Deepnet : an extensible data acquisition and curation framework supporting computer vision deep learning research
Abstract
We present a system, DeepNet, for ingestion, curation, and management of geospatial data images to facilitate a range of geospatial research. The system allows for the semi-autonomous ingestion of geospatial data from a variety of sources while pre-serving data integrity and provenance. This data repository can then be utilized in a number of ways, with our focus being on the creation of high resolution remote sensing imagery data sets (HR-RSIDS) of a variety of modalities. In addition to aggregating public data sets, the framework includes open source research data points which are found during the course of curation and quality assurance of data while utilizing the framework. These features are of particular value since there is a real possibility that features found during the utilization of the system do not exist in any previously ingested geodata source; the result being that the system grows through its utilization without the reliance on third-party geodata data sets being released publicly. By using DeepNet, we can substantially accelerate the production of HR-RSIDS from public data that is available without imagery. While curating the imagery and data does take a non-trivial amount of time, DeepNet has been designed to support multiple-user curation via a web application component. Allowing multiple people the ability to curate the data and imagery simultaneously allows for the curation process to scale linearly with the number of people doing the curation work. Once the data have been curated, creating a new data set from the data can be done in a matter of minutes. This is especially useful because it allows the data sets created with DeepNet to be corrected, augmented, and extended with new data. We present both the design and implementation of the framework, as well as a number of current and potential uses for the data that the framework manages.
Degree
M.S.
Thesis Department
Rights
OpenAccess.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License.