerraTrace is a U.S. patent-pending technology (503289-US01, filed January 2025). It was awarded runner-up at the ACM Computing Systems and Applications Conference in February 2025 and showcased at the Bloomberg Green Festival.