Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Sparse representations in machine learning and remote sensing

Karalas Konstantinos

Full record

---

Summary

Land cover maps are critical for environmental monitoring and urban development among others. Unfortunately, in order to produce such maps significant labor intensive effort is required by human annotators through field-studies. Interestingly, on the other hand, high resolution imaging systems onboard airborne and spaceborne platforms are able to capture rich information in parts of the electromagnetic spectrum that the human eye cannot discern. This remote sensing imagery can be used to overcome the issues associated with field-studies, providing global and up-to-date land cover maps. Typically, during the mapping procedure, each remotely sensed pixel is classified into a single class, leading to very coarse representations. In the past few years, the development of the powerful framework of multi-label learning, where instances may be associated with multiple labels simultaneously, has been successfully applied in various computer vision scenarios. Part of the success is also attributed to the development of hand-crafted features which can dramatically boost the performance under specific conditions, however, these features are very specialized and lack universality.This thesis introduces a radically novel approach for inferring the complex relationships between multispectral satellite imagery and spectral profiles of different surface materials, exploiting the proliferation of remote sensing imagery, through the introduction of the multi-label classification framework. The adoption of this scheme provides a real-world answer to the scale incompatibility problem between remote sensing imagery and ground-based measurements, since they naturally come in different spatial resolutions. Furthermore, instead of relying on specialized features, we propose the application of deep feature learning with stacked sparse autoencoders, in order to automatically extract meaningful features identifying the underlying explanatory patterns hidden in low level satellite data.To validate the merits of the proposed approach, we consider real contemporary data from the European Environment Agency for generating the ground-truth, and multispectral images from the Moderate-resolution Imaging Spectroradiometer sensor for feature extraction. We present results using several state-of-the-art multi-label learning classifiers and evaluate their predictive performance under different challenging scenarios, including cases where training is localized in a specific area and time, while testing takes place on a different location or time instance. Experimental results suggest that the proposed framework can achieve excellent prediction accuracy, even from a limited number of diverse training examples, whereas the application of feature learning leads to more representative features that can significantly boost the performance of multi-label classification problems.

Available Files

Services