Extracting Information from Multimodal Remote Sensing Data for Sea Ice Characterization

• Main Author: Nilsen, Torjus
• Format: Master Thesis
• Language: English
• Published: UiT Norges arktiske universitet 2021
• Subjects: Machine Learning; VDP::Mathematics and natural science: 400::Mathematics: 410::Analysis: 411; Semi-Supervised; Graph based; Multimodal; Remote Sensing; GLCM; MBO; Label propagation; MixMatch; FYS-3941; Arctic; Sea ice
• Online Access: https://hdl.handle.net/10037/21912

Remote sensing is the discipline that studies acquisition, preparation and analysis of spectral, spatial and temporal properties of objects without direct touch or contact. It is a field of great importance to understanding the climate system and its changes, as well as for conducting operations in the Arctic. A current challenge however is that most sensory equipment can only capture one or fewer of the characteristics needed to accurately describe ground objects through their temporal, spatial, spectral and radiometric resolution characteristics. This in turn motivates the fusing of complimentary modalities for potentially improved accuracy and stability in analysis but it also leads to problems when trying to merge heterogeneous data with different statistical, geometric and physical qualities. Another concern in the remote sensing of arctic regions is the scarcity of high quality labeled data but simultaneous abundance of unlabeled data as the gathering of labeled data can be both costly and time consuming. It could therefore be of great value to explore routes that can automate this process in ways that target both the situation regarding available data and the difficulties from fusing of heterogeneous multimodal data. To this end Semi-Supervised methods were considered for their ability to leverage smaller amounts of carefully labeled data in combination with more widely available unlabeled data in achieving greater classification performance. Strengths and limitations of three algorithms for real life applications are assessed through experiments on datasets from arctic and urban areas. The first two algorithms, Deep Semi-Supervised Label Propagation (LP) and MixMatch Holistic SSL (MixMatch), consider simultaneous processing of multimodal remote sensing data with additional extracted Gray Level Co-occurrence Matrix texture features for image classification. LP trains in alternating steps of supervised learning on potentially pseudolabeled data and steps of deciding new labels through node propagation while MixMatch mixes loss terms from several leading algorithms to gain their respective benefits. Another method, Graph Fusion Merriman Bence Osher (GMBO), explores processing of modalities in parallel by constructing a fused graph from complimentary input modalities and Ginzburg-Landau minimization on an approximated Graph Laplacian. Results imply that inclusion of extracted GLCM features could be beneficial for classification of multimodal remote sensing data, and that GMBO has merits for operational use in the Arctic given that certain data prerequisites are met.