Toward Easy Export of Imagery Products and Feature Classes as Training Data for Deep Learning Frameworks

Authors: Dawn J. Wright*, Esri, Thomas Maurer, Esri, Hua Wei, Esri
Topics: Geographic Information Science and Systems, Remote Sensing
Keywords: deep learning, artificial intelligence, GIS
Session Type: Paper
Day: 4/12/2018
Start / End Time: 10:00 AM / 11:40 AM
Room: Grand Ballroom A, Astor, 2nd Floor
Presentation File: No File Uploaded


Whether to train a Deep Learning (DL) model to find objects of interest such as cars or solar panels in satellite or aerial images, or to classify such images into different categories of land-use, or other such tasks, a common starting point is always labeled ground truth or training data. From an industry perspective, an organization such as ESRI has a large user base of roughly 350,000 agencies, universities, non-profits, and other partners, with most of them maintaining and permanently updating their own GIS data. But how to allow this treasure trove of data to be effectively and appropriately used for training new DL models? This talk will provide an overview of new tools to export GIS data from multiple sources into popular DL formats such as KITTI or PASCAL_VOC. These can then be directly used as input to DL frameworks such as Microsoft CNTK or Google TensorFlow in order to train DL models. For example, NAIP images and building footprints of an entire county can be exported as a sequence of equally sized image chips plus one meta data file per image chip containing the bounding boxes around all buildings in KITTI format. From this data a DL model can be trained that detects buildings. The hope is that this new suite of tools will make it easier for DL researchers and students at all levels (from undergraduate to doctoral and beyond) to access existing GIS data and to use them for training new DL models.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login