Land cover classification using Machine Learning Techniques
The project intends to perform and compare different supervised classification methods, using a SENTINEL-2A satellite image to settle on some separable land cover classes within Lunds Kommun, and create a land cover map of the area, by exploring different objects’ reflectance properties, and compare their accuracy.
Four ML techniques were tested to create a Land cover classification of Lund Municipality (Lunds Kommun):
Artificial neural network-based model Multilayer Perceptron (MPL);
Decision tree-based model Random Forest (RF);
Support Vector machine-based model (SVM) with the radial basis function (RBF) kernel;
Minimum Distance (MD).
The first three ones were done using R, and the last one was made in IDRISI TerrSet.
Fieldwork was carried out in order to collect training and evaluation data - 299 GPS points were collected and identified among 6 land cover classes in the field, 80% was used for training and 20% for classification validation.
Based on the visual interpretation of the true and false image composites, the spectral signature of the objects, and the fieldwork, the classes defined were:
Urban Areas
Agriculture Lands
Water
Coniferous Forest
Deciduous Forest
Bare Soil
Among the methods, the MLP showed the best accuracy and Kappa results.