Automatic classification of high resolution land cover using a new data weighting procedure: The combination of k-means clustering algorithm and central tendency measures (KMC-CTM)

Information on a well-scale urban land cover is important for a number of urban planning practices involving tree shade mapping, green space analysis, urban hydrologic modeling and urban land use mapping. In this study, an urban land cover dataset received from the database of UCI (University of California at Irvine) machine learning was used as the urban land cover data. This dataset is the urban area located in Deerfield Beach, FL, USA. Separately, this dataset is a high definition atmospheric image consisting of 9 different urban land covers. The characteristics of a multi-scale spectral, magnitude and formal tectology were used to sort out and classify these different images. The dataset comprises a total of 147 features and land covers of 9 different areas involving trees, grass, soil, concrete, asphalt, buildings, cars, pools and shadows. A new data weighting method was recommended to classify these 9 different patterns automatically. This recommended data weighting method is based on the combination of the measures of central tendency composed of mean value, harmonic value, mode and median along with the k-means clustering method. In the data weighting method, the data sets belonging to each class within the dataset are first calculated by using k-means clustering method, after which the measures of central tendency belonging to each class are calculated, as well. The measure of central tendency belonging to each class is divided by the set central value belonging to the class in question, as the result of which the data weight coefficient of that class has already been calculated. This calculation process is performed separately for 9 different land covers, and afterwards, these data weighting coefficients found are multiplied by the dataset, and thus, the dataset has been weighted. In the second stage, on the other hand, 3 different classification algorithms containing k-NN (k-nearest neighbor), extreme learning machine (ELM) and support vector machine (SVM) were used to classify 9 different urban land covers after the data weighting method. In determining the educational and test data sets, the 10-fold cross validation was used. When classification through raw data was performed along with k-NN (for k = 1), ELM and SVM classification algorithms, the overall classification accuracy obtained was 77.15%, 84.70% and 84.79%, respectively. When classification through data weighting method (the combination of k-means clustering and mode measure) along with k-NN (for k = 1), ELM and SVM classification algorithms was made, the overall classification accuracy obtained proved to be 98.58%, 98.62% and 98.77%, respectively. The obtained results suggest that the urban land cover in an atmospheric image via the recommended data weighting method was classified as 9 different areas with a high classification success rate. © 2015 Elsevier B.V. All rights reserved.