Near-real-time deforestation alerts
Dense time-series of optical and SAR imagery can flag new forest clearance within days of occurrence. The hard problem is not the algorithm; it is cloud cover, revisit gaps, and minimum mappable area.
Sensors
- Sentinel-1 SAR (C-band, ESA): 10 m ground range resolution in IW mode, 6-day repeat at the equator (12-day for a single satellite). Cloud-penetrating by design, making it the primary source for humid tropical alerts where optical data is routinely blocked for weeks at a time. Backscatter drops sharply when dense canopy is removed.
- Landsat 8/9 OLI (USGS/NASA): 30 m multispectral resolution, 16-day revisit per satellite (8-day combined). The long archive from 1972 onward underpins CCDC and CODED baseline models. Free and open; latency from acquisition to public availability is typically under 24 hours.
- Sentinel-2 MSI (ESA): 10 m in visible and near-infrared bands, 5-day revisit at the equator with both satellites. NDVI and NBR time-series from Sentinel-2 feed the GLAD-S2 alert system. Cloud masking with the Sen2Cor or s2cloudless processors is essential before any change detection step.
- Planet SuperDove: 3 m resolution, near-daily revisit globally. Commercially licensed. Dramatically reduces the minimum mappable area and the latency penalty from cloud gaps, but the short archive depth (from 2016 at best) limits baseline construction compared with Landsat.
What the algorithms are actually doing
Two published methods dominate operational deforestation detection. CCDC (Continuous Change Detection and Classification), developed at Boston University, fits harmonic regression models to every pixel's time-series across all available Landsat bands. When new observations fall outside the predicted seasonal envelope by a statistically significant margin, the pixel is flagged as changed. The method is sensitive to gradual degradation as well as abrupt clearance, which is useful, but it requires a dense historical archive to build reliable baselines.
CODED (Continuous Degradation Detection) extends CCDC specifically for tropical forests, adding a spectral unmixing step that separates the signal of bare soil and non-photosynthetic vegetation from photosynthetic canopy. This makes it more sensitive to partial clearance and selective logging than a raw NDVI threshold would be. Both algorithms are open-source and run on Google Earth Engine, which is how research groups and national agencies deploy them at scale.
Operational alert systems and their actual latency
The GLAD alert system, produced by the University of Maryland's Global Land Analysis and Discovery lab, uses Landsat 8/9 and Sentinel-2 to issue alerts at 30 m resolution. Alerts are published within two to three days of a cloud-free observation confirming loss. The word 'confirming' carries weight: in persistently cloudy regions such as the Congo Basin or the western Amazon, a cloud-free observation may not arrive for four to eight weeks, regardless of how fast the algorithm runs.
The RADD (Radar for Detecting Deforestation) alert system, also from the GLAD lab, addresses exactly that problem by substituting Sentinel-1 SAR backscatter for optical reflectance. Because C-band radar penetrates cloud, RADD can issue alerts within six to twelve days of clearance in regions where GLAD-optical would be silent for a month. The trade-off is that SAR backscatter is sensitive to moisture and surface roughness, so heavy rainfall on bare soil can temporarily mimic the backscatter signature of standing forest, producing false positives that require optical confirmation when cloud finally clears.
Fusion approaches that combine RADD timing with GLAD optical confirmation are now standard practice in serious monitoring programmes. Neither system alone is sufficient.
Minimum mappable area: the constraint nobody advertises
At 30 m resolution, the practical minimum mappable clearance event is roughly 0.09 ha (one pixel), but reliable detection in practice requires several contiguous changed pixels, putting the real floor closer to 0.5 to 1 ha for Landsat-based systems. Sentinel-2 at 10 m lowers that to around 0.05 to 0.1 ha. Planet SuperDove at 3 m can detect clearances of a few hundred square metres, which matters for detecting the selective felling patterns that precede large-scale clearance.
These thresholds are not fixed. They depend on the density of the time-series, the quality of cloud masking, and the spectral contrast between the cleared surface and surrounding forest. A freshly graded red-soil road through dark-canopy forest is far easier to detect than a recently burned patch where charred ground has similar near-infrared reflectance to stressed canopy.
Where cloud cover breaks the promise of 'near-real-time'
The phrase 'near-real-time' is technically accurate for the interval between a cloud-free observation and an alert being issued. It says nothing about the interval between the clearance event and the first cloud-free observation. In the humid tropics, Sentinel-2 cloud-free fraction at the equator averages below 20 percent during the wet season. That means the majority of clearance events are first detected weeks after they occur, even with a five-day revisit.
SAR partially compensates, but not completely. Sentinel-1's six-day repeat means a clearance event will be observed within six days by radar, assuming the satellite's acquisition plan covers that tile, which is not universal. In practice, the combination of Sentinel-1 RADD alerts for timing and Sentinel-2 or Landsat for spatial confirmation and attribution gives the best achievable latency for operational programmes without commercial tasking.
What a monitoring programme actually needs to decide
Alert latency, spatial resolution, and false-positive rate form a triangle: improving one usually degrades another. A programme focused on rapid law-enforcement response needs low latency and is willing to accept more false positives that field teams then verify. A programme feeding REDD+ carbon accounting needs high spatial accuracy and low false-positive rates, and can tolerate longer confirmation windows.
The choice of sensor stack follows from that decision, not the other way around. Satellize structures deforestation monitoring engagements around this question first, drawing on the same open constellations (Sentinel-1, Sentinel-2, Landsat) that underpin its crop-estimation work in Tonga, supplemented by commercial tasking where a client's licence covers it. The GLAD and RADD alert feeds are the starting point, not the finished product; attribution, boundary delineation, and integration with national forest inventories require additional processing steps that vary by jurisdiction.
Archive depth and baseline integrity
A change-detection algorithm is only as good as its baseline. CCDC typically requires three to five years of dense Landsat observations to fit a reliable harmonic model. In regions where Landsat archive coverage before 2013 is sparse due to acquisition gaps or persistent cloud, the baseline is weaker and false-positive rates rise. Sentinel-2 archive depth runs from mid-2015 at best, which is sufficient for most current programmes but limits retrospective analysis.
Planet's commercial archive from 2016 onward adds high-resolution context but is not long enough to replace Landsat as a baseline source. For any programme that needs to demonstrate additionality under a carbon standard, the Landsat archive remains the only freely available source with the temporal depth to support a credible pre-project reference period.
Typical figures
| Spatial resolution (optical) | 10 m (Sentinel-2), 30 m (Landsat 8/9), 3 m (Planet SuperDove) |
| Spatial resolution (SAR) | 10 m ground range, Sentinel-1 IW mode |
| Revisit period | 5 days (Sentinel-2, two satellites); 8 days (Landsat combined); 6 days (Sentinel-1); near-daily (Planet) |
| Alert latency after cloud-free observation | 2 to 3 days (GLAD optical); 6 to 12 days (RADD SAR) |
| Effective latency in humid tropics (wet season) | 2 to 8 weeks for optical confirmation; 6 to 12 days for SAR-only |
| Minimum mappable clearance area | ~0.5 to 1 ha (Landsat); ~0.05 to 0.1 ha (Sentinel-2); ~0.01 ha (Planet) |
| Spectral bands used | NIR, SWIR, Red (optical NDVI/NBR); C-band 5.6 GHz VV/VH (SAR) |
| Archive depth | 1972 onward (Landsat); 2015 onward (Sentinel-2); 2014 onward (Sentinel-1); 2016 onward (Planet) |
| Geographic coverage | Global; Sentinel-1 acquisition plan varies by region |
| Delivery formats | GeoTIFF alert rasters, GeoJSON polygon feeds, CSV alert logs, WMS/WMTS tile services |
Analytics Satellize can run
| Daily RADD SAR alert feed | Sentinel-1 backscatter change detection (RADD algorithm, GLAD/UMD) | GeoJSON polygon feed of new alerts, updated on each Sentinel-1 overpass, with confidence score and date of first detection |
| Confirmed optical deforestation alerts | GLAD-Landsat and GLAD-S2 alert fusion with cloud-mask filtering | Weekly GeoTIFF layer of confirmed loss polygons with area estimate and first-detection date |
| CCDC baseline and change map | Continuous Change Detection and Classification on Landsat time-series (Zhu & Woodcock method) | Annual change-detection raster with per-pixel change date, magnitude, and land-cover attribution |
| CODED degradation and clearance map | Spectral unmixing and CODED algorithm on Landsat/Sentinel-2 stack | GIS layer distinguishing abrupt clearance from gradual degradation, with confidence intervals |
| SAR-optical fusion alert with false-positive filter | RADD timing cross-validated against next available Sentinel-2 or Landsat cloud-free observation | Filtered alert shapefile with latency metadata, suitable for law-enforcement dispatch or REDD+ reporting |
| Jurisdictional deforestation rate report | Spatial aggregation of confirmed alerts within administrative or concession boundaries | Monthly PDF and CSV report with hectares cleared by administrative unit, trend chart, and comparison to reference period |
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.