Written and maintained by CASRAI Editorial Board
Published
On this page
Land cover change is a difference in the physical material covering the land surface (forest, cropland, water, built surface and so on) between two or more dates. It is usually measured by classifying satellite or aerial imagery for each date and comparing the results, and the number that comes out is only as trustworthy as the accuracy assessment behind it. This guide defines the term, shows how change is measured, and works through a three-class example in which simply counting changed map pixels understates the true area of loss.
Land cover change defined: land cover versus land use
Land cover describes what is on the surface; land use describes what people do with that land. A pine plantation and a natural pine stand can share one land cover class and have different uses, which is why land use planning and land cover mapping are related but not interchangeable. In the research literature the two are often combined as land use and land cover change (LULCC), but a change in cover is something a sensor can see, while a change in use often needs ancillary records.
Two kinds of change are worth separating. Conversion replaces one class with another (forest to cropland). Modification alters the condition within a class (forest thinning or degradation) and is much harder to detect from a two-date classification. The Intergovernmental Panel on Climate Change’s special report on land use, land-use change and forestry notes that where an activity is forest to non-forest change, most such activities will be readily discernible, but where definitions include land that stays forest, simple forest classification at the two endpoints would be insufficient.
How land cover change is measured
Two-date classification and post-classification comparison
The most common design classifies imagery from an earlier and a later date into the same legend and cross-tabulates the two maps. Each pixel then falls in a from-to class pair; pairs where the class differs are mapped as change. Because each map carries its own classification error, errors in the two dates compound in the change layer.
A concrete dataset: the National Land Cover Database
The U.S. Geological Survey describes the National Land Cover Database (NLCD) as the definitive land cover database for the United States, with 30 m resolution over the conterminous states. Its legacy releases cover nine epochs (2001, 2004, 2006, 2008, 2011, 2013, 2016, 2019 and 2021), the 2021 epoch is described as comparable with 2019 and earlier epochs, and Annual NLCD is the latest release. The legend has 16 classes based on a modified Anderson Level II classification system. Change is shown through a change index across all epochs, and the science products include a Land Cover Change Disturbance Date, a Forest Disturbance Date (1984 to 2021) and a Land Cover Change Count between consecutive epochs. The data table below lists the epoch intervals, which are unequal; any rate of change per year therefore has to be divided by the actual interval, not by a fixed step.
USGS also reports accuracy for the product family. For NLCD 2001 the overall accuracies were 78.7% at Anderson Level II and 85.3% at Level I; for the 1992 product they were 58% and 80%. Those figures show the practical rule behind the rest of this guide: accuracy depends strongly on how finely the legend divides the land, and a change map inherits the weakness.
Why error matrices matter for change area
Olofsson, Foody, Stehman and Woodcock (2013) showed that the area of change counted directly from a map can differ greatly from the true area because of classification error, and recommended a stratified estimator built on a sample of reference labels. Olofsson and colleagues (2014) set out good practices: probability sampling, reference labels that are more accurate than the map, an analysis consistent with the sampling design, and an error matrix reported as estimated area proportions with confidence intervals. Foody (2010) had earlier shown that imperfect reference data can introduce large error into land cover change studies.
Formula for error-adjusted change area
Let a change map have q classes (strata). Draw a probability sample and label each sampled location in the reference data.
- Wi: the share of the mapped area that lies in map class i (mapped area of class i divided by total mapped area).
- nij: the number of sample units mapped as class i whose reference label is class j.
- ni: the number of sample units in map class i, the sum of nij over j.
- pij = Wi × nij / ni: the estimated proportion of the total area that is mapped as i and truly j.
- p·j: the column sum of pij over i, the estimated proportion of the area that truly belongs to class j. Multiply by total area A to get the estimated area of class j.
- Overall accuracy = sum of the diagonal pii; user’s accuracy of class i = pii / pi·, which equals nii / ni; producer’s accuracy of class j = pjj / p·j.
- Standard error of p·k = square root of the sum over i of Wi² × (nik/ni) × (1 − nik/ni) / (ni − 1); an approximate 95% interval is the estimate ± 1.96 × SE.
Worked example: three-class change map
This example is an illustrative composite with invented numbers, built only to show the arithmetic; it does not describe a real study area. A map of 100,000 ha has three classes: Stable (90,000 ha mapped), Loss (6,000 ha) and Gain (4,000 ha), so W = 0.90, 0.06 and 0.04. One hundred locations are sampled in each mapped class and labelled from reference imagery.
| Map class (rows) / reference (columns) | Stable | Loss | Gain | ni |
|---|---|---|---|---|
| Stable | 97 | 2 | 1 | 100 |
| Loss | 10 | 88 | 2 | 100 |
| Gain | 12 | 0 | 88 | 100 |
Step 1, area proportions pij = Wi × nij / ni. Stable row: 0.90 × 0.97 = 0.873, 0.90 × 0.02 = 0.018, 0.90 × 0.01 = 0.009. Loss row: 0.06 × 0.10 = 0.006, 0.06 × 0.88 = 0.0528, 0.06 × 0.02 = 0.0012. Gain row: 0.04 × 0.12 = 0.0048, 0, 0.04 × 0.88 = 0.0352.
Step 2, column sums. Stable 0.873 + 0.006 + 0.0048 = 0.8838; Loss 0.018 + 0.0528 + 0 = 0.0708; Gain 0.009 + 0.0012 + 0.0352 = 0.0454. The three add to 1.0000, as they must.
Step 3, estimated areas: 0.8838 × 100,000 = 88,380 ha stable, 0.0708 × 100,000 = 7,080 ha loss and 0.0454 × 100,000 = 4,540 ha gain. Counting the map directly gives 6,000 ha of loss, so the map understates loss by 1,080 ha (about 15% of the estimate). Two errors pull in opposite directions: 10% of the mapped Loss class is truly stable (commission, which inflates the map), but 2% of the very large Stable class is truly loss, and 0.018 × 100,000 = 1,800 ha of loss hides inside it (omission, which outweighs the commission).
Step 4, accuracies. Overall = 0.873 + 0.0528 + 0.0352 = 0.961. User’s: Stable 97/100 = 0.97, Loss 88/100 = 0.88, Gain 88/100 = 0.88. Producer’s: Stable 0.873 / 0.8838 = 0.988, Loss 0.0528 / 0.0708 = 0.746, Gain 0.0352 / 0.0454 = 0.775.
Step 5, uncertainty for loss (k = Loss). SE² = 0.9² × 0.02 × 0.98 / 99 + 0.06² × 0.88 × 0.12 / 99 + 0.04² × 0 = 0.000160 + 0.0000038 = 0.000164, so SE = 0.01281 and the 95% interval half-width is 1.96 × 0.01281 × 100,000 ≈ 2,512 ha. Estimated loss: about 7,080 ± 2,510 ha. The mapped value of 6,000 ha falls inside that interval, but the interval is wide because a small error rate in the very large Stable class matters.
Interpretation: reading the numbers
| Result | What it tells you | Reading guidance from the example |
|---|---|---|
| Overall accuracy 96.1% | Share of area correctly mapped | Dominated by the large Stable class, so it can look high even when change classes are weak |
| User’s accuracy of Loss 88% | How reliable a pixel mapped as Loss is | 12% of mapped Loss is commission error |
| Producer’s accuracy of Loss 74.6% | How much true loss the map found | About a quarter of true loss is omitted |
| Estimated vs mapped Loss area | Direction of bias | Omission outweighed commission, so the map understates loss |
| 95% interval ± 2,512 ha | Sampling uncertainty of the area | Report with the estimate; a change smaller than the interval cannot be claimed |
Assumptions and common mistakes
- Counting pixels as area. Direct pixel counts are not error-adjusted; the 2013 and 2014 papers recommend the stratified estimate instead.
- Reporting overall accuracy alone. Rare change classes are swamped by the stable majority, so report per-class user’s and producer’s accuracy.
- Weak reference data. The stratified method assumes the reference label is more accurate than the map; Foody (2010) found that imperfect reference data can introduce large error.
- Ignoring the design. Sample weights Wi must be used. Treating a stratified sample as a simple random sample biases both the areas and the accuracies.
- Using two-date classes for modification. As the IPCC report notes, land that stays forest cannot be assessed by classifying the endpoints alone; biomass-based forest definitions need field inventory.
- Comparing epochs as if they were equal. NLCD epochs are two or three years apart, and different product generations are not always comparable.
How to report land cover change
Following Olofsson et al. (2013), report the sampling design, the error matrix, the mapped area per category, and user’s, producer’s and overall accuracy, with confidence intervals for the change areas. An APA-style sentence using the illustrative numbers above would read: “Error-adjusted forest loss was 7,080 ha (95% CI [4,568, 9,592]), compared with 6,000 ha counted directly from the map; overall accuracy was 96.1%, with user’s and producer’s accuracy for loss of 88.0% and 74.6% (Olofsson et al., 2014).” Where the change layer comes from NLCD or a similar product, name the product, epoch pair and legend level. For background, see what is geography, and for the imagery side see what is remote sensing. Elevation or canopy-structure data from LiDAR point clouds can serve as ancillary reference information.
How to calculate it in Excel, R or Python
Excel
Put the count matrix in B2:D4 (rows Stable, Loss, Gain), the mapped area shares W in F2:F4 and ni in E2:E4. In H2 enter =$F2*B2/$E2 and fill across to J4 to get pij. Estimated area of each class is =SUM(H$2:H$4)*100000 in H5, filled right. Overall accuracy is =H2+I3+J4.
R
The CRAN package mapaccuracy provides olofsson(r, m, Nh, margins = TRUE), where r holds the reference labels, m the map labels and Nh the mapped area (or pixel count or proportion) per class. It returns overall, user’s and producer’s accuracy, estimated areas and their standard errors, and the area-proportion error matrix.
library(mapaccuracy)
cls = c("Stable", "Loss", "Gain")
m = rep(cls, each = 100)
r = c(rep(cls, c(97, 2, 1)), rep(cls, c(10, 88, 2)), rep(cls, c(12, 0, 88)))
Nh = c(Stable = 90000, Loss = 6000, Gain = 4000)
olofsson(r, m, Nh)
Python
import numpy as np
n = np.array([[97, 2, 1], [10, 88, 2], [12, 0, 88]])
W = np.array([0.90, 0.06, 0.04])
p = W[:, None] * n / n.sum(axis=1, keepdims=True)
area = p.sum(axis=0) * 100000 # 88380, 7080, 4540
oa = np.trace(p) # 0.961
pa = np.diag(p) / p.sum(axis=0) # producer's accuracy
ni = n.sum(axis=1, keepdims=True)
se = np.sqrt((W[:, None]**2 * (n/ni) * (1 - n/ni) / (ni - 1)).sum(axis=0))
print(area, oa, pa, 1.96 * se * 100000)
Frequently asked questions
What is land cover change?
It is a change in the physical surface material of an area between dates, such as forest becoming cropland, identified by comparing classified imagery or other observations.
What is the difference between land cover change and land use change?
Land cover is what the surface consists of; land use is the human purpose it serves. Cover change can be detected from imagery, whereas use change often needs records beyond the image.
How is land cover change measured?
Typically by classifying imagery from two dates into one legend and cross-tabulating them, then estimating change area from a reference sample so that classification error is accounted for.
Why does the mapped area of change differ from the true area?
Classification error. Olofsson et al. (2013) found that map-derived change area can differ greatly from true area, which is why they recommend a stratified estimator.
How accurate is NLCD?
USGS cites overall accuracies of 78.7% (Anderson Level II) and 85.3% (Level I) for NLCD 2001, and 58% and 80% for the 1992 product.
Sources
- U.S. Geological Survey, National Land Cover Database, https://www.usgs.gov/centers/eros/science/national-land-cover-database (checked 2026-11-15).
- IPCC, Special Report on Land Use, Land-Use Change and Forestry, chapter 3, https://grida.no/climate/ipcc/land_use/139.htm (checked 2026-11-15).
- CRAN, mapaccuracy: olofsson() reference manual, https://search.r-project.org/CRAN/refmans/mapaccuracy/html/olofsson.html (checked 2026-11-15).
- Olofsson, P., Foody, G. M., Stehman, S. V., and Woodcock, C. E. (2013). Making better use of accuracy data in land change studies. Remote Sensing of Environment, 129, 122-131.
- Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., and Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment, 148, 42-57. https://doi.org/10.1016/j.rse.2014.02.015
Data: NLCD legacy epoch intervals, 2001-2021
Intervals between consecutive legacy National Land Cover Database epochs, derived from the nine epoch years listed on the USGS NLCD page. Useful for converting change between epochs into annual rates.
Period covered: 2001/2021. Coverage: Conterminous United States.
| Epoch pair | Start epoch (year) | End epoch (year) | Interval (years) |
|---|---|---|---|
| 2001-2004 | 2001 | 2004 | 3 |
| 2004-2006 | 2004 | 2006 | 2 |
| 2006-2008 | 2006 | 2008 | 2 |
| 2008-2011 | 2008 | 2011 | 3 |
| 2011-2013 | 2011 | 2013 | 2 |
| 2013-2016 | 2013 | 2016 | 3 |
| 2016-2019 | 2016 | 2019 | 3 |
| 2019-2021 | 2019 | 2021 | 2 |
Sources
- USGS National Land Cover Database (accessed 2026-11-15)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







