A rehabilitation report shows woody cover rising for five years running. The curve is smooth, the maps are persuasive, and a decision is waiting on it: release an area from intensive management, accept a milestone, renew funding. Then someone asks a question that has little to do with ecology. Was year five measured and analysed the same way as year one?
A reported difference can reflect actual environmental change; differences in acquisition, sampling, timing or conditions; changes in processing, source datasets, classification or interpretation; or a correction that improves an earlier estimate. Only the first is what monitoring exists to detect, and the fourth is welcome. The risk is that a trend line quietly splices all four together.
The proposition worth testing is that a credible trend needs sufficient continuity, or a documented bridge, when its measurement method changes. The evidence supports it, with three qualifications: reproducible is not the same as accurate, many changes are too small to matter for the decision at hand, and a documented break is sometimes the honest answer.
What can change between monitoring rounds
Acquisition effects arise even when software does not change. In a one-day drone study over a Scottish peatland, changing illumination geometry altered reflectance and NDVI between flights; mean NDVI differences were modest (0.01–0.05), but per-pixel differences were larger in places. In the field, a review of observer error in vegetation surveys reported mean pseudoturnover (apparent species change caused by one observer overlooking what another found) of 10–30% in composition surveys.
Calibration and processing effects can be subtler. A NASA-affiliated team documented systematic negative trends in several MODIS Collection 5 products, including NDVI and EVI, caused by sensor ageing; the Collection 6 calibration removed major calibration trends from the Level 1B data. An uncorrected series risked reading instrument degradation as vegetation change.
When a better dataset changes the trend
The clearest documented example is the University of Maryland tree-cover-loss dataset published through Global Forest Watch. Its producers describe three sources of inconsistency: an algorithm refined after the original 2001–2012 version, Landsat 8 data from 2013 that improved detection of small-scale loss, and year-to-year variation in usable imagery. They advise caution when comparing across the 2011 and 2015 changes, or when a trend shifts at the same time as a method. Their most robust option is a sample-based estimate from reference data, for which they point to good-practice guidance on accuracy and area estimation; they note it is independent of the mapping changes but not feasible for every project.
The limits matter. This is one global product, and its history is not an effect size for any other dataset. The producers say single-year area statistics are little affected, and they describe the later data as more sensitive, not less reliable. Nothing here suggests the earlier numbers were careless. It shows that improvement can change the shape of a series.
Small differences can also matter downstream. An Australian comparison of Landsat 7 and 8 found reflectance differences of a few per cent that were magnified when converted to modelled cover: roughly 5% for NDVI and up to 7% for fractional cover before adjustment. A regression fitted on near-coincident image pairs removed the systematic part; a per-pixel difference remained, and the relationship varied by region. These were early, 2014 results, so they illustrate a principle rather than current magnitudes.
Routine updates still need judgement
Routine maintenance happens beneath every long series. On 20 October 2026 the DEA Sandbox, an analysis environment rather than a data product, moved from Python 3.10 to 3.12 with updates to several analysis packages. Users relying on particular package versions or custom workflows were advised to review the changes and assess any impacts. The alert does not describe a change to DEA data products, and nothing in it suggests environmental outputs differ. It simply shows the kind of change a provider announces and a user must judge. When DEA replaced a discontinued water-vapour input to its surface-reflectance processing earlier in 2026, it flagged interim data and backfilled affected products.
Compare the same observations through both workflows
One approach worth proposing is to process the same archived observations through the old and new workflows, holding other inputs and settings constant where feasible, and to compare results in the units the decision depends on. This is proposed practice, not a requirement drawn from any source cited here.
It can isolate what the processing change did to identical inputs, removing season, weather and survey effort from the comparison. It cannot reveal acquisition differences or ecological change, and it cannot say which workflow is closer to the truth: two workflows can agree and share a bias. Accuracy needs independent reference data.
The sample should cover the conditions and classes that matter, including difficult cases and areas near decision thresholds, because the studies above show effects varying with target, region and conditions, and a few scenes can mislead. Differences are then read against measurement uncertainty and the margin between the current estimate and the management threshold. No universal tolerance applies. Where differences are patterned by class, sit near a threshold, or an adjustment factor is contemplated, specialist statistical and ecological advice is warranted. Retaining inputs, settings and methods notes is what makes the check possible; it is also part of sound GIS and spatial analysis.
For commissioning and handover across providers, see our guide to retaining the project record across contractor changes.
A rehabilitation decision: a hypothetical example
Hypothetical illustration only. It is not a TALONYX project and implies no real results.
A program owner must decide whether a rehabilitation area can move from active maintenance to monitoring only. Reported canopy cover now exceeds the completion criterion, but the provider changed its classification workflow after year four. A matched rerun of archived year-three and year-four imagery could lead to different decisions. If the two workflows agree closely relative to the gap between estimate and criterion, the owner records the change and proceeds. If they diverge mainly in shaded canopy or grass-dominated patches, the apparent gain should not be treated as recovery until field plots or a reference sample confirm it, and the series is reported with a break at the change. If the archived imagery was not kept, no rerun is possible; an independent reference sample would estimate how well each period holds up, and inputs are retained from now on. In none of these cases does the check show whether the ecosystem has recovered. It shows whether the measured change can be trusted as evidence of that, which is the question for rehabilitation and restoration monitoring in the first place.
Reprocessing, overlap or a documented break
Sometimes another route is better. If a provider can reprocess history, that is often the cleaner bridge: the Copernicus Sentinel-2 archive was reprocessed into Collection-1 to give largely consistent processing baselines from 2015. When a sensor or survey method changes, overlapping capture lets users see the offset directly, as when MODIS Collections 5 and 6 ran in parallel. If neither is feasible, a documented break, with the periods reported separately, is a legitimate outcome. Where old inputs or workflows are gone, an independent reference sample is the fallback.
A fair objection is that improved methods should replace weaker ones. They should. Global Forest Watch’s producers accept that perfect consistency is neither always possible nor always desirable; the aim is to flag the change so a revised number is not read as environmental change. Another is cost. The check is for material changes to consequential decisions: a new sensor, calibration, source dataset, classifier or training data, or definition. A routine environment update, like the one above, usually calls for recording versions and rerunning an existing test case, not reanalysing the archive. Nor should unsupported software be kept indefinitely; retaining inputs, parameters and a reference output serves better.
Questions to agree before the next monitoring round
- What may change between rounds—sensor, calibration, season, survey effort, processing, classification, source data—and how will changes be recorded?
- Will inputs, settings and a reference output be retained so a past result can be reproduced?
- If the method changes, will the response be overlap, reprocessing or a declared break, and who decides?
- How large is the margin between the current estimate and the decision threshold, relative to uncertainty?
- Is there independent field or reference evidence, not just agreement between two workflows?
- How will improvements to earlier estimates be reported, so they are not mistaken for environmental change?
Keep the evidence useful through change
Agree how changes will be recorded, checked and reported before the next monitoring round. The aim is evidence that remains useful for the decision, even as methods improve.
Discuss a projectSources and further reading
- DEA Tech Alerts
- DEA Sandbox release 2.2.1
- Global Forest Watch: assessing trends in tree cover loss
- Olofsson et al. 2014, good practices for estimating area and assessing accuracy of land change
- Lyapustin et al. 2014, MODIS C5 calibration degradation
- Flood 2014, Landsat 7 and 8 continuity in the Australian landscape
- Stow et al. 2019, illumination geometry in multispectral drone imagery
- Morrison 2016, observer error in vegetation surveys
- SentiWiki, Sentinel-2 processing and Collection-1
- The Image Is Not the Ecosystem
