The poor generalization of deep convolutional networks to aerial imagery from new geographic locations: An empirical study with solar array detection

Rui Wang, Joseph Camilo, Leslie M. Collins, Kyle Bradbury, Jordan M. Malof

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Convolutional neural networks (CNNs) have recently achieved unprecedented performance for the automatic recognition of objects (e.g., buildings, roads, or vehicles) in color aerial imagery. Although these results are promising, questions remain about their practical applicability. This is because there is a wide variability in the visual characteristics of remote sensing imagery across different geographic locations, and CNNs are often trained and tested on imagery from nearby (or the same) geographic locations. It is therefore unclear whether trained CNNs will perform well on new, previously unseen, geographic locations, which is an important practical consideration. In this work we investigate this problem when applying CNNs for solar array detection on a large aerial imagery dataset comprised of two nearby US cities. We compare the performance of CNNs under two conditions: Training and testing on the same city vs training on one city and testing on another city. We discuss several subtle difficulties with these experiments and make recommendations. We show that there can be substantial performance loss in second case, when compared to the first. We also investigate how much training data is required from the unseen city in order to fine-tune the CNN so that it performs well. We investigate several different fine-tuning strategies, yielding a clear winner.

Original languageEnglish
Title of host publication2017 IEEE Applied Imagery Pattern Recognition Workshop, AIPR 2017
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781538612354
DOIs
StatePublished - Jul 2 2017
Event2017 IEEE Applied Imagery Pattern Recognition Workshop, AIPR 2017 - Washington, United States
Duration: Oct 10 2017Oct 12 2017

Publication series

NameProceedings - Applied Imagery Pattern Recognition Workshop
Volume2017-October
ISSN (Print)2164-2516

Conference

Conference2017 IEEE Applied Imagery Pattern Recognition Workshop, AIPR 2017
Country/TerritoryUnited States
CityWashington
Period10/10/1710/12/17

Keywords

  • Semantic segmentation
  • aerial imagery
  • geographic generalization
  • remote sensing

Fingerprint

Dive into the research topics of 'The poor generalization of deep convolutional networks to aerial imagery from new geographic locations: An empirical study with solar array detection'. Together they form a unique fingerprint.

Cite this