Logging road network extraction for indirect carbon loss quantification
Logging roads cause two to five times more carbon loss than the harvested area alone, yet most project MRV ignores them. Automated extraction from sub-metre optical imagery and SAR coherence change closes that gap.
Sensors
- Maxar WorldView-3: 30 cm panchromatic, 1.24 m multispectral. At 30 cm GSD a road of roughly 1.5 m width is detectable as a linear feature; formal extraction algorithms typically require at least two pixels of road width, putting the practical minimum near 60 cm. Tasked revisit 1 to 4.5 days depending on latitude and collection angle.
- Planet SkySat: 50 cm panchromatic, 1 m colour. Minimum detectable road width for automated linear extraction is approximately 1 to 1.5 m under good contrast conditions. Near-daily tasking of a fixed area of interest is operationally achievable; archive from 2016 onwards.
- Sentinel-1 SAR (C-band): Interferometric Wide swath at 10 m GSD, 6-day repeat at the equator with both satellites. Coherence differencing between pre- and post-construction image pairs detects canopy removal along road corridors even under cloud cover; road widths below roughly 10 to 20 m are not individually resolved but cause coherence loss signatures wider than the physical disturbance due to edge effects.
- ALOS-2 PALSAR-2 (L-band): L-band (1.27 GHz) penetrates canopy and responds to soil moisture and surface roughness changes caused by road construction. Stripmap mode delivers 3 m resolution; ScanSAR 25 m. Coherence loss at L-band persists longer than at C-band in wet tropical forests, making PALSAR-2 better suited for detecting older, partially revegetated roads. Revisit 14 days.
Why the road budget dwarfs the harvest footprint
A felled tree is visible. The road built to reach it is permanent, and its consequences compound. Published work from the Carnegie Airborne Observatory and subsequent studies using Landsat and LiDAR time-series has consistently found that logging road construction triggers carbon losses two to five times larger than the directly harvested area, once edge-effect drying, windthrow, and subsequent smallholder encroachment are counted over a five to ten year window.
The mechanism is straightforward. A road edge exposes previously interior forest to altered humidity, wind and light. Mortality rates within 100 to 300 m of a new road edge are measurably elevated. The road also functions as an access corridor: secondary hunters, farmers and illegal loggers follow within months. None of this appears in conventional project accounting, which typically draws a polygon around the harvested coupe and stops there. The omission is not trivial; it is the dominant source of underestimation in selective-logging MRV.
What a freshly bulldozed road gives away spectrally and geometrically
Logging roads are geometrically distinctive: they are narrow, linear, often branching in herringbone patterns off a main haul road, and they expose bare mineral soil or compacted laterite with a spectral signature sharply different from closed canopy. In very-high-resolution optical imagery at 30 to 50 cm, the road surface itself is resolvable. Automated extraction methods, including convolutional neural network approaches trained on labelled road pixels and classical morphological thinning algorithms, can detect roads as narrow as two pixels wide. At WorldView-3's 30 cm panchromatic resolution that is roughly 60 cm; at SkySat's 50 cm it is approximately 1 to 1.5 m. Skid trails, which may be only 2 to 4 m wide, are detectable at WorldView-3 resolution but will be missed by Sentinel-2 or Landsat entirely.
Canopy shadow complicates extraction in steep terrain. Roads running east-west through valleys can be partially obscured in morning acquisitions. Multi-date compositing, or fusion with SAR-derived disturbance masks, substantially reduces false negatives. Honest practice requires reporting both the detected network and a confidence-weighted estimate of what may have been missed due to shadow or cloud.
SAR coherence loss as a cloud-independent detector
Tropical forests are cloudy. Optical sensors miss months of activity in the wet season, which is precisely when logging operations often accelerate ahead of seasonal access windows. SAR coherence offers a complementary view. When forest canopy is removed to build a road, the interferometric coherence between a pre-event and post-event Sentinel-1 image pair drops sharply over the disturbed corridor. The effect is detectable even when the road itself is narrower than the 10 m pixel, because the surrounding canopy disturbance widens the coherent-loss signature.
ALOS-2 PALSAR-2 at L-band adds a further capability. L-band wavelengths interact with larger woody structure and penetrate partial canopy regrowth. Roads that have been cut but are beginning to revegetate, and which may appear closed to optical sensors within one to two growing seasons, remain detectable in PALSAR-2 coherence stacks for considerably longer. This matters for retrospective audits of project baselines, where the question is not what is happening now but what was built in the three to five years before a carbon project registered.
A known limitation: SAR coherence also decorrelates over wet soil, flooded areas and agricultural fields. Separating road-related coherence loss from seasonal flooding or cultivation requires ancillary land-cover context. Applying a forest mask derived from optical data before the coherence analysis reduces false positives substantially.
From extracted network to quantified indirect emissions
Detecting the road is step one. Translating it into a carbon loss estimate requires a chain of published methods. The Global Road Watch programme, operated by World Resources Institute, provides a framework for road network attribution and change detection at global scale using Landsat-derived products, though its minimum detectable road width of roughly 30 m makes it unsuitable for skid-trail-level accounting. For project-scale MRV, the extracted road centreline network from VHR imagery is buffered by empirically derived edge-effect distances, typically 100 to 300 m depending on forest type and published mortality-rate studies, and the resulting disturbed area polygon is intersected with above-ground biomass maps to estimate carbon at risk.
The Carnegie Airborne Observatory's published work on Borneo and the Amazon established the methodological precedent: road length, road width, buffer distance and biomass density are the four inputs. Uncertainty is real. Buffer distances derived from one forest type may not transfer cleanly to another. Biomass maps carry their own estimation errors. A defensible MRV product reports the road network extraction confidence, the buffer assumption and its literature source, and the biomass map provenance separately, so an auditor can interrogate each component.
Practical limits and what to tell an auditor
No sensor combination is perfect. WorldView-3 and SkySat deliver the geometric resolution needed for skid-trail detection, but tasking cost constrains how frequently large project areas can be revisited. Sentinel-1 covers continuously and freely but cannot resolve individual roads narrower than roughly 20 m. PALSAR-2 has a 14-day repeat and limited commercial availability for tasked acquisitions. A credible MRV design layers the sensors: Sentinel-1 coherence for continuous cloud-independent monitoring of major road construction, VHR optical for periodic high-confidence road network mapping, and PALSAR-2 for retrospective audit of partially revegetated older roads.
Satellize structures this kind of layered sensor programme for carbon project operators and auditors, drawing on open Sentinel and PALSAR-2 archives alongside commercially tasked SkySat or WorldView-3 collections. The Tonga crop-estimation programme demonstrated the organisation's approach to multi-date compositing and change attribution; the same temporal analysis logic applies directly to road network change detection.
An auditor reviewing road-based indirect emissions should expect to receive, at minimum: a dated road network shapefile with confidence scores per segment, a documented buffer methodology citing the peer-reviewed source for the edge-effect distance used, an area-weighted biomass estimate with uncertainty bounds, and a clear statement of what sensor coverage gaps exist in the monitoring period. Anything less is not MRV; it is a map.
Typical figures
| Best spatial resolution (optical) | 30 cm panchromatic (Maxar WorldView-3); 50 cm panchromatic (Planet SkySat) |
| Minimum detectable road width (VHR optical) | ~60 cm at WorldView-3 (2-pixel rule); ~1.0–1.5 m at SkySat under good contrast |
| SAR resolution | 10 m IW (Sentinel-1); 3 m Stripmap / 25 m ScanSAR (ALOS-2 PALSAR-2) |
| Minimum detectable disturbance width (SAR coherence) | ~10–20 m corridor (coherence loss signature typically broader than physical road) |
| Revisit cadence | 6 days (Sentinel-1, dual satellite); 14 days (PALSAR-2); 1–4.5 days tasked (WorldView-3, SkySat) |
| Cloud penetration | Full (Sentinel-1, PALSAR-2 SAR); none (optical sensors) |
| Archive depth | Sentinel-1 from 2014; PALSAR-2 from 2014; SkySat from ~2016; WorldView-3 from 2014 |
| Edge-effect buffer (published range) | 100–300 m from road centreline, depending on forest type and published mortality studies |
| Delivery formats | Road network shapefile / GeoJSON with confidence scores; disturbed-area polygon; carbon-at-risk raster; PDF audit summary |
Analytics Satellize can run
| Road network extraction | CNN-based linear feature detection on VHR panchromatic imagery, with morphological post-processing to enforce network topology | Dated road centreline shapefile, segment-level confidence score, total network length in km |
| SAR coherence disturbance map | Sentinel-1 interferometric coherence differencing (pre/post image pairs); forest mask applied to suppress non-road decorrelation sources | Binary and continuous coherence-loss raster, road-corridor candidate polygons, change date estimate |
| Retrospective road audit (revegetated roads) | ALOS-2 PALSAR-2 L-band coherence time-series analysis over project baseline period | Historical road network reconstruction shapefile for audit period, with estimated construction date range per segment |
| Indirect carbon loss estimate | Road network buffered by published edge-effect distances (100–300 m); intersected with above-ground biomass map; carbon calculated using IPCC Tier 1 or site-specific biomass density | Carbon-at-risk polygon layer, tCO2e estimate with uncertainty range, methodology citation sheet |
| Multi-sensor monitoring stack | Sentinel-1 coherence for continuous coverage; VHR optical for periodic high-confidence snapshots; PALSAR-2 for retrospective gap-fill; results merged into unified change log | Quarterly GIS update package, coverage-gap report, sensor substitution log for auditor transparency |
| Project boundary road-encroachment alert | Road network extraction run against project boundary buffer zone; new road segments flagged against prior baseline network | Alert report with coordinates, imagery date, estimated new road length and associated indirect disturbance area |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.