Object-Based Supervised Machine Learning Regional-Scale Land-Cover Classification Using High Resolution Remotely Sensed Data

High spatial resolution (HR) (1m – 5m) remotely sensed data in conjunction with supervised machine learning classification are commonly used to construct land-cover classifications. Despite the increasing availability of HR data, most studies investigating HR remotely sensed data and associated classification methods employ relatively small study areas. This work therefore drew on a 2,609 km2, regional-scale study in northeastern West Virginia, USA, to investigates a number of core aspects of HR land-cover supervised classification using machine learning. Issues explored include training sample selection, cross-validation parameter tuning, the choice of machine learning algorithm, training sample set size, and feature selection. A geographic object-based image analysis (GEOBIA) approach was used. The data comprised National Agricultural Imagery Program (NAIP) orthoimagery and LIDAR-derived rasters. Stratified-statistical-based training sampling methods were found to generate higher classification accuracies than deliberative-based sampling. Subset-based sampling, in which training data is collected from a small geographic subset area within the study site, did not notably decrease the classification accuracy. For the five machine learning algorithms investigated, support vector machines (SVM), random forests (RF), k-nearest neighbors (k-NN), single-layer perceptron neural networks (NEU), and learning vector quantization (LVQ), increasing the size of the training set typically improved the overall accuracy of the classification. However, RF was consistently more accurate than the other four machine learning algorithms, even when trained from a relatively small training sample set. Recursive feature elimination (RFE), which can be used to reduce the dimensionality of a training set, was found to increase the overall accuracy of both SVM and NEU classification, however the improvement in overall accuracy diminished as sample size increased. RFE resulted in only a small improvement the overall accuracy of RF classification, indicating that RF is generally insensitive to the Hughes Phenomenon. Nevertheless, as feature selection is an optional step in the classification process, and can be discarded if it has a negative effect on classification accuracy, it should be investigated as part of best practice for supervised machine land-cover classification using remotely sensed data