Comparing supervised algorithms in Land Use and Land Cover classification of a Landsat time-series

Machine learning algorithms (MLAs) are often applied to identify Land Use and Land Cover (LULC) changes, but typically to only a limited set of imagery. This leaves the consistency of MLAs performance through time poorly understood. The research objective was therefore to compare the performance of six MLAs across a time-series of Landsat imagery (1979, 1992, 2003, 2014), all processed in the same manner. Here Support Vector Machines (SVM), K-Nearest Neighbours (KNN), Random Forests (RF), Learning Vector Quantization (LVQ), Recursive Partitioning, Regression Trees (RPART) and Stochastic Gradient Boosting (GBM) were evaluated. The results demonstrated that SVM achieved higher overall accuracies and kappa coefficients, and a slightly improved fit at individual class level, than the second best classifier RF. Both classifiers clearly outperformed the other algorithms. These results suggest that SVMs (or RFs) should be prioritised when classifying time-series imagery for LULC change detection.