Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels

Matthieu Molinier; Jorma Kilpi

doi:10.1109/IGARSS.2019.8900328

Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels

Matthieu Molinier (Corresponding author), Jorma Kilpi

Research output: Chapter in Book/Report/Conference proceeding › Conference article in proceedings › Scientific › peer-review

18 Citations (Scopus)

Abstract

Spatial-spectral approaches applied on hyperspectral images (HSI) with limited labels suffer from overfitting when the size of input filters and the percentage of training data increases. In those cases, pixel values corresponding to testing sets are partly or completely seen during training phase, reducing the number independent testing pixels and leading to overoptimistic accuracy assessment. These effects have been demonstrated in several previous works but still require attention. In this work we propose additional visulizations and measures of the overlapping and overfitting effects, demonstrated on common HSI datasets, to increase awareness on these issues.

Original language	English
Title of host publication	Proceedings of IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium
Publisher	IEEE Institute of Electrical and Electronic Engineers
Pages	5049-5052
Number of pages	4
ISBN (Electronic)	978-1-5386-9154-0
ISBN (Print)	978-1-5386-9155-7
DOIs	https://doi.org/10.1109/IGARSS.2019.8900328
Publication status	Published - 14 Nov 2019
MoE publication type	A4 Article in a conference publication
Event	39th IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2019 - Yokohama, Japan Duration: 28 Jul 2019 → 2 Aug 2019 https://igarss2019.org

Conference

Conference	39th IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2019
Abbreviated title	IGARSS 2019
Country/Territory	Japan
City	Yokohama
Period	28/07/19 → 2/08/19
Internet address	https://igarss2019.org

Keywords

Deep Learning
Remote sensing
Hyperspectral imagery
limited training samples
overfitting

Access to Document

10.1109/IGARSS.2019.8900328

Cite this

@inproceedings{3b6dca708ce94ab3a179e49d11494b17,

title = "Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels",

abstract = "Spatial-spectral approaches applied on hyperspectral images (HSI) with limited labels suffer from overfitting when the size of input filters and the percentage of training data increases. In those cases, pixel values corresponding to testing sets are partly or completely seen during training phase, reducing the number independent testing pixels and leading to overoptimistic accuracy assessment. These effects have been demonstrated in several previous works but still require attention. In this work we propose additional visulizations and measures of the overlapping and overfitting effects, demonstrated on common HSI datasets, to increase awareness on these issues.",

keywords = "Deep Learning, Remote sensing, Hyperspectral imagery, limited training samples, overfitting",

author = "Matthieu Molinier and Jorma Kilpi",

note = "Funding Information: This work was supported by VTT TechnicalResearch Centre of Finland Ltd through the Substance node on Deep learning Applications and by the AIROBEST project of the Academy of Finland, Grant No. 317387 Funding Information: This work was supported by VTT Technical Research Centre of Finland Ltd through the Substance node on Deep learning Applications and by the AIROBEST project of the Academy of Finland, Grant No. 317387 Publisher Copyright: {\textcopyright} 2019 IEEE.; 39th IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2019, IGARSS 2019 ; Conference date: 28-07-2019 Through 02-08-2019",

year = "2019",

month = nov,

day = "14",

doi = "10.1109/IGARSS.2019.8900328",

language = "English",

isbn = "978-1-5386-9155-7",

pages = "5049--5052",

booktitle = "Proceedings of IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium",

publisher = "IEEE Institute of Electrical and Electronic Engineers",

address = "United States",

url = "https://igarss2019.org",

}

Molinier, M & Kilpi, J 2019, Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels. in Proceedings of IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium. IEEE Institute of Electrical and Electronic Engineers, pp. 5049-5052, 39th IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2019, Yokohama, Japan, 28/07/19. https://doi.org/10.1109/IGARSS.2019.8900328

Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels. / Molinier, Matthieu (Corresponding author); Kilpi, Jorma.
Proceedings of IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium. IEEE Institute of Electrical and Electronic Engineers, 2019. p. 5049-5052.

Research output: Chapter in Book/Report/Conference proceeding › Conference article in proceedings › Scientific › peer-review

TY - GEN

T1 - Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels

AU - Molinier, Matthieu

AU - Kilpi, Jorma

N1 - Funding Information: This work was supported by VTT TechnicalResearch Centre of Finland Ltd through the Substance node on Deep learning Applications and by the AIROBEST project of the Academy of Finland, Grant No. 317387 Funding Information: This work was supported by VTT Technical Research Centre of Finland Ltd through the Substance node on Deep learning Applications and by the AIROBEST project of the Academy of Finland, Grant No. 317387 Publisher Copyright: © 2019 IEEE.

PY - 2019/11/14

Y1 - 2019/11/14

N2 - Spatial-spectral approaches applied on hyperspectral images (HSI) with limited labels suffer from overfitting when the size of input filters and the percentage of training data increases. In those cases, pixel values corresponding to testing sets are partly or completely seen during training phase, reducing the number independent testing pixels and leading to overoptimistic accuracy assessment. These effects have been demonstrated in several previous works but still require attention. In this work we propose additional visulizations and measures of the overlapping and overfitting effects, demonstrated on common HSI datasets, to increase awareness on these issues.

AB - Spatial-spectral approaches applied on hyperspectral images (HSI) with limited labels suffer from overfitting when the size of input filters and the percentage of training data increases. In those cases, pixel values corresponding to testing sets are partly or completely seen during training phase, reducing the number independent testing pixels and leading to overoptimistic accuracy assessment. These effects have been demonstrated in several previous works but still require attention. In this work we propose additional visulizations and measures of the overlapping and overfitting effects, demonstrated on common HSI datasets, to increase awareness on these issues.

KW - Deep Learning

KW - Remote sensing

KW - Hyperspectral imagery

KW - limited training samples

KW - overfitting

UR - http://www.scopus.com/inward/record.url?scp=85089957798&partnerID=8YFLogxK

U2 - 10.1109/IGARSS.2019.8900328

DO - 10.1109/IGARSS.2019.8900328

M3 - Conference article in proceedings

SN - 978-1-5386-9155-7

SP - 5049

EP - 5052

BT - Proceedings of IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium

PB - IEEE Institute of Electrical and Electronic Engineers

T2 - 39th IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2019

Y2 - 28 July 2019 through 2 August 2019

ER -

Avoiding Overfitting When Applying Spectral-Spatial Deep Learning Methods on Hyperspectral Images with Limited Labels

Abstract

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this