How do you translate satellite imagery into reliable land cover and land use maps? In this Blended Learning course on Image Classification and Land Cover, you’ll learn how to use image classification to automatically recognize and distinguish vegetation, water, soil, and built-up areas, among other features. You’ll learn to evaluate and validate classification results and apply them to spatial issues.
In this Blended Learning course, you’ll learn how to convert satellite images into usable land-cover information. You’ll learn to apply various classification methods, evaluate results, and translate classifications into reliable maps that can be used for nature conservation, agriculture, water management, urban development, and monitoring changes in the landscape, among other applications.
A satellite image consists of pixels with measured values in different spectral bands. Different types of land cover each have their own spectral properties. For example, vegetation reflects radiation differently than water, bare soil, or buildings. These differences form the basis for the automatic classification of satellite images.
You’ll learn how pixels can be categorized into different classes based on their spectral properties. You’ll also be introduced to concepts such as features, classes, training data, decision rules, and classification models. You’ll also learn the difference between land cover and land use and why this distinction is important when interpreting classification results.
In image classification, you determine which category a pixel or group of pixels belongs to. You’ll learn how different spectral bands, band combinations, and derived information can be used to better distinguish between classes.
You’ll be introduced to supervised and unsupervised classification. In supervised classification, you use known examples to train a model to recognize which characteristics belong to a particular class. In unsupervised classification, pixels are automatically grouped based on similarities, after which you investigate what these groups represent in the landscape.
Above all, you’ll learn to assess which approach is best suited to the problem at hand, the available satellite data, and the desired land-cover map.
The quality of a classification is strongly determined by the quality of the data used. That is why you’ll learn how to select representative training areas and how to prevent classes from being too similar spectrally or from being underrepresented.
You’ll be introduced to various classification methods and learn to compare their results. In doing so, you’ll explore how the choice of spectral bands, training data, and classification method influences the final result.
For the analyses, you’ll work with open satellite data and open-source software. QGIS serves as the central working environment, and where relevant, specialized open-source algorithms—such as those from the Orfeo ToolBox—can be used. However, the software itself is not the starting point; the central focus is always on how to achieve a reliable classification.
A land-cover map is only useful if you know how reliable the classification is. That is why you will learn to critically evaluate classification results and compare them with independent reference data.
You’ll be introduced to validation techniques such as confusion matrices and accuracy metrics for individual classes and for the classification as a whole. You’ll investigate where errors occur, which classes are often confused with one another, and how training data or the chosen method can be improved.
In this way, you’ll not only learn how to perform a classification but also how to substantiate the reliability of the results and determine for which applications the map is—or isn’t—suitable.
During the Blended Learning course, you’ll work with realistic satellite images and open datasets. You’ll use image classification to reveal spatial patterns and derive land-cover information that can be used for analysis and decision-making.
You’ll work on assignments such as:
Upon completion, you will be able to independently classify satellite images and convert them into reliable land-cover information. You will be able to select an appropriate classification strategy, evaluate training data, validate results, and use land-cover maps to analyze spatial patterns and trends.
You’ll receive 1-on-1 guidance. After signing up, our course coordinator will contact you to schedule your first session.
Do you have questions about the course content? Or are you unsure whether the course aligns with your learning goals or preferences? Would you prefer an in-house or private course? We’d be happy to help.
Various freely available satellite datasets are available for image classification. Within Europe, the Copernicus Data Space Ecosystem is a key source of Sentinel data. Sentinel-2, in particular, is well-suited for classifying vegetation, water, soil, and built-up areas thanks to its various spectral bands and the regular availability of new images. In addition, Landsat data can be used for analyses over longer periods. In the Blended Learning module, you’ll learn which data sets are suitable for different classification problems.
In image classification, pixels or groups of pixels are categorized into different classes based on their characteristics. In supervised classification, you provide examples of known classes, such as water, vegetation, soil, and buildings. This training data is used to automatically classify other pixels. In unsupervised classification, pixels are first automatically grouped based on similarities, after which you determine which land-cover classes these groups represent.
Land-cover classes are primarily identified using measurements from different spectral bands of a satellite image. Vegetation, water, soil, and built-up areas each respond differently to visible light, near-infrared, and other parts of the electromagnetic spectrum. By combining multiple bands and, where applicable, derived spectral indices, classes can be more clearly distinguished from one another, and the reliability of a classification can be improved.
The main result is a land-cover map in which each pixel or spatial unit is assigned to a class. This produces, for example, a map showing vegetation, water, soil, and built-up areas. From this map, it is then possible to derive areas, spatial patterns, and differences between regions. Classification results can also be combined with other geodata or used as a basis for further spatial analyses.
A classification is verified using independent reference data. By comparing the classified land-cover classes with the actual situation, you can determine where the classification is accurate and where errors occur. A confusion matrix and various accuracy metrics reveal which classes are reliably identified and which are frequently confused with one another. This allows you to assess whether the land-cover map is sufficiently reliable for the intended spatial analysis.
