The K-Closest Resemblance Classifier for Remote Sensing Data

Abstract

The supervised learning algorithms provide a powerful tool to classify and process the remotely sensed imagery data sets. They have the strengths to handle high-dimensional data and to map classes with very complex characteristics. However, the usual supervised machine learning algorithms face issues that limit their applicability especially in dealing with the knowledge interpretation and with imbalanced labeled data sets. To address these issues, the prototype based classifier K-Closest Resemblance K-CR was proposed. K-CR is inspired by the social choice theory and preference modeling, which argues that the classifiers based on preference modeling are simple, do not need to normalize the features, and do not have loss of information during learning. The effectiveness of the proposed classifier is evaluated by comparing with the other well-known existing classifiers for remote sensing data set. The obtained results indicate that our proposed classifier is an effective tool for land cover classification from remote sensing data.