Comparison of Selected Machine Learning Algorithms for Sub-Pixel Imperviousness Change Assessment

The paper presents the comparison of nine machine learning algorithms for sub-pixel impervious surface area change assessment. Predictive models were tuned and trained using the caret package in R environment. Their performance was analyzed based on both cross-validation results and results obtained for validation dataset. A paired t-test was used to determine if the differences between model accuracies are statistically significant. In case of imperviousness mapping for individual time points the regression trees based models outperformed other ones both for cross-validation on calibration dataset and for validation dataset. The Cubist algorithm seems to be the best performed one. The best assessment method for ISA change cannot be unambiguously pointed out. Random Forest gave the lowest RMS errors, random kNN was the best one according to MAE measure and support vector machines with radial basis kernel gave the highest mean value of the R2.