Advanced techniques for classification of polarimetric synthetic aperture radar data

With various remote sensing technologies to aid Earth Observation, radar-based imaging is one of them gaining major interests due to advances in its imaging techniques in form of syn-thetic aperture radar (SAR) and polarimetry. The majority of radar applications focus on mon-itoring, detecting, and classifying local or global areas of interests to support humans within their efforts of decision-making, analysis, and interpretation of Earth’s environment.\nThis thesis focuses on improving the classification performance and process particularly concerning the application of land use and land cover over polarimetric SAR (PolSAR) data. To achieve this, three contributions are studied related to superior feature description and ad-vanced machine-learning techniques including classifiers, principles, and data exploitation.\nFirst, this thesis investigates the application of color features within PolSAR image classi-fication to provide additional discrimination on top of the conventional scattering information and texture features. The color features are extracted over the visual presentation of fully and partially polarimetric SAR data by generation of pseudo color images. Within the experiments, the obtained results demonstrated that with the addition of the considered color features, the achieved classification performances outperformed results with common PolSAR features alone as well as achieved higher classification accuracies compared to the traditional combination of PolSAR and texture features.\nSecond, to address the large-scale learning challenge in PolSAR image classification with the utmost efficiency, this thesis introduces the application of an adaptive and data-driven supervised classification topology called Collective Network of Binary Classifiers, CNBC. This topology incorporates active learning to support human users with the analysis and interpretation of PolSAR data focusing on collections of images, where changes or updates to the existing classifier might be required frequently due to surface, terrain, and object changes as well as certain variations in capturing time and position. Evaluations demonstrated the capabilities of CNBC over an extensive set of experimental results regarding the adaptation and data-driven classification of single as well as collections of PolSAR images. The experimental results verified that the evolutionary classification topology, CNBC, did provide an efficient solution for the problems of scalability and dynamic adaptability allowing both feature space dimensions and the number of terrain classes in PolSAR image collections to vary dynamically.\nThird, most PolSAR classification problems are undertaken by supervised machine learn-ing, which require manually labeled ground truth data available. To reduce the manual labeling efforts, supervised and unsupervised learning approaches are combined into semi-supervised learning to utilize the huge amount of unlabeled data. The application of semi-supervised learning in this thesis is motivated by ill-posed classification tasks related to the small training size problem. Therefore, this thesis investigates how much ground truth is actually necessary for certain classification problems to achieve satisfactory results in a supervised and semi-supervised learning scenario. To address this, two semi-supervised approaches are proposed by unsupervised extension of the training data and ensemble-based self-training. The evaluations showed that significant speed-ups and improvements in classification performance are achieved. In particular, for a remote sensing application such as PolSAR image classification, it is advantageous to exploit the location-based information from the labeled training data. \nEach of the developed techniques provides its stand-alone contribution from different viewpoints to improve land use and land cover classification. The introduction of a new fea-ture for better discrimination is independent of the underlying classification algorithms used. The application of the CNBC topology is applicable to various classification problems no matter how the underlying data have been acquired, for example in case of remote sensing data. Moreover, the semi-supervised learning approach tackles the challenge of utilizing the unlabeled data. By combining these techniques for superior feature description and advanced machine-learning techniques exploiting classifier topologies and data, further contributions to polarimetric SAR image classification are made. According to the performance evaluations conducted including visual and numerical assessments, the proposed and investigated tech-niques showed valuable improvements and are able to aid the analysis and interpretation of PolSAR image data. Due to the generic nature of the developed techniques, their applications to other remote sensing data will require only minor adjustments.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.