Before classifying an image, you must decide what you are mapping: the physical surface (land cover) or how humans use it (land use). This distinction shapes your training data, class labels, and how you interpret results.
Learning objectives
- Distinguish between land cover and land use.
- Recognize why the same spectral signature can represent different classes.
- Choose appropriate class labels for your classification project.
- Understand limitations of spectral-only classification for land use mapping.
Why it matters
Confusing land use with land cover leads to misclassification. A grassy area looks the same to a satellite whether it is a park, a golf course, or a cattle pasture. Understanding this distinction helps you set realistic expectations for your classifier.
Key definitions
| Term | Definition | Examples |
|---|---|---|
| Land Cover | Physical material on the Earth's surface | Forest, water, bare soil, impervious surface, grass |
| Land Use | Human activity or function of the land | Residential, commercial, agriculture, recreation, conservation |
The same cover, different uses
Consider these examples where identical spectral signatures represent different land uses:
| Land Cover | Possible Land Uses |
|---|---|
| Grass/Herbaceous | Park, golf course, pasture, lawn, sports field, airport |
| Forest | Conservation area, timber harvest, recreation, watershed protection |
| Water | Reservoir, aquaculture, natural lake, wastewater treatment |
| Impervious | Residential, commercial, industrial, transportation |
Key insight
Satellites see land cover, not land use. A classifier trained on spectral bands can reliably distinguish forest from water, but it cannot tell a city park from a cattle ranch based on reflectance alone. Land use mapping often requires ancillary data: parcel boundaries, building footprints, road networks, or temporal patterns.
LULC: The hybrid approach
Many classification schemes combine both concepts into "Land Use/Land Cover" (LULC) classes. Common examples include:
- NLCD (National Land Cover Database): Primarily land cover with some use-based classes like "Developed" at different intensities.
- CORINE (Europe): Hierarchical system mixing cover and use.
- FAO LCCS: Classification system designed to separate cover from use when needed.
Try it: Examine a LULC dataset
Load NLCD in Earth Engine and explore the class legend:
// Load NLCD 2021
var nlcd = ee.Image('USGS/NLCD_RELEASES/2021_REL/NLCD/2021');
var landcover = nlcd.select('landcover');
// Add to map with built-in palette
Map.setCenter(-98.5, 39.5, 5);
Map.addLayer(landcover, {}, 'NLCD 2021');
// Print the class names
print('NLCD classes:', landcover.get('landcover_class_names'));
Notice how "Developed" classes mix land cover (impervious percentage) with land use (urban).
Common mistakes
- Training "agriculture" samples on bare fields some months and crops other months.
- Expecting a classifier to distinguish "residential" from "commercial" using only spectral data.
- Mixing classification schemes (for example, calling grass "pasture" in one sample and "lawn" in another).
- Ignoring that the same pixel may be classified differently at different times of year.
Quick self-check
- Can a satellite distinguish a forest used for conservation from one used for logging?
- Why might "Developed, High Intensity" be considered a hybrid LULC class?
- What additional data would help classify grass as either "park" or "pasture"?
Next steps
- Introduction to Image Classification - classification fundamentals
- Supervised vs Unsupervised - choosing an approach
- Discrete vs Continuous Data - understanding output types