Sentinel-2 provides free multispectral imagery at 10 m resolution, making it one of the most accessible sources for land-use and land-cover (LULC) mapping. Operated by the European Space Agency under the Copernicus programme, it revisits any point on Earth every five days, giving you a rich time series to work with. This article walks through a supervised classification workflow using QGIS and the Semi-Automatic Classification Plugin (SCP).
A well-executed LULC map is the starting point for environmental assessments, agricultural planning, urban growth analysis and disaster-risk studies.
Download imagery
Use the Copernicus Open Access Hub or the SCP plugin's built-in download tool to obtain a cloud-free scene for your area of interest. Select a scene with less than 10 % cloud cover and choose the Level-1C product if you intend to apply your own atmospheric correction, or Level-2A if you want surface reflectance straight from the archive.
Pre-processing
If you downloaded a Level-1C product, apply atmospheric correction using the SCP DOS1 algorithm to convert top-of-atmosphere digital numbers to approximate surface reflectance. Stack the bands you need — typically B2, B3, B4 and B8 for a standard RGB-NIR composite — into a single multi-band raster using the SCP Band Set tool.
Useful band combinations
A false-colour composite using NIR, Red and Green (B8-B4-B3) makes vegetation appear bright red and is useful for distinguishing vegetated from non-vegetated surfaces. A shortwave infrared composite (B12-B8A-B4) helps separate bare soil, built-up areas and moisture-stressed vegetation.
Training samples
Digitise representative polygons for each land-cover class — forest, cropland, water, built-up and bare soil — using field knowledge or high-resolution basemaps as a reference. Aim for at least 30 polygons per class, distributed across the study area to capture spectral variability within each class.
Classification
Run a Random Forest or Maximum Likelihood classification from the SCP Classification tool. After the classification completes, review the accuracy assessment report. Aim for an overall accuracy above 85 % and a Kappa coefficient above 0.80 before proceeding. If accuracy is low, add more training samples for the poorly classified classes and re-run.
Post-processing
Apply a majority filter with a 3×3 kernel to remove isolated pixels that are likely misclassifications. Calculate class areas using the SCP Land Cover Change tool or QGIS's raster layer statistics. Export the final classified raster as a GeoTIFF and produce a summary statistics table for your report.
Have a mapping challenge?
Let's turn your spatial data into an accurate, useful and visually compelling map.
