Extracting Supervised Learning Classifiers from Possibly Incomplete Remotely Sensed Data

Mapping and classification of human settlements from remotely sensed data has attracted a lot of attention in recent years. Real world data, however, often suffer from corruptions or noise but not always known. This is the heart of information-based remote sensing models. This paper investigates the impact of incomplete remotely sensed data in the evaluation of machine learning techniques (classifiers) for the task of predicting or classifying pixels into different landcover region types. Six classifiers are empirically evaluated by artificially simulating different missing data proportions, patterns and mechanisms using a multispectral image dataset. A 4-way repeated measures design is employed to analyse the data. The simulation results suggest classifiers as having their strengths and limitations in terms of dealing with the incomplete data problem with the artificial neural network classifier as substantially inferior and naive Bayes classifier and support vector machines representing superior approaches.