Minimal Sampling Classifier.

In the context of nature conservation, mapping of habitats and classification of landscape types is an important action. The visual mapping of natural or near-natural landscapes requires a high level of expert knowledge and is very time-consuming and demanding. AI-based methods are increasingly being used for automatic classification of aerial and satellite images. However, the results are often not comprehensible or implausible for the expert.

MiSa.C is designed to interactively incorporate users' domain expert knowledge in an machine learning classification process and thus allows the user to benefit from artificial intelligence and still having full control over the mapping results. The only input data required is the image dataset to be classified and one reference point per land cover class. The tool uses image statistics, as well as machine learning, to automatically create an extended set of unbiased and comprehensive training data and model results.

Since at its core it uses machine learning, MiSa.C aims at the use of multiple image sources, provided in a 4D data cube (2D space, time, and an observation value), e.g. a combination of optical imagery, radar imagery and a digital elevation model or other, for more effective classifications and a broader set of use-case scenarios.

MiSa.C is a three-step classification process via an web-app running in the cloud. Its major advantage is the ability to infuse expert knowledge into the classification process. Either via parameter tuning or through a threshold to control the delineation of the land cover class, MiSa.C assists users with different levels of expertise during the classification process.

If you want to test MiSa.C, please request a free account at fernlab@gfz-potsdam.de.

Find more information about MiSa.C in our git documentation at

FERN.Lab process chain

Open data
Commercial data
Method development
Data homogenization
Time Series Analysis

GUI development
API development