Ensemble of classifiers for remote sensed hyperspectral land cover analysis: An approach based on Linear Programming and Weighted Linear Combination

Hyperspectral images have been considered as one of the most important tool for remote sensed land cover analysis. Such images have information about materials on earth´s surface expressed in many wavelengths that allow us to identify and classify those materials with more accuracy. In this work we used a combination of several classification methods in order to produce an accurate thematic map based on the remote sensed hyperspectral image classification. To perform the combination, three types of feature representation and two learning algorithms (Support Vector Machines (SVM) and Backpropagation Multilayer Perceptron Neural Network (MLP)) were used yielding six classification methods. Our approach proposal is based onWeighted Linear Combination (WLC), in which weights are found using Linear Programming (LP) - WLC-LP. Experiments are carried out using two well-known databases: Indian Pines, acquired by AVIRIS sensor; and Pavia University, acquired by ROSIS sensor. Results show the efficiency of our proposed approach which significantly reduces the time required to found optimal weights for the combiner compared to a previous approach based on Genetic Algorithm.