Roads are expanding faster than at any point in human history, and many of them don't appear on maps. The ghost roads study in 2024 used ~7,000 hours of volunteer effort to hand-map roads across the tropical Asia-Pacific, finding 3--6.6 times more roads than exist in current global road datasets.
Road building almost always preceded local forest loss, making unmapped roads one of the best early warning signals we have for deforestation. Mapping them quickly and globally is therefore a conservation priority and not just a "cartographic tidyup"!
The current SOTA detects roads directly from Sentinel imagery with CNN segmentation models, such as monitoring road development in the Congo Basin, mapping roads in the Brazilian Amazon, and the ghost roads team's own remote road mapping with AI. The 10m resolution of Sentinel does a surprisingly good job, esp. for new roads where there's an initial canopy opening. It's also free and global.
The Congo Basin work is the most directly comparable baseline. Slagter et al. fuse Sentinel-1 and Sentinel-2 at 10m to detect road openings monthly, reporting an F1 of 0.909 with a 4.2% false detection rate and a 14.9% missed detection rate. The S1 radar channel is essential under the Congo's near-permanent cloud!
The resulting forest roads dataset
is CC-BY and revised regularly, with the current release covering 68000+ jm
road development from 2019--2025 as vector line features, each carrying
a NetworkID for the connected network it belongs to. It can
be browsed live.
WRI have since folded it into a longer Congo Basin forest roads layer on Global Forest Watch that stitches together hand-digitised Landsat roads back to 1976 (!), crowdsourced logging-road mapping up to 2018, and the ML detections from 2019 onwards. Nearly a quarter of the roads it maps sit inside forest that was still classified as intact in 2016.
1 Project overview
This project would investigate whether the TESSERA foundation model can be better/faster than raw Sentinel imagery. Tessera compresses a full year of Sentinel-1 and Sentinel-2 observations into a 128-dimensional embedding for every 10m pixel on the planet (easily accessible via GeoTessera Python library released for geospatial embeddings), so a road classifier also gets the radar signal through cloud and a year of temporal dynamics rather than a single optical scene.
That temporal signal might also allow us to separate paved from unpaved surfaces, and to spot informal logging tracks that only appear seasonally. Road detection is already one of the downstream benchmarks in the Tessera v2 evaluation suite (via the CityRep urban representation benchmark), so there is a baseline to start from.
An interesting twist with roads is that they are not independent pixels. Almost every road connects to another road! A per-pixel classifier throws that structural prior away, but a generative approach could exploit it by starting from a known road network (e.g. OpenStreetMap) and tracing outwards into the unknown. Exactly which ML technique fits best is an open question the project would settle, but candidates include:
- Connectivity-aware segmentation by adding topology-preserving losses to a conventional segmentation head over embedding patches so that predicted roads form connected networks rather than a speckle.
- Iterative tracing in the style of RoadTracer, where an agent walks outwards from known road seeds and decides at each step whether the road continues through embedding space.
- Generative graph completion, treating the mapped network as a partial graph and training a model to propose the missing edges, sampling multiple plausible completions to get uncertainty over where ghost roads might lurk.
- Temporal change detection, differencing yearly Tessera embeddings to catch new canopy openings as roads are actively constructed.
2 Validation strategies
Validation can be done since the ghost roads team's hand-digitised dataset gives us high-quality labels for where roads actually are across the Asia-Pacific tropics, and beyond what OpenStreetMap coverage would suggest for those years. We have Tessera embeddings from 2017-2025.
The Congo Basin data gives us a second, independent tropical test region in a biome the Asia-Pacific labels don't cover. Its labels are already connected graphs rather than a raster, which is the structural prior our generative approaches above aim to learn from. The region also has an unusually rich temporal story to test against: Kleinschroth et al. found that road length inside logging concessions doubled between 2003-2018 while nearly half of those roads were subsequently abandoned. 30% of the openings Slagter et al. detected turned out to be previously abandoned roads being reopened.
Distinguishing new construction from a road that is reopening, or an open road from a revegetation scar, are both temporal classification problems that the annualised Tessera should be able to do better than plain optical snapshots.
Note that the Congo product detects at monthly granularity whereas a Tessera embedding compresses a whole year. Part of the project is working out where on that curve the useful tradeoff between latency and accuracy lies. Fesenmyer et al. calibrate these road development maps into annual timber harvest volumes and carbon emissions for six Congo Basin countries, so any improvement in road detection goes straight through into emissions accounting and policy impact.
3 See also
Andres Zuñiga-Gonzalez has done prior work on road mapping with OSM labels over Tessera embeddings; see his weeknotes on detecting new roads and categorising OSM road features, and his osm-rasterizer tool for generating training rasters from OSM features.
The Global habitat maps from Tessera embeddings work covers the broader classifier-on-embeddings methods we're developing as part of the wider TESSERA programme.