What is Land Cover Classification? Classes and Data

Land cover classification is the process of labeling each part of the Earth's surface by the physical material that covers it, such as water, tree canopy, grass, crops, bare soil, or pavement. In remote sensing, a classifier assigns every pixel or image segment of satellite or aerial imagery to one class from a fixed legend, and the result is a land cover map. National agencies, insurers, planners, and climate scientists use these maps to measure where forests, farms, and cities are, and how they change.

Key takeaways

• Land cover is what physically covers the ground. Land use is what people do with it.
• A land cover classification system, such as Anderson, NLCD, CORINE, or the FAO LCCS, defines the classes and how they nest.
• Classifiers range from unsupervised clustering and supervised random forests to deep learning segmentation models.
• Every map needs an accuracy assessment against independent reference samples. Three 10 m global maps scored between 65% and 75% overall accuracy against the same reference data.
• Open maps range from 1 m regional products to 10 m and 30 m national and global ones, and Annual NLCD covers every year from 1985 to 2025.
• In Wherobots, ESA WorldCover rasters and Overture land cover polygons are queryable in SQL, and RasterFlow runs segmentation models over imagery.

Land cover vs land use

Land cover and land use describe the same ground from two angles. Overture Maps defines land cover as the physical thing covering the land, and land use as the human use to which the land is being put.

	Land cover	Land use
Describes	Physical surface material	Human purpose
Examples	Tree canopy, grass, water, pavement	Housing, farming, recreation, industry
Primary source	Satellite and aerial imagery	Parcels, zoning, surveys, imagery
Golf course	Grass and tree canopy	Recreation

Imagery records land cover directly, because pixel values respond to surfaces. Land use needs context: a field of grass reads the same in imagery whether it is a park, a pasture, or a vacant lot. Many products combine the two and are labeled land use land cover (LULC).

Land cover classes and classification systems

A land cover classification system defines the classes and how they nest, so maps from different agencies and years stay comparable. Legends follow purpose: NASA's Earth Observatory notes that land cover maps range from two categories for flood maps to seventeen for standard global maps.

The different types of land cover recur across systems: water, forest or tree cover, shrubland, grassland, cropland, wetland, developed or built-up land, barren land, and snow and ice. The best-known five-class level is the top of CORINE Land Cover, Europe's inventory: artificial surfaces, agricultural areas, forests and semi-natural areas, wetlands, and water bodies, split into 15 classes at level two and 44 at level three.

Land cover classification from remote sensor data began as a federal standards problem. The Anderson system, USGS Professional Paper 964 by Anderson, Hardy, Roach, and Witmer (1976), fixed the first two levels of a national legend so that maps from different agencies could be added together, and left the third and fourth levels open for agencies to extend.

The National Land Cover Database (NLCD) still uses a modified Anderson Level II, with two-digit codes:

Code	NLCD class	Code	NLCD class
11	Open Water	52	Shrub/Scrub
21	Developed, Open Space	71	Grassland/Herbaceous
22	Developed, Low Intensity	81	Pasture/Hay
23	Developed, Medium Intensity	82	Cultivated Crops
24	Developed High Intensity	90	Woody Wetlands
31	Barren Land	95	Emergent Herbaceous Wetlands
41, 42, 43	Deciduous, Evergreen, Mixed Forest	12	Perennial Ice/Snow

The developed classes are split by impervious surface: Developed, Open Space has less than 20% impervious cover, and Developed, Low Intensity has 20% to 49%.

Global products use the UN Food and Agriculture Organization's Land Cover Classification System (LCCS), which Di Gregorio and Jansen published in 2000. It takes the opposite approach to Anderson's fixed levels: a dichotomous, modular-hierarchical system in which each class is built from independent diagnostic criteria, such as life form and cover density, so any legend can be translated into any other. ESA WorldCover defines its 11 classes with LCCS, from Tree cover and Cropland to Built-up and Mangroves. MODIS global maps use the 17-class IGBP legend, and the Dynamic World taxonomy stays close to the IPCC land categories: forest land, grassland, cropland, wetland, settlement, and other.

How land cover classification works

Every land cover map follows the same steps, whether a person or a neural network draws the boundaries.

1. Choose the imagery. Pick a sensor and season, then build a cloud-free composite, often a per-pixel median across many dates.
2. Collect training labels. Annotate sample areas with the target classes, by hand or from an existing map.
3. Extract features. Use spectral bands, indices such as NDVI, texture, elevation, and the time series of each pixel.
4. Classify. Apply a classifier to every pixel or segment.
5. Clean up. Remove isolated pixels and enforce a minimum mapping unit.
6. Assess accuracy. Compare the map against independent reference samples.

The American Museum of Natural History's guide to land cover classification methods by Ned Horning, the top result for the term, sorts classifiers into a few families:

• Unsupervised classification. A clustering algorithm such as k-means or ISODATA groups similar pixels, and an analyst names each cluster afterward.
• Supervised classification. The analyst supplies labeled sample pixels and the algorithm is trained on them. Breiman's random forests (Machine Learning, 2001) became a standard supervised classifier for land cover, because they handle many correlated features and report variable importance. Belgiu and Drăguţ's review (2016) documents their spread through remote sensing.
• Object-based image analysis (OBIA). Pixels are first grouped into segments, which are classified by their spectra, shape, and context. Blaschke's Object based image analysis for remote sensing (2010) traces how the unit of analysis moved from the pixel to the segment as pixels became smaller than the objects being mapped.
• Deep learning segmentation. A convolutional network such as the encoder-decoder U-Net (Ronneberger, Fischer, and Brox, 2015) labels every pixel from its neighborhood, which separates a shadowed roof from water and a lawn from a crop field. Zhu and colleagues' Deep learning in remote sensing (IEEE Geoscience and Remote Sensing Magazine, 2017) and Ma and colleagues' meta-analysis (ISPRS Journal, 2019) record the field's move to these networks.

Resolution decides which classes are possible. At 30 m a pixel mixes a house, its yard, and the street, so NLCD measures developed intensity by percent impervious cover. At 1 m, the Chesapeake Conservancy maps structures, roads, and tree canopy over roads as separate classes. The header image above shows the same strip of Rockville, Maryland at 1 m and 10 m: the 1 m map separates single houses and roads, and the 10 m map merges them into built-up blocks.

Accuracy assessment compares the map with reference labels at sample locations and tabulates the result in an error matrix, the method Congalton set out in his 1991 review. The matrix gives overall accuracy, user's accuracy (how often a mapped class is right on the ground), and producer's accuracy (how much of a true class the map found). Olofsson and colleagues' good practices (2014) add a probability sampling design and area estimates with confidence intervals, because pixel counts from a map carry its errors.

Land cover classification examples

These well-known maps show the range of satellite image land cover classification:

• NLCD and Annual NLCD. The USGS maps the conterminous US at 30 m from Landsat. Annual NLCD Collection 1.2, released in June 2026, covers every year from 1985 to 2025 with 16 classes and uses deep learning models.
• ESA WorldCover. A 10 m global map from Sentinel-2 and Sentinel-1 radar. ESA reports independently validated overall accuracy of 74.4% for the 2020 map and 76.7% for 2021.
• Dynamic World. Brown and colleagues (2022) trained a deep learning model on 10 m Sentinel-2 imagery and publish class probabilities for each new Sentinel-2 image, so a map can cover any date range.
• Esri and Impact Observatory 10 m maps. Karra and colleagues trained a segmentation model on over 5 billion human-labeled Sentinel-2 pixels (IGARSS 2021) for annual global maps.
• Global forest change. Hansen and colleagues mapped 21st-century forest loss and gain at 30 m from Landsat (Science, 2013).
• Chesapeake Bay 1 m land cover. The Chesapeake Conservancy maps the watershed from aerial imagery and lidar, with 900 times more detail than 30 m NLCD.
• Pretrained models. Esri's Land Cover Classification (Landsat 8) model is a U-Net trained on NLCD 2016, so it outputs NLCD classes from new Landsat scenes.

Uses follow from the classes. Impervious surface drives runoff, so land cover feeds flood risk models and stormwater fees. Forest extent and loss feed carbon accounting. Vegetation near structures drives wildfire exposure in catastrophe models. Planners start tree canopy targets and heat mapping from a land cover classification map, and the Chesapeake Bay Program uses its 1 m data to track development and habitat change across the watershed.

Land cover datasets

Resolution, update frequency, and class detail trade off against each other.

Dataset	Coverage	Resolution	Classes	Time span
Annual NLCD	Conterminous US	30 m	16	1985 to 2025, yearly
USGS LCMAP	Conterminous US	30 m	8	1985 to 2021, yearly
ESA WorldCover	Global	10 m	11	2020 and 2021
Dynamic World	Global	10 m	9, as probabilities	Each Sentinel-2 image
Impact Observatory 10 m LULC	Global	10 m	9	2017 to 2023, yearly
CORINE Land Cover	Europe	Vector inventory	44	Since 1990, updated periodically
Chesapeake LULC	Chesapeake Bay watershed	1 m	13 to over 50	2013/14 to 2021/22

Land cover change

Land cover change is the shift from one class to another between dates, such as forest to development or pasture to cropland. Annual NLCD publishes a land cover change product that concatenates the before and after codes: 8123 means Pasture/Hay became Developed, Medium Intensity, and 7182 means Grassland/Herbaceous became Cultivated Crops.

Comparing two independent maps exaggerates change, because each map's errors register as false change. ESA warns that differences between its 2020 and 2021 WorldCover maps mix real change with algorithm changes. USGS LCMAP models each pixel's time series instead, with Zhu and Woodcock's Continuous Change Detection and Classification (CCDC) algorithm, which fits a harmonic model to every clear Landsat observation and flags a break when new observations depart from the prediction. The Chesapeake Conservancy reports its 2013 to 2022 change data captures 96% of land cover changes with a mapped accuracy of 77%. Change detection covers the methods in depth.

Land cover classification in Wherobots

Wherobots holds land cover in two forms. Overture land cover in the Havasu catalog holds ESA WorldCover classes as polygons, with Overture land use alongside. Overture carries land cover polygons for several map resolutions, marked by cartography.min_zoom and cartography.max_zoom, so an area summary picks one resolution first. The WorldCover rasters themselves are public Cloud Optimized GeoTIFFs, and WherobotsDB reads them in place as out-of-database rasters, fetching only the tiles a query touches.

This query counts the WorldCover 2021 pixels of each class in a box around Rockville, Maryland, the place in the header image:

-- WorldCover 2021 class counts for a box around Rockville, Maryland
WITH box AS (
  SELECT ST_GeomFromText('POLYGON((-77.21 39.04, -77.09 39.04, -77.09 39.13, -77.21 39.13, -77.21 39.04))', 4326) AS g
),
wc AS (  -- 1024-pixel tiles of the 3 by 3 degree file, read on demand
  SELECT RS_TileExplode(RS_FromPath('s3://esa-worldcover/v200/2021/map/ESA_WorldCover_10m_2021_v200_N39W078_Map.tif'), 1024, 1024) AS (x, y, rast)
),
c AS (   -- RS_Clip(raster, band, geometry, allTouched, noDataValue, crop)
  SELECT /*+ BROADCAST(box) */ RS_Clip(wc.rast, 1, box.g, false, 0, true) AS c
  FROM wc CROSS JOIN box
  WHERE RS_Intersects(wc.rast, box.g)
)
SELECT v AS worldcover_class, COUNT(*) AS pixels,
       ROUND(100 * COUNT(*) / SUM(COUNT(*)) OVER (), 1) AS pct
FROM (SELECT explode(RS_BandAsArray(c, 1)) AS v FROM c)
WHERE v > 0
GROUP BY v
ORDER BY pixels DESC;

The box holds 1,555,200 pixels, and the query returns their class counts in about 20 seconds: tree cover 64.3%, built-up 23.2%, grassland 11.2%, permanent water 0.6%, bare or sparse vegetation 0.4%, and cropland 0.2%. Tiling the file first matters here. A WorldCover file covers 3 by 3 degrees, about 36,000 by 36,000 pixels, and clipping the whole file in one call runs out of memory; RS_TileExplode splits it into 1024-pixel tiles, and only the tiles that intersect the box are clipped.

Swapping the box for parcel, tract, or watershed polygons turns the same pattern into a zonal summary, the raster-vector join behind most land cover analysis. It gives the per-polygon class shares that flood, wildfire, and canopy analyses start from.

Two WorldCover versions over Rockville

ESA produced WorldCover 2020 with algorithm v100 and WorldCover 2021 with v200, with better cropland and wetland mapping in v200. Both maps share the same 10 m grid, so a pixel-by-pixel cross-tabulation measures how much of a place changes label between versions. This query clips both tiles to the same box around Rockville and counts every pair of 2020 and 2021 classes:

-- Cross-tabulate WorldCover 2020 (v100) and 2021 (v200) pixels in one box
WITH box AS (
  SELECT ST_GeomFromText('POLYGON((-77.21 39.04, -77.09 39.04, -77.09 39.13, -77.21 39.13, -77.21 39.04))', 4326) AS g
),
t20 AS (
  SELECT RS_TileExplode(RS_FromPath('s3://esa-worldcover/v100/2020/map/ESA_WorldCover_10m_2020_v100_N39W078_Map.tif'), 1024, 1024) AS (x, y, rast)
),
t21 AS (
  SELECT RS_TileExplode(RS_FromPath('s3://esa-worldcover/v200/2021/map/ESA_WorldCover_10m_2021_v200_N39W078_Map.tif'), 1024, 1024) AS (x, y, rast)
),
pairs AS (  -- the two files share one grid, so tile (x, y) covers the same ground in both
  SELECT /*+ BROADCAST(box) */
         RS_BandAsArray(RS_Clip(a.rast, 1, box.g, false, 0, true), 1) AS v20,
         RS_BandAsArray(RS_Clip(b.rast, 1, box.g, false, 0, true), 1) AS v21
  FROM t20 a JOIN t21 b ON a.x = b.x AND a.y = b.y
  CROSS JOIN box
  WHERE RS_Intersects(a.rast, box.g)
),
cells AS (
  SELECT z.v20 AS class_2020, z.v21 AS class_2021
  FROM (SELECT explode(arrays_zip(v20, v21)) AS z FROM pairs)
  WHERE z.v20 > 0 AND z.v21 > 0
)
SELECT class_2020, class_2021, COUNT(*) AS pixels,
       ROUND(100 * COUNT(*) / SUM(COUNT(*)) OVER (), 2) AS pct
FROM cells
GROUP BY class_2020, class_2021
ORDER BY pixels DESC;

The query runs in about 20 seconds. Of the 1,555,200 pixels, 91.0% keep the same label and 9.0%, or 140,716 pixels, change. The largest moves are tree cover to built-up (2.6%), grassland to tree cover (1.6%), tree cover to grassland (1.0%), and bare or sparse vegetation to built-up (0.9%). Built-up land rises from 20.1% of the box in the 2020 map to 23.2% in 2021, and bare land falls from 1.8% to 0.4%. The 9.0% measures label disagreement between the two maps. It mixes real change with the change of algorithm, and real change can also hide under the same label in both maps, so without reference labels it is neither an upper nor a lower bound on how much of Rockville changed.

The diagonal of the result is agreement and every other cell is a label that changed. Read against an Olofsson-style sample, the off-diagonal cells split into real change and map error; read alone, they show how much of a change map can come from the classifier.

Classifying new imagery with RasterFlow

Imagery for custom classification is in the catalog too: Sentinel-2 seasonal mosaics at 10 m and USDA NAIP at 30 cm to 1 m. RasterFlow runs semantic segmentation, the task that classifies each pixel, over mosaics at scale. It builds cloud-free Sentinel-2 composites with all 12 L2A bands at 10 m, runs a land cover model through the bring-your-own-model path, and vectorizes the classes into polygons. This call builds a harvest-season mosaic for an area of interest, written to your own bucket as Zarr:

# Build a cloud-free Sentinel-2 composite as the input for a land cover model
from datetime import datetime
from rasterflow_remote import RasterflowClient, DatasetEnum

client = RasterflowClient()
result = client.build_mosaics(
    datasets=[DatasetEnum.S2_MED_HARVEST],
    aoi="s3://my-company-data/aois/project.parquet",  # one or multiple geometries
    start=datetime(2024, 1, 1),
    end=datetime(2025, 1, 1),
    bucket="s3://my-company-data/rasterflow/results",
)
print(result.first_row_mosaic)

The change detection workflow stacks two seasons or years into one inference pass, and for label-free change RasterFlow scores AlphaEarth embedding distance across years. Classified polygons load into WherobotsDB, where a spatial join summarizes land cover by parcel, watershed, or county. WherobotsDB is built by the original creators of Apache Sedona, 100% code compatible across all spatial functions.

Land cover classification research

Land cover research follows four threads. Global maps moved from 1 km to 10 m, and regional maps to 1 m. Classifiers moved from clustering and random forests to the deep segmentation models described above, trained on billions of labeled pixels. Accuracy assessment moved from the error matrix to sample-based area estimation, which shows that 10 m global maps still disagree with each other. And the analysis moved to scale, because every map ends in a raster-vector join against parcels, fields, or watersheds.

From 1 km to 1 m: five decades of land cover maps

After the Anderson and LCCS legends, global maps followed the sensors. Loveland and colleagues built IGBP DISCover (International Journal of Remote Sensing, 2000) from 1 km AVHRR data with the IGBP legend. Friedl and colleagues' MODIS Collection 5 land cover (Remote Sensing of Environment, 2010) moved to 500 m. Gong and colleagues' FROM-GLC (2013) and Chen and colleagues' GlobeLand30 (2015) reached 30 m with Landsat, and Buchhorn and colleagues' Copernicus Global Land Cover Layers (2020) added cover fractions at 100 m. The 10 m generation came with Sentinel-2: ESA WorldCover (Zanaga and colleagues), Dynamic World (Brown and colleagues, 2022), and the Impact Observatory and Esri maps (Karra and colleagues, 2021).

The 1 m generation is regional. Robinson and colleagues' Large Scale High-Resolution Land Cover Mapping with Multi-Resolution Data (CVPR 2019) trained on the Chesapeake Conservancy's 1 m labels and used 30 m NLCD as a weak signal to extend the model to the rest of the conterminous US.

Accuracy: why two good maps disagree

Congalton's error matrix and Foody's Status of land cover classification accuracy assessment (2002) set the vocabulary. Pontius and Millones' Death to Kappa (2011) argued that the kappa statistic misleads and split disagreement into quantity (wrong class proportions) and allocation (right proportions, wrong places). Stehman and Foody's Key issues in rigorous accuracy assessment of land cover products (2019), with the Olofsson good practices, made probability sampling and design-based area estimates the standard.

The 10 m global maps show why. Venter and colleagues compared Dynamic World, WorldCover, and Esri Land Cover (Remote Sensing, 2022) against global reference data and found overall accuracy of 75% for Esri, 72% for Dynamic World, and 65% for WorldCover. ESA's own validation sample gives WorldCover the higher scores listed in the examples above, so the score depends on the reference sample as much as on the map.

Resolution and the mixed pixel

Woodcock and Strahler's The factor of scale in remote sensing (1987) showed that the right pixel size depends on the information wanted, the method, and the spatial structure of the scene. Fisher's The pixel: a snare and a delusion (1997) argued that treating a pixel as one uniform patch of ground is a conceptual trap. At 10 m, a pixel on a suburban street holds part of a roof, a lawn, a tree crown, and asphalt, and the map can record only one label. NLCD's impervious-cover classes and the Chesapeake 1 m program are the two responses described above.

Finer pixels bring a different problem: occlusion. Robinson, Corley, and colleagues' Seeing the roads through the trees (2024) built the ChesapeakeRSC benchmark from 1 m NAIP imagery of Maryland, with a class for tree canopy over roads. A U-Net trained on it reached 84% recall on visible roads and 63.5% on roads under canopy, because the pixels above a covered road look like trees. RasterFlow ships the ChesapeakeRSC model, and its output for Maryland is in the Havasu catalog.

Open problems

• Map disagreement. Independent 10 m maps of the same year disagree on class areas, so users who pick a different map get a different answer.
• Area estimation. Pixel counting is biased by map error. Design-based estimators fix that, but they need a probability sample of reference labels for every study area.
• Change versus version. Comparing two maps confounds real change with algorithm change, the effect the Rockville cross-tabulation above measures.
• Labels. Deep models need many labels, and labels are expensive: the Chesapeake Conservancy spent 10 months and $1.3 million on its 1 m map of the Chesapeake Bay watershed, as Robinson and colleagues report.
• Legends. A legend built for one purpose, such as carbon accounting, splits classes differently from one built for stormwater, and translating between legends loses detail.

What scale changes

• The pixel count. WorldCover 2021 ships as 2,651 Cloud Optimized GeoTIFF tiles of 3° by 3°, each at 10 m, and Robinson and colleagues' 1 m US map holds over 8 trillion pixels.
• The join. Decisions are made per parcel, field, tract, or watershed, so every map ends in a raster-vector join. Jia Yu, Jinxuan Wu, and Mohamed Sarwat introduced GeoSpark (ACM SIGSPATIAL 2015) for distributed spatial joins, and Yu, Zongsi Zhang, and Sarwat described its partitioning and indexing in Spatial data management in Apache Spark: the GeoSpark perspective and beyond (GeoInformatica, 2019). GeoSpark became Apache Sedona, and Yu and Sarwat founded Wherobots.
• Training. Kanchan Chowdhury and Mohamed Sarwat's GeoTorch (ACM SIGSPATIAL 2022), now GeoTorchAI, adds raster transforms and benchmark datasets for training land cover models at scale.

Related terms

• Satellite imagery analysis: the computer vision tasks behind land cover maps
• Change detection: comparing imagery and maps over time
• NDVI: the vegetation index used as a classification feature
• Earth observation data: the imagery land cover maps are made from
• Raster vs vector data: land cover exists as both
• Sentinel-2: the 10 m imagery behind WorldCover and Dynamic World
• Landsat: the 30 m archive behind NLCD

Read more from Wherobots

• Raster processing at scale: the out-of-database architecture behind WherobotsDB
• AlphaEarth Foundations for agriculture change detection
• Zonal statistics on AlphaEarth embeddings
• Fields of the World: a GeoAI pipeline on RasterFlow
• Build a Sentinel-2 mosaic with RasterFlow

Query land cover rasters and polygons with spatial SQL with a Wherobots free trial at cloud.wherobots.com.