A forest of 10,000 square kilometres is 100 million Sentinel-2 pixels in each 10 m band. With two satellites and a 290 km swath, the Copernicus Sentinel-2 mission description gives a five-day revisit at the Equator. Imagery is not the bottleneck. The hard part is turning that stack of reflectance values into decisions a range officer can act on: which compartments are under stress, where new plantation will survive and where water should be held back.
Regional forest monitoring combines Sentinel-2 and Landsat 8/9 optical imagery, Sentinel-1 radar for cloudy seasons, per-pixel vegetation index time series, terrain models and field plots. Red-edge indices catch stress that NDVI misses in closed canopy. Outputs such as health maps and plantation and check dam sites go into one spatial database, served through APIs and used offline in the field.
What the programme set out to do
The programme covered 10,000 square kilometres of forest and ran over three years. It had three analytical goals: assess forest health across the whole area, identify places suitable for plantation, and identify locations for check dams, which are small structures built across seasonal streams to slow runoff, hold water and let it soak into the ground. Each goal needed satellite and field data used together, because neither answers these questions on its own.
Two delivery requirements shaped the system as much as the analysis: one spatial database serving all results to other systems through APIs, and a mobile app that field staff could use offline, because much of the forest has weak or no coverage.
Choosing the satellite sensors
Sentinel-2 carries the MultiSpectral Instrument (MSI), with 13 bands. Four (blue, green, red, near-infrared) are at 10 m ground sample distance. Six are at 20 m: four narrow bands in the vegetation red-edge domain (B5, B6, B7 and B8A, centred near 705, 740, 783 and 865 nm) and two shortwave infrared bands (B11 and B12). These matter for forests because the transition from red absorption to near-infrared reflectance shifts as chlorophyll content changes, often before stress is visible in the canopy.
Landsat provides the long baseline. Landsat 8 and 9 carry the Operational Land Imager, whose nine spectral bands are all at 30 m except the 15 m panchromatic band, and together give eight-day coverage, as set out in the USGS Landsat 9 mission page. The grid is coarser than Sentinel-2, but the archive allows comparison with conditions years before a programme starts.
Optical sensors fail under cloud, which in wet-season tropical climates can mean weeks without a usable scene. Sentinel-1 fills that gap. Its C-band synthetic aperture radar (wavelength about 5.55 cm) provides its own illumination and images through cloud, day or night. In Interferometric Wide mode it covers a 250 km swath at 5 by 20 m with VV and VH polarisation, according to the Sentinel-1 mission description. VH backscatter responds to canopy structure and helps detect clearing.
| Sensor | Resolution used for forests | Revisit | Main role |
|---|---|---|---|
| Sentinel-2 MSI | 10 m (visible, NIR), 20 m (red-edge, SWIR) | 5 days (two satellites, Equator) | Health indices, red-edge stress detection |
| Landsat 8/9 OLI | 30 m (15 m panchromatic) | About 8 days combined | Historical baseline, long time series |
| Sentinel-1 C-band SAR | 5 by 20 m (IW mode) | 12 days per satellite, 6 days with two | Cloud-season monitoring, clearing detection |
Vegetation indices and where they stop working
The normalised difference vegetation index is the starting point. NDVI is (NIR minus red) divided by (NIR plus red), which for Landsat 8 and 9 means bands 5 and 4, with values from minus 1 to 1, as documented on the USGS Landsat NDVI product page. Healthy, dense vegetation sits high on the scale and bare soil sits near zero. NDVI is simple, well understood and comparable across sensors once the data are atmospherically corrected to surface reflectance.
Its main weakness in forests is saturation. Once the canopy closes, red reflectance is already near its minimum and NDVI stops rising, so a moderately stressed closed forest and a healthy one can return almost the same value. Mutanga and Skidmore's study of narrow-band indices measured this on dense grass canopy: standard NDVI explained little of the biomass variation (R² of 0.26), while NDVI-type indices built from narrow red-edge bands performed far better. The study was on pasture, not forest, but the mechanism is the same.
Two other indices help. The enhanced vegetation index adds a blue band and correction terms; the USGS Landsat EVI page describes it as correcting for some atmospheric conditions and canopy background noise and being more sensitive in areas of dense vegetation. The normalised difference red-edge index (NDRE) replaces the red band with a red-edge band, typically Sentinel-2 B5 with B8a, so it stays sensitive to chlorophyll changes after NDVI has flattened. A practical split is NDVI for broad classes and red-edge indices to separate condition within closed canopy.
Time series, phenology and change
A single image is easy to misread: a deciduous forest in the dry season can look degraded when it is simply leafless. The answer is a time series. Build a per-pixel profile of index values across several years, fit the seasonal curve (green-up, peak, senescence), and compare each new observation with what is expected for that pixel at that time of year. A drop against the pixel's own baseline means far more than a low absolute value.
Cloud masking and compositing come first. Cloudy and shadowed pixels are removed using each scene's quality band, such as the Sentinel-2 scene classification layer, and the rest are combined into monthly or seasonal median composites. Sentinel-1 backscatter can fill gaps where optical data are missing. The result is a regular cube of observations that supports both trend analysis and abrupt change detection, such as clearing or fire scars.
For forest loss specifically, global products give useful context. The global forest cover change maps by Hansen and colleagues, published in Science in 2013, mapped forest loss and gain at 30 m from Landsat for 2000 to 2012. Such datasets are a good cross-check, not a substitute, because their tree cover definition may not match a department's own forest classes.
Terrain inputs for plantation and check dams
Plantation suitability and check dam placement depend as much on terrain as on vegetation. Both start from a digital elevation model, from which slope, aspect, flow direction, flow accumulation, stream networks and stream order are derived. Aspect affects moisture and sun exposure, and so seedling survival. Slope limits where planting can happen and controls erosion risk. Flow accumulation shows where water concentrates and where a small structure can hold it, the same storage logic behind our waterbody register case study.
Siting is a multi-criteria problem: each criterion becomes a normalised suitability layer, and the layers are weighted and combined. A clear example is the study of GIS-based site selection for check dams by Rahmati and colleagues in Sustainability (2019). It used slope, stream power index, sediment transport index, topographic wetness index, drainage density, stream order and other terrain factors, set weights through the analytic hierarchy process, and combined them with weighted linear combination. Slope and stream power index received the highest weights.
Plantation suitability follows the same pattern with different criteria: land cover and canopy density from imagery, moisture proxies, slope and aspect, distance from existing forest, and exclusion layers for settlements, water and protected zones. The output is a ranked surface, not a yes or no map, so field teams confirm the highest-scoring candidates before any planting budget is committed.
Field data and ground truth
Satellite layers are only as good as the field data used to train and check them, which is why this programme used field datasets alongside the imagery. Typical plots record species, canopy condition, regeneration, disturbance and GNSS position. They train the classifiers that produce forest health classes, and held-back plots test the result with a confusion matrix and per-class precision, recall and F1.
The relationship runs both ways. Satellite results point field teams to blocks where red-edge values have dropped against the seasonal baseline. The visit confirms, corrects or rejects the result, and the record goes back into the database. Over a multi-year programme this loop builds a labelled dataset that improves later models.
Serving data through a database and APIs
In this programme all results were held in one spatial database. In systems of this kind, every layer shares one coordinate reference system and attribute schema. Rasters are stored as tiled, overviewed files and vector features (compartments, plots, candidate sites) as tables, with each record carrying its source, date and processing version.
Other systems read the data through APIs rather than file exports. OGC API - Features defines a REST interface for querying individual features over HTTP, which suits plot records and candidate sites, while map tiles serve raster layers. Our Smartout platform follows the same pattern today: one spatial model, change history on every layer, and APIs so planning and reporting tools use the same version of the data.
An offline mobile app for field staff
Field teams in forests often work beyond reliable mobile coverage, so the field app was built to work offline. In an offline-first design, staff download the layers and tasks for their area before a trip: base imagery, compartments, candidate sites and forms. The app records observations, photographs and GNSS positions locally, then synchronises with the central database when a connection returns.
Offline design affects data quality as much as convenience. Good field apps validate entries on the device, time-stamp and geolocate every record, and apply clear rules when two users edit the same feature. Without it, observations end up on paper and are typed in later. We apply the same thinking in remote and low-connectivity deployments.
What we delivered across three programmes
The 10,000 square kilometre programme described above is one of three. A separate conservation programme ran over four years and spanned 660 forest divisions. It assessed forest health, identified places suitable for plantation and identified places suitable for wildlife introduction, again using satellite and field datasets together. It followed the same pattern of a spatial database, API access and an offline mobile app, applied across a far larger number of administrative units.
Earlier work assessed the impact of afforestation using satellite data. That experience shaped how we approach plantation suitability: knowing what has and has not established in the past is one of the best guides to where planting should go next. All three programmes drew on the processing methods we continue to refine.
Forests as terrain for drones and robots
The model that supports forest planning is also the map that machines need. Drones and unmanned ground vehicles (UGVs) need a terrain twin: DTM, canopy height, tracks, streams and obstacles in one coordinate reference system. Each layer has a specific job in a mission.
For drones, the DTM drives terrain-following flight, and a canopy surface model sets the minimum safe altitude above the crowns. Compartment boundaries and protected zones become geofences. A drop in red-edge index against a compartment's seasonal baseline, or a fall in Sentinel-1 VH backscatter that suggests clearing, can become the next flight: the flagged area is turned into a survey polygon and the drone captures centimetre-level imagery that resolves individual crowns.
For UGVs, slope, surface class and mapped tracks give a traversability layer, so a route avoids gradients the vehicle cannot climb. GNSS is often degraded under closed canopy, so a ground robot can localise by matching its lidar against a prior point cloud and uses the twin to anchor its local map to the forest's coordinate system.
Fire response shows how the layers connect. NASA's Fire Information for Resource Management System (FIRMS) distributes active fire detections from MODIS and VIIRS globally within three hours of satellite observation. In a workflow of this kind, an alert inside a forest boundary is matched to the twin, which supplies terrain, access and recent vegetation condition. A drone team verifies it, and its imagery updates the shared model.
To be precise about scope: today Smartout maintains the operational twin, fuses satellite, drone and field data, keeps change history and serves it through APIs. The next layer, which we are building towards, is a robot perception engine that consumes camera and lidar feeds and updates the twin in near real time. Mission planning, route optimisation and navigation are longer-term directions, described on our robotics solutions page.
Limits of satellite forest health monitoring
Satellite data have clear limits. A 10 m Sentinel-2 pixel mixes several tree crowns, so individual tree health needs drone or field data. Cloud reduces optical coverage when change is fastest, NDVI saturates in closed canopy, and radar on steep slopes needs terrain correction. Classifiers may not transfer between forest types, and siting depends on expert-chosen weights. We therefore treat every satellite product as a prioritisation tool that shows field teams where to look first, not a final verdict.
Frequently asked questions
Which satellite is best for forest health monitoring?
Sentinel-2 is usually the main source because of its 10 m bands, 20 m red-edge bands and five-day revisit with two satellites. Landsat 8/9 adds a long historical record at 30 m. Sentinel-1 radar fills gaps when cloud blocks optical sensors. Most regional programmes use all three together rather than relying on a single sensor.
Why does NDVI underperform in dense forest?
Once a canopy closes, red reflectance is already very low and NDVI stops increasing, so different levels of forest condition return similar values. This is called saturation. Red-edge indices such as NDRE, and EVI with its correction terms, stay more sensitive in dense vegetation and help separate healthy from stressed closed canopy.
How is afforestation site selection done with satellite data?
Afforestation site selection combines land cover and canopy density from imagery with terrain layers from a DEM, such as slope and aspect, plus moisture proxies and exclusion zones. Each factor becomes a weighted suitability layer, and the combined result ranks candidate areas. Field teams then confirm the highest-ranked sites before planting is approved.
What is a check dam and how are sites chosen?
A check dam is a small structure across a seasonal stream that slows runoff and holds water so it can soak into the ground. Sites are chosen with multi-criteria analysis using DEM derivatives such as slope, flow accumulation, stream order and wetness indices, weighted by expert judgement and then verified in the field.
Put your forest data in one model
If you manage forest, conservation or watershed programmes and want satellite, terrain and field data in one model your teams and machines can use, we can show you how we approach it. Book a technical call and bring your area, your data and the decisions you need to make.