Data Reduction and Deep-Learning Based Recovery for Geospatial Visualization and Satellite Imagery
| dc.contributor.advisor | Mondal, Debajyoti | |
| dc.contributor.committeeMember | Stavness, Ian | |
| dc.contributor.committeeMember | Jin, Lingling | |
| dc.contributor.committeeMember | Liu, Juxin | |
| dc.creator | Tasnim, Jarin | |
| dc.date.accessioned | 2021-03-16T21:03:47Z | |
| dc.date.available | 2021-03-16T21:03:47Z | |
| dc.date.created | 2020-12 | |
| dc.date.issued | 2021-03-16 | |
| dc.date.submitted | December 2020 | |
| dc.date.updated | 2021-03-16T21:03:47Z | |
| dc.description.abstract | The storage, retrieval and distribution of data are some critical aspects of big data management. Data scientists and decision-makers often need to share large datasets and make decisions on archiving or deleting historical data to cope with resource constraints. As a consequence, there is an urgency of reducing the storage and transmission requirement. A potential approach to mitigate such problems is to reduce big datasets into smaller ones, which will not only lower storage requirements but also allow light load transfer over the network. The high dimensional data often exhibit high repetitiveness and paradigm across different dimensions. Carefully prepared data by removing redundancies, along with a machine learning model capable of reconstructing the whole dataset from its reduced version, can improve the storage scalability, data transfer, and speed up the overall data management pipeline. In this thesis, we explore some data reduction strategies for big datasets, while ensuring that the data can be transferred and used ubiquitously by all stakeholders, i.e., the entire dataset can be reconstructed with high quality whenever necessary. One of our data reduction strategies follows a straightforward uniform pattern, which guarantees a minimum of 75% data size reduction. We also propose a novel variance based reduction technique, which focuses on removing only redundant data and offers additional 1% to 2% deletion rate. We have adopted various traditional machine learning and deep learning approaches for high-quality reconstruction. We evaluated our pipelines with big geospatial data and satellite imageries. Among them, our deep learning approaches have performed very well both quantitatively and qualitatively with the capability of reconstructing high quality features. We also show how to leverage temporal data for better reconstruction. For uniform deletion, the reconstruction accuracy observed is as high as 98.75% on an average for spatial meteorological data (e.g., soil moisture and albedo), and 99.09% for satellite imagery. Pushing the deletion rate further by following variance based deletion method, the decrease in accuracy remains within 1% for spatial meteorological data and 7% for satellite imagery. | |
| dc.format.mimetype | application/pdf | |
| dc.identifier.uri | http://hdl.handle.net/10388/13285 | |
| dc.subject | Data Reduction | |
| dc.subject | Reconstruction | |
| dc.subject | Deep Learning | |
| dc.subject | SRGAN | |
| dc.subject | Image Inpainting | |
| dc.subject | Geospatial Visualization | |
| dc.title | Data Reduction and Deep-Learning Based Recovery for Geospatial Visualization and Satellite Imagery | |
| dc.type | Thesis | |
| dc.type.material | text | |
| thesis.degree.department | Computer Science | |
| thesis.degree.discipline | Computer Science | |
| thesis.degree.grantor | University of Saskatchewan | |
| thesis.degree.level | Masters | |
| thesis.degree.name | Master of Science (M.Sc.) |