Using Sentinel-2 for Forest Monitoring

Updated 2026-11-02 · 4 min read · By the akar.earth specialist team

Sentinel-2 forest monitoring in practice: bands, 10 m resolution, cloud handling in Indonesia, change detection and what verifiers expect to see.

Sentinel-2 is the default free satellite for forest monitoring because it combines 10 m resolution, red-edge and shortwave infrared bands, and a revisit of about five days. It is excellent for mapping forest cover, canopy loss and regrowth over project areas and supply chains, but it does not measure biomass directly and it is limited by cloud in much of Indonesia.

What Sentinel-2 gives you

Sentinel-2 is part of the EU Copernicus programme. Each satellite carries the MultiSpectral Instrument (MSI) with 13 bands:

The archive starts in 2015 (Sentinel-2A), with denser coverage from 2017 once Sentinel-2B joined. That matters for the EUDR cut-off of 31 December 2020 and for ARR eligibility checks: Sentinel-2 can show what land looked like at the cut-off, at 10 m. For earlier history, use Landsat at 30 m.

A practical workflow

1. Define the area. Use accurate project or plot polygons. For EUDR, plots over 4 ha need polygons; see polygon requirements.
2. Use Level-2A surface reflectance, which is atmospherically corrected, rather than top-of-atmosphere Level-1C.
3. Mask clouds and shadows using the scene classification layer or a dedicated cloud probability product. Default masks miss thin cloud and haze, which is common during the dry-season burning period.
4. Composite. Build monthly, quarterly or annual median composites to fill gaps. In very cloudy areas, an annual best-pixel composite may be the only reliable option.
5. Calculate indices. NDVI saturates in dense canopy; NBR (using SWIR) is better for burn and clearing detection; red-edge indices help separate plantations from natural forest.
6. Classify or detect change. Train a classifier with labelled points, or compare composites across dates. Validate with independent reference points.
7. Report area and uncertainty using a sample-based accuracy assessment, not just pixel counts.

Where Sentinel-2 works well

• Forest/non-forest and canopy cover mapping for ARR eligibility and baseline land cover. See land eligibility.
• Clearing detection for REDD+ activity data and supply-chain monitoring. Integrated alert systems already use Sentinel-2; see deforestation alerts.
• Tracking canopy closure in reforestation, including the vegetative stocking index style of analysis used in VM0047.
• Separating land uses such as oil palm, rubber and natural forest, which matters for EUDR satellite verification.

Where it struggles

• Cloud cover. Sumatra, Kalimantan and Papua can go weeks without a clear view. Combine with Sentinel-1 radar, which sees through cloud.
• Degradation. Selective logging that removes a few trees per hectare often leaves a canopy that still looks closed at 10 m.
• Small parcels and agroforestry. A 0.5 ha coffee or cocoa garden under shade trees is about 50 pixels. Mixed pixels make “forest vs. agroforestry” calls uncertain.
• Biomass. Reflectance saturates as canopies close. For stocks, see remote sensing biomass.
• Super-resolution. Some vendors super-resolve Sentinel-2 to finer grids. This can help visual interpretation, but the information content is still bounded by the original 10 m and 20 m data. Treat super-resolved products as interpretation aids and ask how they were validated.

Pairing Sentinel-2 with other sensors

Sentinel-2 rarely works alone in Indonesia. Three pairings cover most needs:

• Sentinel-1 radar for cloud-free change detection. Radar responds to structure and moisture rather than colour, so it fills gaps in the wet season and during haze.
• Landsat for history. Landsat’s 30 m archive reaches back decades, which matters when land eligibility or a REDD+ reference period starts before 2015. Harmonised Landsat–Sentinel products make the two easier to combine.
• Very-high-resolution imagery (commercial satellites or drones) for checking individual alerts, small parcels and training labels.

Plan the stack before analysis starts, and record which sensor answered which question. Verifiers find mixed-sensor workflows convincing when the role of each source is explicit.

Common mistakes

• Comparing images from wet and dry seasons and calling the difference “deforestation”.
• Using NDVI thresholds copied from temperate studies.
• Ignoring geometric offsets between scenes when analysing small plots.
• Reporting pixel-count area of change without an accuracy assessment, which biases area estimates.
• Not keeping scene IDs, which makes the result impossible to reproduce.

What good evidence looks like

A verifier or EUDR auditor should be able to rerun your analysis. Keep the list of scene or tile IDs and dates, the cloud mask method, composite rules, the training and validation point sets, the classifier settings and code version, and an accuracy assessment with confusion matrix. Archive the before and after images for every flagged change.

For a comparison with commercial imagery, see Sentinel vs Planet. Sentinel-2 is one layer in a full monitoring system; the dMRV hub explains how it fits with field data, and our dMRV implementation service sets it up for projects and supply chains.