Post-closure land cover trajectory and compliance verification
Closed mine sites carry financial bonds and legal rehabilitation milestones that can span decades. Time-series classification of Landsat and Sentinel-2 imagery turns the satellite archive into an independent, auditable record of land cover change from bare ground through to forest canopy.
Sensors
- Landsat TM / ETM+ / OLI (USGS / NASA): 30 m multispectral resolution, 16-day revisit per satellite (8-day with two satellites in tandem). Archive from 1972 onward provides the only consistent global baseline covering the full closure arc of mines permitted in the 1980s and 1990s. Bands include SWIR-1 and SWIR-2, essential for separating bare soil from sparse vegetation and for detecting moisture in rehabilitated substrates.
- Sentinel-2 MSI (ESA Copernicus): 10 m (visible/NIR) and 20 m (red-edge, SWIR) resolution, 5-day revisit at mid-latitudes with both satellites. Available from 2015. The four red-edge bands (B5, B6, B7, B8A) improve discrimination between early pioneer cover and established shrub, and support more precise NDVI and EVI time series than Landsat alone.
- Planet SuperDove: 3–5 m resolution, near-daily revisit. Useful for resolving spatial heterogeneity within a single waste-rock bench or small rehabilitation plot that a 10 m pixel would average into ambiguity. Adds cost; best applied to contested areas or pre-bond-release verification rather than routine annual tracking.
- ALOS-2 PALSAR-2 (JAXA): L-band SAR at 3–10 m resolution depending on mode. L-band penetrates canopy to return a backscatter signal correlated with above-ground biomass. Provides a structural cross-check on optical greenness indices: a site can look green in NDVI while carrying very little woody biomass, a distinction that matters for bond-release thresholds tied to forest-class targets.
Why a single vegetation survey is not enough
Most mining jurisdictions require periodic vegetation surveys as a condition of rehabilitation bond release. Those surveys are point-in-time. They cannot tell a regulator whether a site is on a credible recovery trajectory, whether a green flush after an unusually wet season masks underlying failure, or whether a site that looked healthy three years ago has since regressed. The question is not just 'what is growing there now' but 'where is this site going over the next decade'.
Satellite time series answer that question directly. When you stack annual or seasonal land cover classifications across a fifteen- or thirty-year archive, you get a trajectory: the rate of bare-ground reduction, the timing of shrub establishment, whether the pioneer-to-shrub transition stalled. That trajectory can be compared against the regulatory milestone schedule written into the closure plan, producing a compliance gap map rather than a compliance snapshot.
What the Landsat archive actually gives you
Landsat's value here is not its resolution. Thirty metres is coarse enough that a single pixel on a rehabilitated waste-rock dump may mix bare substrate, sparse grass and a shrub. The value is temporal depth. Landsat 5 TM operated from 1984 to 2013; Landsat 7 ETM+ from 1999; Landsat 8 and 9 OLI from 2013 and 2021 respectively. For a mine that closed in 1995, you can reconstruct the land cover state at closure, at five years, at ten years, and every year since. No ground survey programme has that kind of retrospective reach.
The USGS Collection 2 surface-reflectance archive applies consistent atmospheric correction across the entire record, which matters enormously for inter-annual comparison. Without it, you are comparing apples and oranges every time the satellite's sensor ages or the atmosphere behaves differently. Even so, analysts must account for sensor-to-sensor radiometric differences when crossing the TM-to-OLI transition, and cloud contamination in persistently overcast climates (tropical highlands, maritime temperate zones) can produce gaps of one to three years in useful observations. That is a real limit, not a footnote.
Classifying land cover states across decades
The standard approach stacks all cloud-free observations within an annual or seasonal window, composites them to reduce noise, then classifies each pixel into a discrete cover class: bare ground, sparse herbaceous, dense herbaceous, shrub, woodland or forest. Random forest and support vector machine classifiers trained on reference data from the site's own early post-closure imagery tend to outperform generic global land cover products, because mine rehabilitation substrates (crushed waste rock, applied topsoil, seeded grasses) have spectral signatures that differ from natural soils in the surrounding landscape.
Sentinel-2 red-edge bands add meaningful discrimination between early pioneer cover and established shrub that Landsat's six multispectral bands cannot reliably resolve. A practical workflow fuses both: Landsat for the long baseline back to closure, Sentinel-2 from 2015 onward for finer spatial and spectral detail. The fusion requires careful cross-calibration; published methods based on the harmonic analysis of time series (HANTS) and the Continuous Change Detection and Classification (CCDC) algorithm are well-documented in the peer-reviewed literature and are the defensible choices for regulatory submission.
ALOS-2 PALSAR-2 L-band backscatter adds a structural dimension that optical indices cannot. A rehabilitated site planted with fast-growing Acacia may show high NDVI within three years while carrying above-ground biomass far below the target forest class. L-band backscatter, particularly the HV cross-polarisation channel, correlates with woody biomass at the stand level and provides an independent check on whether optical greenness reflects genuine structural recovery.
Turning classification into a compliance argument
A land cover map is not a compliance document. The analytical step that makes it one is aligning the classified trajectory against the milestones written into the approved closure plan: 'at least 70% shrub cover by year five', 'no bare ground exceeding 5% by year ten', 'forest canopy equivalent to reference ecosystem by year twenty'. Each milestone becomes a spatial query against the time-series stack.
The output is a milestone compliance matrix: each spatial unit (bench, catchment, dump face) against each milestone year, flagged as compliant, non-compliant, or insufficient data. Regulators and bond holders can read this directly. It is also auditable in a way that a consultant's field survey is not: the satellite data, the classifier, the training samples and the threshold rules are all reproducible and version-controlled. That auditability is increasingly what regulators in Australia, Canada and South Africa are asking for as bond values on large sites climb into the hundreds of millions of dollars.
Honest limits of the method
Cloud cover is the most persistent operational problem. In the wet tropics, annual cloud-free composites may rest on fewer than five usable observations per year, introducing classification uncertainty that widens confidence intervals on cover estimates. Analysts should report the number of clear observations contributing to each annual composite; regulators should treat classifications built on fewer than three observations with appropriate scepticism.
Thirty-metre pixels cannot resolve fine-scale spatial heterogeneity within a rehabilitation plot. A 30 m pixel that is 40% bare and 60% grass will classify differently depending on the classifier's decision boundary, and two adjacent pixels with the same fractional cover may classify differently due to topographic shading. Sub-pixel spectral unmixing can help, but it introduces its own assumptions. Planet SuperDove at 3–5 m resolves this for targeted areas, at higher cost.
Finally, spectral greenness is not ecological function. A site can achieve regulatory cover thresholds while hosting a monoculture of invasive grass with negligible biodiversity value. Satellite classification verifies structure and cover, not species composition or ecosystem function. It is a necessary condition for bond release, not a sufficient one.
From archive to annual compliance report
Satellize runs this workflow on open constellations (Landsat Collection 2, Sentinel-2 Level-2A) and adds commercial Planet tasking where spatial resolution is the binding constraint. The output is a structured GIS layer and annual compliance report keyed to the client's specific milestone schedule, not a generic land cover product. The same analytical infrastructure that supports Satellize's crop-estimation programme in Tonga handles the multi-year classification stacks and change-detection logic that closure compliance requires.
For mining companies approaching bond release, the most useful first step is a retrospective trajectory analysis: take the site's closure date, pull the full archive, classify it, and measure where the site actually sits against its milestone schedule. That analysis typically takes four to six weeks and produces the evidentiary baseline from which annual monitoring proceeds. It also identifies, early, whether a site is on track or whether the closure plan needs revision before the regulator asks the same question.
Typical figures
| Spatial resolution (routine monitoring) | 30 m (Landsat); 10–20 m (Sentinel-2 MSI) |
| Spatial resolution (targeted verification) | 3–5 m (Planet SuperDove) |
| Revisit cadence | 8 days (Landsat 8+9 combined); 5 days (Sentinel-2A+B combined); near-daily (Planet) |
| Archive depth | Landsat from 1972 (global consistent surface reflectance from 1984 with TM); Sentinel-2 from 2015 |
| Key spectral bands for cover classification | Red, NIR, SWIR-1, SWIR-2 (all Landsat/Sentinel-2); red-edge B5–B8A (Sentinel-2 only) |
| Biomass structural proxy | ALOS-2 PALSAR-2 L-band HV backscatter; 3–10 m depending on acquisition mode |
| Minimum detectable cover change | Approximately 5–10% absolute change in fractional cover at 30 m; finer at 10 m with Sentinel-2 (site-dependent) |
| Cloud contamination impact | Wet-tropical sites may yield fewer than 5 usable annual observations; classification confidence degrades below 3 observations per annual composite |
| Delivery formats | GeoTIFF land cover raster (annual); vector milestone compliance layer; tabular compliance matrix (CSV/XLSX); PDF annual report |
| Latency (annual update) | 4–8 weeks after end of reporting period, depending on cloud cover and composite window |
Analytics Satellize can run
| Multi-decadal land cover trajectory stack | Annual cloud-free compositing (median or percentile) followed by random forest classification trained on site-specific reference data; CCDC for continuous change detection across the archive | Annual GeoTIFF raster stack from closure date to present, with per-pixel class and classification confidence |
| Milestone compliance matrix | Spatial query of classified trajectory against closure plan milestones (cover thresholds, class targets, spatial units); gap analysis against regulatory schedule | Vector GIS layer and tabular report flagging each spatial unit as compliant, non-compliant or data-insufficient at each milestone year |
| Bare-ground recession rate | Time series of bare-ground fractional cover derived from spectral unmixing (MESMA or linear mixture models) applied to Landsat and Sentinel-2 composites | Annual bare-ground fraction map with trend line and projected milestone attainment date |
| Vegetation structural class transition map | Change-vector analysis on NDVI, EVI and SWIR indices to detect pioneer-to-shrub and shrub-to-woodland transitions; validated against ALOS-2 PALSAR-2 L-band HV backscatter as biomass proxy | Transition event map (year and direction of class change) per spatial unit; summary statistics for bond-release submission |
| Regression risk flag | Detection of statistically significant negative NDVI trends within previously classified vegetated pixels using seasonal decomposition (STL or HANTS); distinguishes drought-driven from structural regression | Annual alert layer highlighting areas of confirmed or probable cover regression, with supporting time-series plots |
| Reference ecosystem comparison | Classification of off-site reference polygons (undisturbed analogues specified in the closure plan) using the same annual compositing and classification pipeline; computation of similarity index between rehabilitated and reference cover distributions | Comparative cover distribution table and spatial similarity score, suitable for regulatory equivalence demonstration |
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.