Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests

Matti Mõttus; Matthieu Molinier; Eelis Halme; Hai Cu; Jorma Laaksonen

doi:10.1109/IGARSS47720.2021.9554257

Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests

Matti Mõttus, Matthieu Molinier, Eelis Halme, Hai Cu, Jorma Laaksonen

BA6408 Remote sensing

Aalto University

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

1 Citation (Scopus)

Abstract

Very high resolution remote sensing data of forests, where individual tree crowns are separable, contains structural information on tree size and density. Such information is complementary to the spectral signatures currently used in forestry applications. Advanced machine learning methods, e.g. convolutional neural networks (CNNs), offer an automated and standardized way of retrieving both spectral and structural information from imagery. A key characteristic in CNNs is patch size, which should be large enough to include dominant structural scale, yet as small as possible to avoid unnecessary averaging. Our results show that the patch should be larger than one tree, but increasing it excessively reduces retrieval accuracy. Furthermore, large patch sizes can cause loss of independence between training and validation data, leading to overestimating model performance.

Original language	English
Title of host publication	2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS
Publisher	IEEE Institute of Electrical and Electronic Engineers
Pages	2035-2038
Number of pages	4
ISBN (Electronic)	978-1-6654-0369-6, 978-1-6654-0368-9
ISBN (Print)	978-1-6654-4762-1
DOIs	https://doi.org/10.1109/IGARSS47720.2021.9554257
Publication status	Published - 2021
MoE publication type	A4 Article in a conference publication
Event	IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021 - Brussels, Belgium, Brussels, Belgium Duration: 11 Jul 2021 → 16 Jul 2021

Conference

Conference	IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021
Country/Territory	Belgium
City	Brussels
Period	11/07/21 → 16/07/21

Keywords

Convolutional Neural Networks
Deep learning
forest variable prediction
Hyper-spectral data
Patch size
TAIGA dataset
Very high-resolution imagery

Access to Document

10.1109/IGARSS47720.2021.9554257

Cite this

@inproceedings{81ae4aebf6994181b7a5632109dead0e,

title = "Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests",

abstract = "Very high resolution remote sensing data of forests, where individual tree crowns are separable, contains structural information on tree size and density. Such information is complementary to the spectral signatures currently used in forestry applications. Advanced machine learning methods, e.g. convolutional neural networks (CNNs), offer an automated and standardized way of retrieving both spectral and structural information from imagery. A key characteristic in CNNs is patch size, which should be large enough to include dominant structural scale, yet as small as possible to avoid unnecessary averaging. Our results show that the patch should be larger than one tree, but increasing it excessively reduces retrieval accuracy. Furthermore, large patch sizes can cause loss of independence between training and validation data, leading to overestimating model performance.",

keywords = "Convolutional Neural Networks, Deep learning, forest variable prediction, Hyper-spectral data, Patch size, TAIGA dataset, Very high-resolution imagery",

author = "Matti M{\~o}ttus and Matthieu Molinier and Eelis Halme and Hai Cu and Jorma Laaksonen",

note = "Funding Information: ∗Thanks to Academy of Finland for funding AIROBEST (grant 317387). Publisher Copyright: {\textcopyright} 2021 IEEE; IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021 ; Conference date: 11-07-2021 Through 16-07-2021",

year = "2021",

doi = "10.1109/IGARSS47720.2021.9554257",

language = "English",

isbn = "978-1-6654-4762-1",

pages = "2035--2038",

booktitle = "2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS",

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

address = "United States",

}

Mõttus, M , Molinier, M , Halme, E, Cu, H & Laaksonen, J 2021, Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests. in 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS. IEEE Institute of Electrical and Electronic Engineers, pp. 2035-2038, IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021, Brussels, Belgium, 11/07/21. https://doi.org/10.1109/IGARSS47720.2021.9554257

Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests. / Mõttus, Matti ; Molinier, Matthieu ; Halme, Eelis et al.
2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS. IEEE Institute of Electrical and Electronic Engineers, 2021. p. 2035-2038.

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

TY - GEN

T1 - Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests

AU - Mõttus, Matti

AU - Molinier, Matthieu

AU - Halme, Eelis

AU - Cu, Hai

AU - Laaksonen, Jorma

PY - 2021

Y1 - 2021

N2 - Very high resolution remote sensing data of forests, where individual tree crowns are separable, contains structural information on tree size and density. Such information is complementary to the spectral signatures currently used in forestry applications. Advanced machine learning methods, e.g. convolutional neural networks (CNNs), offer an automated and standardized way of retrieving both spectral and structural information from imagery. A key characteristic in CNNs is patch size, which should be large enough to include dominant structural scale, yet as small as possible to avoid unnecessary averaging. Our results show that the patch should be larger than one tree, but increasing it excessively reduces retrieval accuracy. Furthermore, large patch sizes can cause loss of independence between training and validation data, leading to overestimating model performance.

AB - Very high resolution remote sensing data of forests, where individual tree crowns are separable, contains structural information on tree size and density. Such information is complementary to the spectral signatures currently used in forestry applications. Advanced machine learning methods, e.g. convolutional neural networks (CNNs), offer an automated and standardized way of retrieving both spectral and structural information from imagery. A key characteristic in CNNs is patch size, which should be large enough to include dominant structural scale, yet as small as possible to avoid unnecessary averaging. Our results show that the patch should be larger than one tree, but increasing it excessively reduces retrieval accuracy. Furthermore, large patch sizes can cause loss of independence between training and validation data, leading to overestimating model performance.

KW - Convolutional Neural Networks

KW - Deep learning

KW - forest variable prediction

KW - Hyper-spectral data

KW - Patch size

KW - TAIGA dataset

KW - Very high-resolution imagery

UR - http://www.scopus.com/inward/record.url?scp=85128002788&partnerID=8YFLogxK

U2 - 10.1109/IGARSS47720.2021.9554257

DO - 10.1109/IGARSS47720.2021.9554257

M3 - Conference article in proceedings

SN - 978-1-6654-4762-1

SP - 2035

EP - 2038

BT - 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS

PB - IEEE Institute of Electrical and Electronic Engineers

T2 - IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021

Y2 - 11 July 2021 through 16 July 2021

ER -

Patch size selection for analysis of sub-meter resolution hyperspectral imagery of forests

Abstract

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this