Comparison of Sentinel-2 and Landsat 8 imagery for forest variable prediction in boreal region

Research output: Contribution to journalArticleScientificpeer-review

3 Citations (Scopus)

Abstract

We compared the performance of Sentinel-2 and Landsat 8 data for forest variable prediction in the boreal forest of Southern Finland. We defined twelve modelling setups to train multivariable prediction models with either multilayer perceptron (MLP) or regression tree models with the brute force forward selection method. The reference data consisted of 739 circular field plots that had been collected by the Finnish Forest Centre concurrently with the Sentinel-2 and Landsat 8 acquisitions. The input data were divided into training, validation and test sets of equal sizes for 100 iterations in each modelling setup. The predicted forest variables were stem volume (V), stem diameter (D), tree height (H) and basal area (G), and their species-wise components for pine (Pine), spruce (Spr) and broadleaved (BL) trees. We recorded the performance figures and the best predictive image bands for each modelling setup. The best average performance over the 100 modelling iterations was obtained using all Sentinel-2 bands. The plot-level relative root mean square errors (RMSE%) of the field observed mean were 38.4% for average stem diameter, 42.5% for stem basal area/ha, 30.4% for average tree height, and 59.3% for growing stock volume/ha with variables including all tree species. The corresponding best figures with all Landsat 8 bands were RMSE% = 44.6%, 50.2%, 36.6% and 72.2%, respectively. The Sentinel-2 outperformed Landsat 8 also when using near-equivalent image bands and Sentinel-2 data down-sampled to 30 m pixel resolution. The relative systematic error (bias%) did not show any significant differences between Sentinel-2 and Landsat 8 data: the average of the absolute value of bias% was 0.8% for Sentinel-2 and 1.2% for Landsat 8. The best predictive Sentinel-2 image band was the red-edge 1 (B05_RE1), when variable totals including all species were estimated. The short-wave infrared bands (B11_SWIR1 & B12_SWIR2) and the visible green band (B03_Green) were also among the best predictors. The median number of predictors in the best performing models was 4–6 for the Sentinel-2 and 4–5 for the Landsat 8 models, respectively. We conclude that Sentinel-2 Multispectral Instrument (MSI) data can be recommended as the principal Earth observation data source in forest resources assessment.

Original languageEnglish
Pages (from-to)257-273
Number of pages17
JournalRemote Sensing of Environment
Volume223
DOIs
Publication statusPublished - 15 Mar 2019
MoE publication typeNot Eligible

Fingerprint

Landsat
imagery
prediction
Mean square error
stem
stems
Systematic errors
basal area
Multilayer neural networks
modeling
Pinus
Pixels
Earth (planet)
Infrared radiation
comparison
forest resources
boreal forest
selection methods
train
boreal forests

Keywords

  • Boreal forest
  • Forest variables
  • Forestry
  • Landsat 8
  • Sentinel-2

Cite this

@article{a117ec70187f4e818219dd62c30fb5e6,
title = "Comparison of Sentinel-2 and Landsat 8 imagery for forest variable prediction in boreal region",
abstract = "We compared the performance of Sentinel-2 and Landsat 8 data for forest variable prediction in the boreal forest of Southern Finland. We defined twelve modelling setups to train multivariable prediction models with either multilayer perceptron (MLP) or regression tree models with the brute force forward selection method. The reference data consisted of 739 circular field plots that had been collected by the Finnish Forest Centre concurrently with the Sentinel-2 and Landsat 8 acquisitions. The input data were divided into training, validation and test sets of equal sizes for 100 iterations in each modelling setup. The predicted forest variables were stem volume (V), stem diameter (D), tree height (H) and basal area (G), and their species-wise components for pine (Pine), spruce (Spr) and broadleaved (BL) trees. We recorded the performance figures and the best predictive image bands for each modelling setup. The best average performance over the 100 modelling iterations was obtained using all Sentinel-2 bands. The plot-level relative root mean square errors (RMSE{\%}) of the field observed mean were 38.4{\%} for average stem diameter, 42.5{\%} for stem basal area/ha, 30.4{\%} for average tree height, and 59.3{\%} for growing stock volume/ha with variables including all tree species. The corresponding best figures with all Landsat 8 bands were RMSE{\%} = 44.6{\%}, 50.2{\%}, 36.6{\%} and 72.2{\%}, respectively. The Sentinel-2 outperformed Landsat 8 also when using near-equivalent image bands and Sentinel-2 data down-sampled to 30 m pixel resolution. The relative systematic error (bias{\%}) did not show any significant differences between Sentinel-2 and Landsat 8 data: the average of the absolute value of bias{\%} was 0.8{\%} for Sentinel-2 and 1.2{\%} for Landsat 8. The best predictive Sentinel-2 image band was the red-edge 1 (B05_RE1), when variable totals including all species were estimated. The short-wave infrared bands (B11_SWIR1 & B12_SWIR2) and the visible green band (B03_Green) were also among the best predictors. The median number of predictors in the best performing models was 4–6 for the Sentinel-2 and 4–5 for the Landsat 8 models, respectively. We conclude that Sentinel-2 Multispectral Instrument (MSI) data can be recommended as the principal Earth observation data source in forest resources assessment.",
keywords = "Boreal forest, Forest variables, Forestry, Landsat 8, Sentinel-2",
author = "Heikki Astola and Tuomas H{\"a}me and Laura Sirro and Matthieu Molinier and Jorma Kilpi",
note = "Project 121842",
year = "2019",
month = "3",
day = "15",
doi = "10.1016/j.rse.2019.01.019",
language = "English",
volume = "223",
pages = "257--273",
journal = "Remote Sensing of Environment",
issn = "0034-4257",
publisher = "Elsevier",

}

TY - JOUR

T1 - Comparison of Sentinel-2 and Landsat 8 imagery for forest variable prediction in boreal region

AU - Astola, Heikki

AU - Häme, Tuomas

AU - Sirro, Laura

AU - Molinier, Matthieu

AU - Kilpi, Jorma

N1 - Project 121842

PY - 2019/3/15

Y1 - 2019/3/15

N2 - We compared the performance of Sentinel-2 and Landsat 8 data for forest variable prediction in the boreal forest of Southern Finland. We defined twelve modelling setups to train multivariable prediction models with either multilayer perceptron (MLP) or regression tree models with the brute force forward selection method. The reference data consisted of 739 circular field plots that had been collected by the Finnish Forest Centre concurrently with the Sentinel-2 and Landsat 8 acquisitions. The input data were divided into training, validation and test sets of equal sizes for 100 iterations in each modelling setup. The predicted forest variables were stem volume (V), stem diameter (D), tree height (H) and basal area (G), and their species-wise components for pine (Pine), spruce (Spr) and broadleaved (BL) trees. We recorded the performance figures and the best predictive image bands for each modelling setup. The best average performance over the 100 modelling iterations was obtained using all Sentinel-2 bands. The plot-level relative root mean square errors (RMSE%) of the field observed mean were 38.4% for average stem diameter, 42.5% for stem basal area/ha, 30.4% for average tree height, and 59.3% for growing stock volume/ha with variables including all tree species. The corresponding best figures with all Landsat 8 bands were RMSE% = 44.6%, 50.2%, 36.6% and 72.2%, respectively. The Sentinel-2 outperformed Landsat 8 also when using near-equivalent image bands and Sentinel-2 data down-sampled to 30 m pixel resolution. The relative systematic error (bias%) did not show any significant differences between Sentinel-2 and Landsat 8 data: the average of the absolute value of bias% was 0.8% for Sentinel-2 and 1.2% for Landsat 8. The best predictive Sentinel-2 image band was the red-edge 1 (B05_RE1), when variable totals including all species were estimated. The short-wave infrared bands (B11_SWIR1 & B12_SWIR2) and the visible green band (B03_Green) were also among the best predictors. The median number of predictors in the best performing models was 4–6 for the Sentinel-2 and 4–5 for the Landsat 8 models, respectively. We conclude that Sentinel-2 Multispectral Instrument (MSI) data can be recommended as the principal Earth observation data source in forest resources assessment.

AB - We compared the performance of Sentinel-2 and Landsat 8 data for forest variable prediction in the boreal forest of Southern Finland. We defined twelve modelling setups to train multivariable prediction models with either multilayer perceptron (MLP) or regression tree models with the brute force forward selection method. The reference data consisted of 739 circular field plots that had been collected by the Finnish Forest Centre concurrently with the Sentinel-2 and Landsat 8 acquisitions. The input data were divided into training, validation and test sets of equal sizes for 100 iterations in each modelling setup. The predicted forest variables were stem volume (V), stem diameter (D), tree height (H) and basal area (G), and their species-wise components for pine (Pine), spruce (Spr) and broadleaved (BL) trees. We recorded the performance figures and the best predictive image bands for each modelling setup. The best average performance over the 100 modelling iterations was obtained using all Sentinel-2 bands. The plot-level relative root mean square errors (RMSE%) of the field observed mean were 38.4% for average stem diameter, 42.5% for stem basal area/ha, 30.4% for average tree height, and 59.3% for growing stock volume/ha with variables including all tree species. The corresponding best figures with all Landsat 8 bands were RMSE% = 44.6%, 50.2%, 36.6% and 72.2%, respectively. The Sentinel-2 outperformed Landsat 8 also when using near-equivalent image bands and Sentinel-2 data down-sampled to 30 m pixel resolution. The relative systematic error (bias%) did not show any significant differences between Sentinel-2 and Landsat 8 data: the average of the absolute value of bias% was 0.8% for Sentinel-2 and 1.2% for Landsat 8. The best predictive Sentinel-2 image band was the red-edge 1 (B05_RE1), when variable totals including all species were estimated. The short-wave infrared bands (B11_SWIR1 & B12_SWIR2) and the visible green band (B03_Green) were also among the best predictors. The median number of predictors in the best performing models was 4–6 for the Sentinel-2 and 4–5 for the Landsat 8 models, respectively. We conclude that Sentinel-2 Multispectral Instrument (MSI) data can be recommended as the principal Earth observation data source in forest resources assessment.

KW - Boreal forest

KW - Forest variables

KW - Forestry

KW - Landsat 8

KW - Sentinel-2

UR - http://www.scopus.com/inward/record.url?scp=85060538152&partnerID=8YFLogxK

U2 - 10.1016/j.rse.2019.01.019

DO - 10.1016/j.rse.2019.01.019

M3 - Article

VL - 223

SP - 257

EP - 273

JO - Remote Sensing of Environment

JF - Remote Sensing of Environment

SN - 0034-4257

ER -