On Classification with Unreliable Labels for Environmental and Medical Applications

Information technology is undergoing yet another revolution, dubbed the data revolution. The recent advancements in sensing technology, storage systems, computational systems, and mathematical tools are enabling the realization of systems that can observe the world at a low cost, store large amounts of data, and run complex algorithms efficiently to process these large datasets. A key component in the design of such systems is the validation process where we need to evaluate the system on datasets representative of real life. In this dissertation, we consider environmental and medical classification problems where the validation process is challenging due to the difficulty of collecting class labels and ground truth.;We divide our work into three parts. In the first part, we present a system for tree type classification using satellite or aerial images. The system is used to update the current forest maps of the National French Forest Inventory (IFN).;In the second part, we present three motion recognition systems using wearable accelerometers designed for healthcare and medical applications. The first system is designed to monitor the workplace activities and study the seated posture habits of the user. The second one is designed to recognize the activity of the user from a set of 14 common daily activities. The third system is designed for stumble detection in analyzing the gait of the user, and studying the effect of frequent stumbles on the risk of falling. We also present two large datasets collected for the validation of the systems.;In the third part, we present a novel algorithm to collect data that optimizes the model selection in the maximum likelihood framework, for linear regression models used in spatial process estimation.