Remote Sensing Classification
Remote Sensing Image Classification · Also known as: land cover classification, image classification, satellite image classification, spectral classification
Remote sensing classification assigns discrete thematic labels — such as forest, urban, water, or cropland — to pixels in a satellite or aerial image based on their spectral, spatial, and temporal properties. It underpins land-use/land-cover mapping, change detection, environmental monitoring, and disaster response at local to global scales.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use remote sensing classification when you need to map land cover, land use, vegetation types, urban extent, or other surface phenomena over large areas that cannot be surveyed on foot. It is ideal when high-resolution multispectral or hyperspectral imagery is available and sufficiently labeled training samples can be collected. Avoid it when cloud cover persistently obscures the study area (use SAR-based methods instead), when required thematic classes are spectrally inseparable, when the spatial resolution of available imagery is too coarse for the minimum mapping unit, or when ground truth data are too scarce to train and validate a reliable classifier.
Strengths & limitations
- Enables synoptic land cover mapping over vast areas at relatively low cost compared with field surveys.
- Temporal repeatability of satellite sensors supports change detection and time-series analysis.
- Wide choice of algorithms — from classical maximum likelihood to deep learning — allows tuning to data complexity and sample size.
- Produces spatially explicit, GIS-compatible raster or vector outputs ready for further spatial analysis.
- Applicable to multiple sensor types: optical, SAR, LiDAR, thermal, and hyperspectral.
- Supports multi-date fusion and object-based image analysis (OBIA) for more detailed feature extraction.
- Classification accuracy is strongly dependent on the quality and representativeness of training samples; biased samples produce biased maps.
- Spectral confusion among land cover types with similar reflectance (e.g., bare soil vs. dry vegetation) can be difficult to resolve without additional bands or ancillary data.
- Atmospheric correction errors, sensor artifacts, and mixed pixels at class boundaries degrade map quality.
- Deep learning classifiers require large labeled datasets and significant computational resources.
- Accuracy assessment results can be misleading if validation samples are not collected independently of training samples.
Frequently asked
What is the difference between supervised and unsupervised classification?
Supervised classification uses analyst-labeled training samples to teach the algorithm what each land cover class looks like spectrally; the algorithm then assigns every pixel to one of the predefined classes. Unsupervised classification groups pixels by statistical similarity (clustering) without prior labels; the analyst interprets and names the resulting clusters after the fact.
How many training samples do I need?
A commonly cited rule of thumb is at least 10 times as many training pixels per class as there are spectral bands, with 50–100 pure pixels per class as a practical minimum. Machine learning classifiers (random forests, SVMs) and deep learning models generally need substantially more samples, especially for spectrally complex classes.
What is overall accuracy and kappa, and which should I report?
Overall accuracy is the proportion of validation pixels correctly classified. Cohen's kappa adjusts for chance agreement. Best practice is to report overall accuracy together with per-class producer's and user's accuracies from the full confusion matrix, as these reveal which classes are confused with one another.
Can I classify imagery without field-collected training data?
Yes, using unsupervised methods (k-means, ISODATA) or by deriving training data from existing land cover maps, OpenStreetMap, or volunteer geographic information. However, classification accuracy is typically lower than with high-quality field-verified training samples.
What is object-based image analysis (OBIA) and when is it preferable?
OBIA first segments the image into spectrally and spatially homogeneous objects (polygons) and then classifies objects using shape, texture, and contextual features in addition to spectral values. It is preferable over pixel-based methods when image resolution is very high (sub-meter) and spatial context — building shape, parcel size — is needed to distinguish classes.
Sources
- Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2015). Remote Sensing and Image Interpretation (7th ed.). Wiley. ISBN: 978-1118343289
- Remote sensing. Wikipedia. link ↗
How to cite this page
ScholarGate. (2026, June 3). Remote Sensing Image Classification. ScholarGate. https://scholargate.app/en/spatial-analysis/remote-sensing-classification
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Hot Spot AnalysisSpatial analysis↔ compare
- Local Indicators of Spatial AssociationSpatial analysis↔ compare
- Multiscale Geographically Weighted RegressionSpatial analysis↔ compare
- Network-Based Spatial AnalysisSpatial analysis↔ compare