Selective illegal logging detection in closed-canopy forest
Selective logging removes individual trees while leaving surrounding canopy intact, defeating coarse deforestation alerts. Sub-hectare gap analysis from Sentinel-1 SAR and high-resolution optical imagery can find the trails and gaps that give it away.
Sensors
- Sentinel-1 C-band SAR (ESA): 10 m ground range resolution in IW mode, 6-day repeat at the equator with both satellites active. C-band (5.405 GHz) penetrates light cloud but interacts primarily with upper canopy; logging roads and compacted skid trails produce distinctive low-backscatter linear features against the higher-return surrounding forest. Free and open archive from 2014.
- ALOS-2 PALSAR-2 L-band SAR (JAXA): L-band (1.2 GHz) penetrates deeper into the canopy volume than C-band, reaching the trunk-ground double-bounce layer. This makes it more sensitive to structural disturbance beneath a partially intact canopy. Fine-beam single polarisation offers 3 m resolution; standard modes run at 10 m. Revisit is 14 days, limiting temporal cadence.
- Planet SuperDove (PlanetScope): 3 m resolution, 8-band multispectral (including red-edge and NIR), near-daily revisit over most tropical forest regions. Canopy gaps as small as 100–200 m² are detectable in cloud-free acquisitions. Cloud cover in humid tropical forests is the principal operational constraint; usable clear-sky imagery may arrive only every 10–30 days in practice.
- Landsat 8/9 OLI (USGS/NASA): 30 m multispectral resolution, 8-day combined revisit with Landsat 8 and 9. Sufficient for detecting aggregated disturbance patches above roughly 0.1 ha but misses individual tree extraction. Useful as a temporal baseline and for contextualising gap clusters identified by finer sensors. Free archive extends to 1972 across the full Landsat series.
Why standard deforestation alerts miss the problem
The major operational forest-alert systems, including GLAD from the University of Maryland and Brazil's PRODES, are calibrated to detect substantial canopy loss. GLAD alerts flag pixels where canopy cover drops sharply; the detection threshold in published validation work sits around 25–50% canopy loss per 30 m pixel. Selective logging, by definition, removes individual stems while leaving the surrounding canopy largely intact. A single extracted tree may disturb 200–400 m² of canopy. That is a fraction of one Landsat pixel and produces no alert.
The economic incentive is precisely this invisibility. High-value tropical hardwoods, including ramin, merbau and certain dipterocarp species, command prices that make extraction of a handful of trees per hectare commercially worthwhile. The operation leaves a green canopy overhead and moves on. By the time a coarse-resolution alert fires, the timber is already on a truck.
What a skid trail gives away in radar
Logging roads and skid trails are the operational signature that remote sensing can actually find. Even when the canopy closes back over a trail within months, the compacted soil and disturbed root zone persist. In Sentinel-1 C-band imagery, freshly cut trails appear as linear low-backscatter anomalies, typically 3–8 dB below the surrounding forest return in VV polarisation, because compacted bare soil scatters far less than the rough, multi-layered forest surface.
ALOS-2 PALSAR-2 adds a complementary view. L-band energy reaches the trunk-ground interface, where double-bounce scattering from standing trunks produces a characteristic high-return signal. When trunks are removed, that double-bounce disappears. Published studies using PALSAR-2 data in Indonesian Borneo have detected selective-logging disturbance at spatial scales of roughly 0.1–0.5 ha, below the threshold of most optical alert systems. The key limitation: PALSAR-2's 14-day revisit means a trail must persist long enough to be imaged. Rapid canopy regrowth in humid tropics can obscure C-band signatures within 6–12 weeks, making near-real-time tasking important.
Optical gap analysis and its cloud problem
At 3 m resolution, Planet SuperDove can resolve individual canopy gaps created by tree extraction. The detection approach is straightforward: subtract a pre-disturbance canopy height or NDVI surface from a post-event image and flag pixels that have shifted from high-canopy to low-canopy or exposed-soil spectral signatures. Red-edge bands (705 nm and 740 nm in SuperDove's band set) are particularly sensitive to early canopy stress and partial shading changes that precede full gap formation.
The honest operational constraint is cloud. Humid tropical forests, which are precisely where selective logging pressure is highest, can be cloud-covered for 200 or more days per year. In practice, usable cloud-free SuperDove imagery over a given site may arrive only every two to four weeks, and sometimes less frequently during monsoon seasons. This does not make optical analysis useless; it makes SAR the primary detection layer and optical the confirmation layer, not the other way around.
Separating logging from natural treefall: the false-positive problem
Natural treefall gaps are a genuine ambiguity. In a mature tropical forest, individual trees fall at a background rate that creates canopy gaps of 50–500 m² continuously. These gaps produce radar and optical signatures that are, at the single-pixel level, nearly indistinguishable from a single extracted tree. This is not a minor caveat; it is the central statistical challenge of the method.
Published peer-reviewed work addresses this through several confidence filters. First, geometric linearity: logging trails are linear or curvilinear features extending tens to hundreds of metres, whereas natural treefall gaps are roughly circular or irregular. A connected-component analysis that flags only gap clusters with high length-to-width ratios substantially reduces false positives. Second, temporal clustering: selective logging operations create multiple gaps over a short period in a spatially coherent pattern, whereas natural treefall is spatially random. Third, proximity to known or suspected access routes. Studies applying these filters to Sentinel-1 time series in the Congo Basin and Amazon report precision values in the range of 60–80% at the gap-cluster level, with recall varying substantially by forest type and season. Those figures are honest benchmarks; they are not good enough for legal prosecution on their own, but they are sufficient to direct field verification teams to the right locations.
Building an operational detection pipeline
A practical pipeline layers the sensors by their strengths. Sentinel-1 runs as the continuous monitoring backbone: every 6-day pass is processed for backscatter change relative to a 90-day rolling baseline. Pixels showing sustained low-backscatter linear features trigger a candidate alert. PALSAR-2 acquisitions, when available, are used to cross-check double-bounce loss. SuperDove imagery is pulled for cloud-free windows over candidate sites to confirm optical gap signatures and assess gap geometry.
Alert confidence is assigned in tiers. A linear SAR anomaly alone is low confidence. A linear SAR anomaly coinciding with an optical gap cluster and proximity to a mapped or inferred access road is high confidence. Only high-confidence alerts are passed to enforcement partners for field verification. This tiered approach is consistent with the confidence-threshold frameworks described in published remote-sensing literature on disturbance detection, and it keeps the false-positive burden on field teams manageable.
Satellize runs this kind of multi-sensor disturbance pipeline on open Sentinel and Landsat archives, with commercial Planet tasking added under client licence where rapid confirmation is needed. The architecture is the same one underlying the crop-estimation work in Tonga: open data as the baseline, commercial data as the precision layer.
What the method cannot do
Selective logging detection from space has real ceilings. A single extracted tree in dense forest is below the detection limit of every operational satellite system currently available. The method detects the infrastructure of extraction, not the extraction event itself. It cannot reliably distinguish legal selective harvest from illegal extraction without ancillary data on concession boundaries and harvest permits. And in forests with high natural disturbance rates, such as those subject to frequent windthrow or pest outbreaks, the false-positive rate rises even with geometric and temporal filters applied.
Archive depth helps. Sentinel-1's record from 2014 and Landsat's from the early 1970s allow analysts to establish a long-term baseline of natural gap dynamics for a given forest patch, which tightens the anomaly threshold. But no satellite system currently provides the sub-metre resolution that would make individual-tree extraction directly visible. That gap may narrow as commercial very-high-resolution SAR constellations mature, but it has not closed yet.
Typical figures
| Spatial resolution (SAR) | 10 m (Sentinel-1 IW mode); 3–10 m (ALOS-2 PALSAR-2 fine-beam) |
| Spatial resolution (optical) | 3 m (Planet SuperDove); 30 m (Landsat 8/9 OLI) |
| Revisit cadence | 6 days (Sentinel-1, dual satellite); 14 days (ALOS-2); near-daily (Planet, cloud-permitting); 8 days (Landsat 8+9 combined) |
| Minimum detectable disturbance | Approximately 0.1–0.5 ha gap cluster (SAR, published studies); ~200 m² individual gap (Planet optical, cloud-free) |
| Key spectral / frequency bands | C-band 5.405 GHz VV/VH (Sentinel-1); L-band 1.2 GHz HH/HV (PALSAR-2); 8-band 400–900 nm incl. red-edge (SuperDove); OLI bands 1–7 (Landsat) |
| Cloud penetration | Full (SAR); none (optical — primary operational constraint in humid tropics) |
| Archive depth | Sentinel-1 from 2014; Landsat from 1972; ALOS-2 from 2014; Planet from ~2016 |
| Alert latency (SAR baseline) | 12–48 hours after Sentinel-1 acquisition, depending on processing pipeline |
| Typical delivery formats | GeoTIFF change layers, GeoJSON alert polygons, KMZ for field teams, PDF summary reports |
| Concession boundary integration | Required as ancillary input to distinguish illegal from permitted harvest; not derivable from satellite data alone |
Analytics Satellize can run
| SAR backscatter change map | Bi-temporal and multi-temporal log-ratio change detection on Sentinel-1 VV/VH time series, with 90-day rolling baseline normalisation | GeoTIFF raster of backscatter anomaly magnitude, updated every 6 days per Sentinel-1 pass |
| Linear feature extraction (trail detection) | Directional morphological filtering and Hough-transform-based linearity scoring applied to SAR anomaly layer to separate trail signatures from circular natural-gap signatures | GeoJSON polyline layer of candidate logging trails with linearity confidence score |
| Canopy gap cluster alerts | Connected-component labelling of sub-canopy disturbance pixels; gap clusters scored by area, geometry and proximity to access routes | Tiered alert feed (low / medium / high confidence) as GeoJSON polygons with metadata, delivered to client GIS or API endpoint |
| Optical gap confirmation layer | NDVI and red-edge difference analysis on Planet SuperDove cloud-free composites; spectral unmixing to estimate exposed-soil fraction within candidate gap pixels | GeoTIFF confirmation raster overlaid on SAR alert polygons; ingested automatically into alert-confidence scoring |
| Multi-year disturbance baseline | Landsat OLI time-series analysis using published vegetation-index change methods (e.g. LandTrendr algorithm class) to establish site-specific natural-gap background rate | Historical disturbance frequency map (GeoTIFF) used to calibrate anomaly thresholds for operational SAR monitoring |
| Enforcement-ready site report | Spatial join of high-confidence alerts against concession boundary and protected-area GIS layers supplied by client; photographic evidence compiled from cloud-free optical acquisitions | PDF report per alert site: coordinates, detection date, sensor evidence, confidence tier, nearest access route, recommended field-verification priority |
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.