Written and maintained by CASRAI Editorial Board
Published
On this page
A land cover map is a raster or vector map that assigns every location a class describing the physical surface, such as forest, open water, impervious surface or cropland, usually by analysing satellite or aerial imagery. Because every class is defined by a rule and every map contains classification errors, a land cover map is only as useful as its legend and its accuracy assessment: the error matrix, overall accuracy and the per-class user’s and producer’s accuracies. This guide explains what the classes mean, how the maps are made, and how to compute and report their accuracy with a worked example.
What is a land cover map?
A land cover map records what physically covers the ground. The US National Oceanic and Atmospheric Administration puts it this way: land cover indicates the physical land type such as forest or open water, and its examples include forests, wetlands, impervious surfaces and agriculture. Maps of this kind are made by analysing satellite and aerial imagery, and tracking change requires land cover maps for several years (NOAA National Ocean Service, Land cover vs. land use).
In a research setting the map is a measurement instrument: each cell carries a label, and the label is the output of a classification procedure with a known error rate. That is why a map should be cited with its product name, year, cell size and class scheme, and why its accuracy figure belongs next to any result computed from it. Spatial analyses built on such a map, including autocorrelation and spatial sampling, are covered in spatial data analysis, and the imagery behind it in what is remote sensing.
Land cover vs. land use
Land cover and land use are different things, and a land cover map cannot tell you the second. NOAA defines land use as how people are using the land, with examples such as development, conservation and mixed uses, and states that land use cannot be determined from satellite imagery. A field of grass, for instance, can be cover class “grassland” while its use is pasture, a park or a utility corridor. Studies that need use, such as zoning or planning questions, need ancillary records; see land use planning for that side.
Classes and legends: what a class label actually means
A land cover class is not a natural category but a rule. The National Land Cover Database (NLCD), produced by the Multi-Resolution Land Characteristics (MRLC) Consortium, uses a classification system modified from the Anderson Land Cover Classification System, and its legend defines classes with numeric thresholds. Developed Open Space is where impervious surfaces account for less than 20% of total cover; Low Intensity is 20% to 49%; Medium Intensity is 50% to 79%. Deciduous Forest is dominated by trees generally taller than 5 metres with more than 20% of total vegetation cover; Shrub/Scrub is shrubs under 5 metres with canopy typically above 20% of total vegetation; Grassland/Herbaceous is generally more than 80% of total vegetation.
The data table below lists ten NLCD classes with the threshold that defines each. The practical consequence is that two maps can both show a “forest” class and still disagree about what is forest, because one may use a 10% canopy rule and another 20%. Before comparing maps, compare legends, not just names.
How land cover maps are produced
Production moves from imagery to a classifier to a validated product. The Penn State e-Education case study of the NLCD describes two generations. NLCD 1992 used leaf-off 1992 Landsat Thematic Mapper bands: an unsupervised classification produced 100 spectral clusters, which analysts then assigned to classes using aerial photographs, giving 21 Level II classes that could be aggregated to nine Level I classes. From NLCD 2001 onward the inputs grew to 18 or more layers, including multi-season Landsat 5 and 7 imagery and elevation derivatives, and the classifier became a supervised decision tree trained on orthophotos, local data, field points and Forest Inventory and Analysis data, with the urban classes taken from an imperviousness layer. Every edition was mosaicked into the Albers Conic Equal Area projection on NAD83 with 30 m cells.
Two methodological points follow. A supervised classifier needs labelled training samples, and the samples used to train a map must not be reused to assess it. And a 30 m cell is a mixed pixel: a cell that is part road and part lawn is forced into a single class, which is one source of the errors measured below.
Accuracy assessment: the error matrix
How accurate is a land cover map?
Accuracy is measured by comparing the mapped class with a reference class at a sample of locations, then tabulating the pairs in an error matrix (also called a confusion matrix). In the Penn State course text on image classification, the matrix matches each sample’s assigned class with its ground truth, rows hold the actual classes, columns hold the predicted classes, and the diagonal holds the correct classifications. Published figures show why this matters: the same case study reports a 38% to 62% chance that an NLCD 1992 pixel was correct, with USGS advising users to aggregate classes, and overall Level II accuracies of 79% for NLCD 2001 and 78% for NLCD 2006 (Wickham et al. 2013).
Formulas
Let n be the total number of reference samples, nii the diagonal count for class i (map and reference agree), ni+ the row total (reference samples of class i) and n+i the column total (samples the map labelled i).
- Overall accuracy = Σ nii ÷ n, the proportion of all samples correctly classified.
- Producer’s accuracy (class i) = nii ÷ ni+. It reflects omission errors: reference locations of the class that the map left out. Omission error = 1 − producer’s accuracy.
- User’s accuracy (class i) = nii ÷ n+i. It reflects commission errors: locations the map put in the class that do not belong. Commission error = 1 − user’s accuracy.
These definitions match those in the Penn State GEOG 892 chapter on digital image classification for land use and land cover assessment, where overall accuracy is the correct samples divided by all samples, producer’s accuracy reflects omission and user’s accuracy reflects commission.
Worked example with real arithmetic
The matrix below is an illustrative composite invented for this guide; it does not describe any real map. Imagine 200 reference samples checked against a three-class map. Rows are the reference class, columns the mapped class.
| Reference Map | Forest | Water | Developed | Row total |
|---|---|---|---|---|
| Forest | 70 | 3 | 7 | 80 |
| Water | 2 | 38 | 0 | 40 |
| Developed | 8 | 1 | 71 | 80 |
| Column total | 80 | 42 | 78 | 200 |
Overall accuracy: the diagonal is 70 + 38 + 71 = 179, so 179 ÷ 200 = 0.895, or 89.5%.
Producer’s accuracy (diagonal ÷ row total): Forest 70 ÷ 80 = 87.5%; Water 38 ÷ 40 = 95.0%; Developed 71 ÷ 80 = 88.75%. User’s accuracy (diagonal ÷ column total): Forest 70 ÷ 80 = 87.5%; Water 38 ÷ 42 = 90.48%; Developed 71 ÷ 78 = 91.03%. Water has the best producer’s accuracy but the lowest user’s accuracy: almost every real water sample was found, but 4 of the 42 cells the map called water were something else (3 reference Forest and 1 reference Developed).
Kappa corrects agreement for chance. With row and column totals above, expected agreement is (80×80 + 40×42 + 80×78) ÷ 200² = 14,320 ÷ 40,000 = 0.358, so κ = (0.895 − 0.358) ÷ (1 − 0.358) = 0.837. Kappa is a chance-corrected agreement index described in Cohen’s kappa for two raters; here the two “raters” are the map and the reference data.
Interpretation: reading the numbers
| What you see | What it tells you |
|---|---|
| High overall accuracy, one weak class | The weak class is small or hard; do not use the map to study that class without checking its own accuracy. |
| Low producer’s accuracy for a class | The map omits real examples of the class, so the mapped area of it is likely an underestimate. |
| Low user’s accuracy for a class | Many mapped cells of the class are wrong, so a cell picked at random from that class is less trustworthy. |
| Accuracy reported only at an aggregated level | Class-level accuracy at the detailed level is lower; the NLCD 1992 advice to aggregate classes exists for this reason. |
Assumptions and common mistakes
- Treating land cover as land use. A cover class says what is on the surface, not how it is used.
- Comparing maps with different legends. Class names match while thresholds differ; compare the definitions first.
- Detecting change between editions without checking comparability. Differences in inputs and method between editions can look like real change; check the product’s own documentation before differencing two years.
- Reusing training samples for validation. The reference sample should be independent of the data used to build the classifier.
- Quoting overall accuracy alone. Report per-class user’s and producer’s accuracy as well.
- Ignoring cell size. A 30 m cell mixes surface types; very small features may not be representable.
How to report a land cover map in a methods section
State the product name, edition or year, cell size, class scheme and the source of the accuracy figure. For example: “Land cover was taken from the National Land Cover Database for [year] (30 m cells; Anderson-derived legend) and aggregated to three classes. In an independent sample of 200 locations, overall accuracy was 89.5% and κ = 0.84; user’s accuracy ranged from 87.5% to 91.0% and producer’s accuracy from 87.5% to 95.0%.” Replace the figures with those of your own validation or the product’s published assessment, and cite the dataset in your reference style.
How to calculate accuracy in Excel, R or Python
Each snippet assumes the matrix above, with reference classes in rows.
Excel
Put the matrix in B2:D4 (reference in rows, map in columns). Overall accuracy: =(B2+C3+D4)/SUM(B2:D4). Producer’s accuracy for Forest: =B2/SUM(B2:D2). User’s accuracy for Forest: =B2/SUM(B2:B4). Fill the formulas down the diagonal for the other classes.
R
m = matrix(c(70,3,7, 2,38,0, 8,1,71), nrow = 3, byrow = TRUE)
sum(diag(m)) / sum(m) # overall accuracy: 0.895
diag(m) / rowSums(m) # producer's accuracy
diag(m) / colSums(m) # user's accuracy
Python
m = [[70, 3, 7], [2, 38, 0], [8, 1, 71]] # rows = reference
n = sum(map(sum, m))
rows = [sum(r) for r in m]
cols = [sum(r[j] for r in m) for j in range(3)]
diag = [m[i][i] for i in range(3)]
print(sum(diag) / n) # overall: 0.895
print([d / r for d, r in zip(diag, rows)]) # producer's
print([d / c for d, c in zip(diag, cols)]) # user's
For library-based confusion matrices see scikit-learn confusion matrix; note that scikit-learn puts true labels in rows by default, matching the layout used here.
Sources
- NOAA National Ocean Service, Land cover vs. land use
- Multi-Resolution Land Characteristics Consortium, NLCD class legend and description
- Penn State e-Education, Case Study: Image Classification for the National Land Cover Dataset
- Penn State e-Education, Digital Image Classification for Land Use Land Cover (LULC) Assessment, Part 2
Data: NLCD land cover classes and their defining thresholds
Ten classes from the National Land Cover Database legend, with the class group and the numeric or descriptive threshold the legend uses to define each class. Compiled from the MRLC class legend and description page.
Coverage: United States.
| NLCD code | Class | Group | Defining threshold in the legend |
|---|---|---|---|
| 11 | Open Water | Water | Generally less than 25% cover of vegetation or soil |
| 12 | Perennial Ice/Snow | Water | Perennial ice and/or snow, generally greater than 25% of total cover |
| 21 | Developed, Open Space | Developed | Impervious surfaces less than 20% of total cover |
| 22 | Developed, Low Intensity | Developed | Impervious surfaces 20% to 49% of total cover |
| 23 | Developed, Medium Intensity | Developed | Impervious surfaces 50% to 79% of total cover |
| 24 | Developed, High Intensity | Developed | Impervious surfaces 80% to 100% of total cover |
| 31 | Barren Land | Barren | Bedrock, desert pavement, scarps, talus, slides, volcanic material, glacial debris; less than 15% vegetation |
| 41 | Deciduous Forest | Forest | Trees generally taller than 5 m and more than 20% of total vegetation cover |
| 52 | Shrub/Scrub | Shrubland | Shrubs under 5 m tall, canopy typically more than 20% of total vegetation |
| 71 | Grassland/Herbaceous | Herbaceous | Graminoid or herbaceous vegetation, generally more than 80% of total vegetation |
Sources
- MRLC NLCD class legend and description (accessed 2026-11-18)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







