Sensors
- Planet SkySat: Delivers 50 cm pan-sharpened optical imagery with on-demand tasking. At that resolution, a freshly cut skid trail 3-4 m wide is detectable as a linear brightness anomaly against canopy. Revisit is task-dependent; the constellation of 21 satellites allows same-day collection over most tropical latitudes.
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral. The SWIR bands (eight bands, 3.7 m) can distinguish bare mineral soil on road surfaces from surrounding vegetation. Revisit is 1-4.5 days depending on latitude and off-nadir tolerance. Archive depth extends to 2014.
- ESA Sentinel-1 (C-band SAR): 6-day repeat (12-day single-pass) at 10 m resolution in IW mode. Coherence differencing between repeat passes detects surface disturbance including road clearing, but C-band penetrates only the top 1-2 m of canopy, limiting detection of trails under closed cover. Free and open.
- JAXA ALOS-2 PALSAR-2 (L-band SAR): L-band (1.27 GHz) penetrates closed tropical canopy to some depth, improving sensitivity to sub-canopy disturbance compared with C-band. Finest stripmap mode: 3 m resolution. Revisit is 14 days. Coherence loss from road construction is detectable but interpretation is ambiguous where ground moisture varies.
Why roads are the signal worth watching
Selective logging and illegal concession expansion follow a predictable physical sequence. A road is cut first. Timber is extracted along skid trails radiating from that road. A log-landing clearing accumulates felled trunks before trucking. Canopy loss at scale comes later, sometimes weeks later. By the time a 30 m Landsat pixel registers deforestation, the access infrastructure has already been in place long enough to have enabled substantial extraction.
That makes road detection a leading indicator rather than a lagging one. A government agency or concession auditor who can identify a new road segment the week it is cut has options. An agency working from annual tree-cover-loss maps does not. This page covers the sensors and methods that make early road detection possible, and is honest about where they fail.
What a freshly cut road looks like from orbit
In optical imagery, a new logging road appears as a narrow linear feature of high reflectance, typically bare mineral soil or compacted laterite, against the dark, absorptive canopy surrounding it. At 50 cm resolution (Planet SkySat), a road 3 m wide subtends roughly six pixels across its width. That is enough to distinguish a road from a stream or a shadow, provided the geometry is right and cloud cover is absent. At WorldView-3's 31 cm panchromatic resolution, road edges, tyre tracks and the disturbed soil of a log-landing clearing are individually resolvable.
Published work in the Brazilian Amazon has demonstrated automated road extraction from sub-metre imagery using convolutional neural network architectures trained on manually digitised road networks. Detection rates in open-canopy or recently cleared areas exceed 80% in favourable conditions. The honest qualification: accuracy drops sharply wherever the road passes under a canopy bridge or is obscured by shadow. Optical sensors see the surface. If the canopy closes over a track, the track disappears from the image.
SAR coherence: seeing through cloud, not through canopy
Synthetic aperture radar is the standard counter to cloud cover, which renders optical monitoring unreliable across most of the humid tropics for months at a time. Sentinel-1's 6-day repeat cycle allows coherence differencing: two passes over the same area are compared pixel by pixel, and areas where the surface has changed between passes show a loss of coherence. Road construction, clearing and soil compaction all produce coherence loss in C-band imagery.
The limitation is significant and should not be glossed over. C-band wavelengths (approximately 5.6 cm) interact primarily with the upper canopy. A skid trail cut beneath a closed canopy may produce no detectable coherence change at C-band because the canopy above it remains intact. L-band SAR from ALOS-2 PALSAR-2 penetrates further into the forest volume, and published studies in Southeast Asian concessions have shown improved detection of sub-canopy disturbance at L-band. But even L-band coherence change is ambiguous: soil moisture variation between passes produces coherence loss that mimics disturbance. Validation against optical imagery or field data is necessary before treating an L-band anomaly as confirmed road activity.
The practical workflow for most monitoring programmes combines both: SAR coherence to flag candidate areas through cloud, optical tasking to confirm and characterise once a clear window opens.
Log landings: the clearest target
If skid trails are difficult to detect, log-landing clearings are not. A typical Amazonian log landing is 30-100 m across, fully cleared of canopy, and covered in bare or compacted soil. At 10 m Sentinel-1 resolution it registers as a coherence-loss patch. At 30 m Landsat resolution it is detectable as a spectral anomaly. At sub-metre resolution the stacked timber, vehicle tracks and disturbed soil are individually visible.
Log landings therefore serve as anchor points for road network reconstruction. A detected landing can be used to seed a road-tracing algorithm working outward along the linear features that connect to it. This approach, used in published research across the Brazilian state of Pará, significantly improves road-network completeness compared with pixel-level classification alone.
Honest limits of the method
Three constraints bound what is achievable. First, cloud cover. Tropical cloud cover can block optical sensors for 60-90% of available acquisition opportunities in peak wet season. SAR partially compensates, but with the coherence ambiguities described above. No current spaceborne system reliably detects a narrow skid trail under closed canopy in continuous cloud.
Second, resolution versus coverage. WorldView-3 at 31 cm provides the clearest imagery but covers a limited area per tasking pass and carries a commercial cost per square kilometre. Sentinel-1 covers large areas freely but at 10 m resolution, where a 3 m road is sub-pixel. The choice of sensor is always a trade-off between spatial detail and the area that can be monitored affordably.
Third, the difference between detection and legal attribution. A road detected in a concession boundary does not automatically establish who built it, when, or whether it violates a permit. Satellite evidence is an input to an investigation, not a conclusion. Satellize's road-detection analytics, used alongside open Sentinel data and commercial tasking on client licence, are designed to produce GIS-ready evidence layers that support that investigation, not to replace it. The Tonga crop-estimation programme is a different domain, but the same principle applies: the analytic output is a starting point for a decision, not the decision itself.
From detection to a monitoring programme
A practical road-monitoring programme for a tropical forest jurisdiction typically runs in three tiers. Sentinel-1 coherence change runs continuously across the full area of interest, flagging candidate disturbance patches every 6-12 days at no imagery cost. Confirmed candidates above a size threshold trigger commercial optical tasking, either Planet SkySat for rapid response or WorldView-3 for highest-detail characterisation. Detected road segments and landings are delivered as vector GIS layers with timestamps, confidence scores and area statistics.
Archive depth matters. WorldView-3 archive extends to 2014; ALOS-2 to 2014; Sentinel-1 to 2014. Road networks that predate a concession permit can be reconstructed from archive imagery, establishing a baseline against which current activity is compared. That historical dimension is often what a regulator or auditor needs most.
Typical figures
| Finest optical resolution available | 31 cm panchromatic (Maxar WorldView-3); 50 cm pan-sharpened (Planet SkySat) |
| SAR resolution (Sentinel-1 IW mode) | 10 m range × 10 m azimuth (after multi-looking) |
| SAR resolution (ALOS-2 PALSAR-2 fine stripmap) | 3 m single-look |
| Sentinel-1 coherence revisit | 6 days (dual-satellite, same geometry); 12 days single satellite |
| Minimum detectable road width (optical, sub-metre) | Approximately 3 m in open canopy; undetectable under closed canopy |
| Minimum detectable disturbance (Sentinel-1 coherence) | Clearings roughly 0.1 ha and above; narrow trails under closed canopy not reliably detected |
| Spectral bands relevant to road detection | Panchromatic, red, NIR for bare-soil contrast; SWIR (WorldView-3) for soil mineralogy; C-band (5.6 cm) and L-band (23.6 cm) for SAR coherence |
| Archive depth | WorldView-3: 2014 to present; Sentinel-1: 2014 to present; ALOS-2: 2014 to present |
| Typical delivery format | Vector road-network GIS layer (GeoPackage or Shapefile), raster coherence-change map (GeoTIFF), timestamped detection report |
| Cloud cover impact on optical sensors | 60-90% of acquisitions unusable in peak wet season across humid tropics |
Analytics Satellize can run
| New road segment alerts | Sentinel-1 coherence differencing (6-day repeat); linear feature extraction on flagged patches | Polygon and centreline GIS layer with detection date, confidence score and area; delivered within 24 hours of SAR acquisition |
| Sub-metre road network map | CNN-based road extraction trained on manually labelled SkySat or WorldView-3 imagery; applied to tasked commercial scenes | Vector road centreline network with width estimates and log-landing polygons; GeoPackage format |
| Log-landing change detection | Multi-date optical classification of bare-soil spectral signature; cross-validated against SAR coherence loss | Timestamped polygon layer of new and expanded log landings; area and perimeter statistics per feature |
| Historical road baseline reconstruction | Archive optical and SAR time-series analysis to date road construction; compared against permit records | Dated road-network chronology report; GIS layer with estimated construction year per segment |
| Concession-boundary compliance layer | Spatial intersection of detected road network with published concession and protected-area boundaries | Flagged violation polygons with area of encroachment; PDF summary for regulatory submission |
| Ongoing monitoring subscription | Automated Sentinel-1 pipeline with threshold-triggered commercial optical tasking; human-reviewed before delivery | Monthly GIS update plus alert feed for new detections above agreed size threshold |
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.