Multimodal classification of remote sensing images

Abstract

Geographic Information Systems (GIS) are computer-based tools that analyzes, stores, manipulates and visualizes geographic information on maps. Enabling people to more easily see, interpret, and understand patterns and relationships, they are important tools, which helps to make decisions on diverse issues related to public policies of urban and environmental monitoring. Furthermore, they are essential tools for economic development planning and analysis in activities of agriculture and mineral extraction. In this context, Remote Sensing Images (RSIs) have been used as a major source of data, particularly with respect to the creation of thematic maps. This process is usually modeled as a supervised classification problem where the system needs to learn the patterns of interest provided by the user and assign a class to the rest of the image regions. Associated with the nature of RSIs, there are several challenges that can be highlighted: (1) they are georeferenced images, i.e., a geographic coordinate is associated for each pixel; (2) the data commonly captures specific frequencies across the electromagnetic spectrum instead of the visible spectrum, which requires the development of specific algorithms to describe patterns; (3) the detail level of each data may vary, resulting in images with different spatial and pixel resolution, but covering the same area; (4) due to the high pixel resolution images, efficient processing algorithms are desirables. Thus, it is very common to have images obtained from different sensors, which could improve the quality of thematic maps generated. However, this requires the creation of techniques to properly encode and combine the different properties of the images. Thus, this dissertation proposes two new techniques for classification of regions in remote sensing images that manage to encode features extracted from different sources of data, spectral and spatial domains. The major objective is the development of a framework able to exploit the diversity of these different types of features to achieve high degrees of accuracy in the creation of thematic maps. The first approach proposed, as named Dynamic Majority Vote, combine the data from different domains using a strategy of fusion at the decision level, which exploits the specialty of different classifiers and combine them for a final decision for each pixel in the thematic map. This approach was created to handle only with multiclass problems and was evaluated in a urban land cover scenario dataset. The second approach proposed, is a combination of the pixel, feature and decision levels fusion, which exploit different types of features, extracted from various sensors using an boosting-based approach. Differently of the Dynamic Majority Vote, this method can handle with binary and multiclass problems, and was evaluated in the same urban land cover multiclass scenario dataset, and also in a coffee crop recognition binary problem.