Iterative K - Nearest Neighbors Algorithm (IKNN) for submeter spatial resolution image classification obtained by Unmanned Aerial Vehicle (UAV)

This study proposes a classification technique named Iterative K - Nearest Neighbors algorithm (IKNN) for submeter spatial resolution images acquired by Unmanned Aerial Vehicles (UAV). The method is based on the development of simple solutions for some limitations found in the traditional K - Nearest Neighbors algorithm (KNN). The main changes with respect to the traditional one are: (i) handle the high dimensionality of the data and the overlapping of the features by computing Gini Importances (GI); and (ii) selecting the number of KNN through an iterative algorithm according each classification rate at each iteration. Considering the GI indices as features weights, the IKNN method achieved a reasonable reduction in dimensionality of the data and overlapping among features. Experiments using the proposed method with confidence threshold equal to 60% resulted in a proportion correct (PC) of 90%, which was superior comparing to Support Vector Machine (SVM) and simple KNN methods.