Authors: Zhenlong LI*, University of South Carolina, Huan Ning, University of South Carolina
Topics: Geographic Information Science and Systems, Cyberinfrastructure, Spatial Analysis & Modeling
Keywords: training dataset, deep learning, GIS, remote sensing, classification, big data, neural network
Session Type: Paper
Start / End Time: 12:40 PM / 2:20 PM
Room: Virginia B, Marriott, Lobby Level
Presentation File: No File Uploaded
Deep learning has witnessed significant advancements over the past several years. Many researchers have been using deep learning for remote sensing imagery classification. Most of those studies focus on building the workflow and increasing the classification accuracy based on certain public datasets, such as SpaceNet. In real world, most classification users focus on a particular field or geographical region. If they want to use deep learning for image classification, they need to create customized training dataset to satisfy various purposes, like city planning, illegal building monitor, or precise agriculture in different regions. However, there is still no guideline of how to build the training dataset for industry experts who are not machine learning specialist. To fill this gap, we explore how to create a training dataset by using satellite imagery combined with existing GIS data sources, such as land use data, address points, and geotagged tweets. The results show that the existing GIS data can serve as practical sources for building a training dataset. Comparing the deep learning classification results from the training datasets built with traditional methods, we conclude serval rules on building a customized training dataset to get an optimal classification accuracy.