Combinação de múltiplas abordagens de classificação para interpretação de imagens hiperespectrais de sensoriamento remoto

Abstract

In the past few decades hyperspectral data analysis has become the main source for classification of remote sensed images. A hyperspectral image is composed of hundreds of spectral channels, where each channel refers to a specific wavelength. Such a large amount of information may lead us to a deeper investigation of the materials on Earth's surface, and thus, a more precise interpretation of them. Land cover classification is still a challenge task, and producing accurate thematic map is a common goal among researchers. Although each pixel in a hyperspectral image has a detailed spectral information, the joint of both spectral and spatial information is required due to its relations among adjacent pixels. However, joining such information is a widely studied and opened topic in the remote sensing community. In this work we applied meta-learning techniques to combine multiple classification approaches aiming to procuce a more accurate thematic maps. For this purpose, four classification approaches were used in the combination, in which two of them is also a part of this work. In an attempt to maximize the information gain in the combination, each one of the classification approaches has its own feature representation and also uses learning algorithms from different nature. We used two feature representation based on only spectral info and another two based on both spectral-spatial info. Three different learning algorithms were applied. The Support Vector Machines (SVM) and K-Nearest Neighbor (KNN) classifiers were used to classify spectral based data, while Multilayer Perceptron Neural Network (MLP) and, once again, SVM on spectral-spatial based data. Two methods to combine the four classification approaches were proposed. Our firts proposal is based on on Weighted Linear Combination (WLC), in which weights are found by using a Genetic Algorithms (GA) - WLC-GA. The second one, Stacking-SVM, performs a nonlinear combination by means of Stacked Generalization strategy. In order to do it, the SVM learning algorithm with RBF kernel was used. Experiments were carried out with two well-known datasets: Indian Pines and Pavia University. In order to evaluate the robustness of the proposed combiner, experiments using different training sizes on different scenarios were conducted. It was observed that the proposed WLC-GA method achieved promising results on both datasets used, owning the highest accuracy levels among Stacking-SVM and other traditional methods, such as Majority Vote (MV) and Average rules. With only 5% of training samples, the WLC-GA method was able to find optimal, or suboptimal, set of weights which allows an accurate combination of the classification approaches, and hence, a more precise thematic maps, overcoming some drawbacks of other combination methods