Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure

Research output: Chapter in Book/Report/Conference proceedingConference abstract in proceedingsScientific

Abstract

Timely and accurate information on forest above-ground biomass (AGB) is required for understanding carbon balance, future climate, the sustainability of current politics and the emerging bioeconomy. Optical Earth Observation (EO) in the visible to short infrared spectral region can contribute to this challenge. This is supported by the more frequent and higher quality EO data, mainly contributed by the new Sentinel-series satellites developed and launched by the European Space Agency (ESA). The new satellites deliver data with better spatial and spectral resolutions in a rapidly increasing volume, calling for new processing algorithms. Current algorithms fall short in making a full use of the higher information content of the new data, and have trouble handling the data volume produced by current and forthcoming satellite instruments. Artificial intelligence (AI) has shown promise to overcome many of the shortcomings in existing empirical methods for retrieving forest characteristics, such as its structure and biomass, from EO data. It can also handle efficiently and rapidly large data volumes. However, despite optimistic case studies, no large-scale applications using optical EO for biomass retrieval have emerged. Up to now, AI has been applied to simple canopies with little structure and no woody biomass. Recent developments in deep learning have made it feasible to utilize it for analyzing complex structured vegetation with significant AGB storage such as the boreal forest. We hypothesize that with a carefully selected approach, deep learning approaches can retrieve the forest AGB which is indirectly, but physically, related to the structural parameters affecting forest reflectance, e.g., crown volume and shape, or tree density. We propose to use a well-validated physically-based forest reflectance model to simulate the spectral reflectance factors of all possible European boreal forest canopies under different illumination conditions. The model will be used to simulate the reflectance signatures of a wide range of forests with structural parameters taken from forestry databases in Sweden, Finland, Estonia and Russia. We will use the simulated reflectance signatures, corresponding to satellite-measured signals, to train a combination of deep learning algorithms with several desired properties such as : - Recurrent Neural Networks or Long Short-Term Memory (LSTM) networks for regression - Fully Convolutional Networks (FCN) or Convolutional Neural Networks (CNN) for their ability to extract hierarchical visual features from multispectral images - semi-supervised approaches to deal with the limited availability of labelled training data - deep probabilistic approaches to provide confidence intervals on the estimated outputs Once trained, we will apply these algorithms to optical EO data from the European boreal zone, and hyperspectral data from Finland. The AI retrieval results will be compared against forestry data from test sites in each of these regions. We will also investigate the potential of Generative Adversarial Networks (GANs) to augment the available reference forest data and generate more diverse and realistic images than from simulations based solely on physical models. AIROBEST will provide a new look on our forests on a continental scale. It will deepen the apprehension of AI methods in Earth Observation and our understanding of the changes in environment. The algorithms produced in the project will be integrated with other EO-based forest monitoring tools developed at VTT, creating a capacity for comprehensive forest information retrieval for research, governmental bodies, and industry. This study is part of the AIROBEST project on Artificial Intelligence for Retrieval of Forest Biomass & Structure, funded by the Finnish Academy (grant 317387) as part of the programme on Novel Applications of Artificial Intelligence in Physical Sciences and Engineering Research (AIPSE).
Original languageEnglish
Title of host publicationESA Living Planet Symposium 2019
Place of PublicationMilano
PublisherEuropean Space Agency ESA
Publication statusPublished - 16 May 2019
MoE publication typeNot Eligible
EventESA Living Planet Symposium 2019 - Milano, Italy
Duration: 13 May 201917 May 2019
https://lps19.esa.int

Conference

ConferenceESA Living Planet Symposium 2019
CountryItaly
CityMilano
Period13/05/1917/05/19
Internet address

Fingerprint

learning
artificial intelligence
biomass
reflectance
aboveground biomass
boreal forest
forestry
method
physical science
multispectral image
carbon balance
spectral reflectance
forest canopy
spectral resolution
confidence interval
train
satellite data
spatial resolution
politics
canopy

Cite this

Molinier, M., Astola, H., & Mõttus, M. (2019). Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure. In ESA Living Planet Symposium 2019 Milano: European Space Agency ESA.
@inbook{3dbbcf85f2b047c79d09820aa996c260,
title = "Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure",
abstract = "Timely and accurate information on forest above-ground biomass (AGB) is required for understanding carbon balance, future climate, the sustainability of current politics and the emerging bioeconomy. Optical Earth Observation (EO) in the visible to short infrared spectral region can contribute to this challenge. This is supported by the more frequent and higher quality EO data, mainly contributed by the new Sentinel-series satellites developed and launched by the European Space Agency (ESA). The new satellites deliver data with better spatial and spectral resolutions in a rapidly increasing volume, calling for new processing algorithms. Current algorithms fall short in making a full use of the higher information content of the new data, and have trouble handling the data volume produced by current and forthcoming satellite instruments. Artificial intelligence (AI) has shown promise to overcome many of the shortcomings in existing empirical methods for retrieving forest characteristics, such as its structure and biomass, from EO data. It can also handle efficiently and rapidly large data volumes. However, despite optimistic case studies, no large-scale applications using optical EO for biomass retrieval have emerged. Up to now, AI has been applied to simple canopies with little structure and no woody biomass. Recent developments in deep learning have made it feasible to utilize it for analyzing complex structured vegetation with significant AGB storage such as the boreal forest. We hypothesize that with a carefully selected approach, deep learning approaches can retrieve the forest AGB which is indirectly, but physically, related to the structural parameters affecting forest reflectance, e.g., crown volume and shape, or tree density. We propose to use a well-validated physically-based forest reflectance model to simulate the spectral reflectance factors of all possible European boreal forest canopies under different illumination conditions. The model will be used to simulate the reflectance signatures of a wide range of forests with structural parameters taken from forestry databases in Sweden, Finland, Estonia and Russia. We will use the simulated reflectance signatures, corresponding to satellite-measured signals, to train a combination of deep learning algorithms with several desired properties such as : - Recurrent Neural Networks or Long Short-Term Memory (LSTM) networks for regression - Fully Convolutional Networks (FCN) or Convolutional Neural Networks (CNN) for their ability to extract hierarchical visual features from multispectral images - semi-supervised approaches to deal with the limited availability of labelled training data - deep probabilistic approaches to provide confidence intervals on the estimated outputs Once trained, we will apply these algorithms to optical EO data from the European boreal zone, and hyperspectral data from Finland. The AI retrieval results will be compared against forestry data from test sites in each of these regions. We will also investigate the potential of Generative Adversarial Networks (GANs) to augment the available reference forest data and generate more diverse and realistic images than from simulations based solely on physical models. AIROBEST will provide a new look on our forests on a continental scale. It will deepen the apprehension of AI methods in Earth Observation and our understanding of the changes in environment. The algorithms produced in the project will be integrated with other EO-based forest monitoring tools developed at VTT, creating a capacity for comprehensive forest information retrieval for research, governmental bodies, and industry. This study is part of the AIROBEST project on Artificial Intelligence for Retrieval of Forest Biomass & Structure, funded by the Finnish Academy (grant 317387) as part of the programme on Novel Applications of Artificial Intelligence in Physical Sciences and Engineering Research (AIPSE).",
author = "Matthieu Molinier and Heikki Astola and Matti M{\~o}ttus",
year = "2019",
month = "5",
day = "16",
language = "English",
booktitle = "ESA Living Planet Symposium 2019",
publisher = "European Space Agency ESA",
address = "France",

}

Molinier, M, Astola, H & Mõttus, M 2019, Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure. in ESA Living Planet Symposium 2019. European Space Agency ESA, Milano, ESA Living Planet Symposium 2019, Milano, Italy, 13/05/19.

Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure. / Molinier, Matthieu (Corresponding author); Astola, Heikki; Mõttus, Matti.

ESA Living Planet Symposium 2019. Milano : European Space Agency ESA, 2019.

Research output: Chapter in Book/Report/Conference proceedingConference abstract in proceedingsScientific

TY - CHAP

T1 - Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure

AU - Molinier, Matthieu

AU - Astola, Heikki

AU - Mõttus, Matti

PY - 2019/5/16

Y1 - 2019/5/16

N2 - Timely and accurate information on forest above-ground biomass (AGB) is required for understanding carbon balance, future climate, the sustainability of current politics and the emerging bioeconomy. Optical Earth Observation (EO) in the visible to short infrared spectral region can contribute to this challenge. This is supported by the more frequent and higher quality EO data, mainly contributed by the new Sentinel-series satellites developed and launched by the European Space Agency (ESA). The new satellites deliver data with better spatial and spectral resolutions in a rapidly increasing volume, calling for new processing algorithms. Current algorithms fall short in making a full use of the higher information content of the new data, and have trouble handling the data volume produced by current and forthcoming satellite instruments. Artificial intelligence (AI) has shown promise to overcome many of the shortcomings in existing empirical methods for retrieving forest characteristics, such as its structure and biomass, from EO data. It can also handle efficiently and rapidly large data volumes. However, despite optimistic case studies, no large-scale applications using optical EO for biomass retrieval have emerged. Up to now, AI has been applied to simple canopies with little structure and no woody biomass. Recent developments in deep learning have made it feasible to utilize it for analyzing complex structured vegetation with significant AGB storage such as the boreal forest. We hypothesize that with a carefully selected approach, deep learning approaches can retrieve the forest AGB which is indirectly, but physically, related to the structural parameters affecting forest reflectance, e.g., crown volume and shape, or tree density. We propose to use a well-validated physically-based forest reflectance model to simulate the spectral reflectance factors of all possible European boreal forest canopies under different illumination conditions. The model will be used to simulate the reflectance signatures of a wide range of forests with structural parameters taken from forestry databases in Sweden, Finland, Estonia and Russia. We will use the simulated reflectance signatures, corresponding to satellite-measured signals, to train a combination of deep learning algorithms with several desired properties such as : - Recurrent Neural Networks or Long Short-Term Memory (LSTM) networks for regression - Fully Convolutional Networks (FCN) or Convolutional Neural Networks (CNN) for their ability to extract hierarchical visual features from multispectral images - semi-supervised approaches to deal with the limited availability of labelled training data - deep probabilistic approaches to provide confidence intervals on the estimated outputs Once trained, we will apply these algorithms to optical EO data from the European boreal zone, and hyperspectral data from Finland. The AI retrieval results will be compared against forestry data from test sites in each of these regions. We will also investigate the potential of Generative Adversarial Networks (GANs) to augment the available reference forest data and generate more diverse and realistic images than from simulations based solely on physical models. AIROBEST will provide a new look on our forests on a continental scale. It will deepen the apprehension of AI methods in Earth Observation and our understanding of the changes in environment. The algorithms produced in the project will be integrated with other EO-based forest monitoring tools developed at VTT, creating a capacity for comprehensive forest information retrieval for research, governmental bodies, and industry. This study is part of the AIROBEST project on Artificial Intelligence for Retrieval of Forest Biomass & Structure, funded by the Finnish Academy (grant 317387) as part of the programme on Novel Applications of Artificial Intelligence in Physical Sciences and Engineering Research (AIPSE).

AB - Timely and accurate information on forest above-ground biomass (AGB) is required for understanding carbon balance, future climate, the sustainability of current politics and the emerging bioeconomy. Optical Earth Observation (EO) in the visible to short infrared spectral region can contribute to this challenge. This is supported by the more frequent and higher quality EO data, mainly contributed by the new Sentinel-series satellites developed and launched by the European Space Agency (ESA). The new satellites deliver data with better spatial and spectral resolutions in a rapidly increasing volume, calling for new processing algorithms. Current algorithms fall short in making a full use of the higher information content of the new data, and have trouble handling the data volume produced by current and forthcoming satellite instruments. Artificial intelligence (AI) has shown promise to overcome many of the shortcomings in existing empirical methods for retrieving forest characteristics, such as its structure and biomass, from EO data. It can also handle efficiently and rapidly large data volumes. However, despite optimistic case studies, no large-scale applications using optical EO for biomass retrieval have emerged. Up to now, AI has been applied to simple canopies with little structure and no woody biomass. Recent developments in deep learning have made it feasible to utilize it for analyzing complex structured vegetation with significant AGB storage such as the boreal forest. We hypothesize that with a carefully selected approach, deep learning approaches can retrieve the forest AGB which is indirectly, but physically, related to the structural parameters affecting forest reflectance, e.g., crown volume and shape, or tree density. We propose to use a well-validated physically-based forest reflectance model to simulate the spectral reflectance factors of all possible European boreal forest canopies under different illumination conditions. The model will be used to simulate the reflectance signatures of a wide range of forests with structural parameters taken from forestry databases in Sweden, Finland, Estonia and Russia. We will use the simulated reflectance signatures, corresponding to satellite-measured signals, to train a combination of deep learning algorithms with several desired properties such as : - Recurrent Neural Networks or Long Short-Term Memory (LSTM) networks for regression - Fully Convolutional Networks (FCN) or Convolutional Neural Networks (CNN) for their ability to extract hierarchical visual features from multispectral images - semi-supervised approaches to deal with the limited availability of labelled training data - deep probabilistic approaches to provide confidence intervals on the estimated outputs Once trained, we will apply these algorithms to optical EO data from the European boreal zone, and hyperspectral data from Finland. The AI retrieval results will be compared against forestry data from test sites in each of these regions. We will also investigate the potential of Generative Adversarial Networks (GANs) to augment the available reference forest data and generate more diverse and realistic images than from simulations based solely on physical models. AIROBEST will provide a new look on our forests on a continental scale. It will deepen the apprehension of AI methods in Earth Observation and our understanding of the changes in environment. The algorithms produced in the project will be integrated with other EO-based forest monitoring tools developed at VTT, creating a capacity for comprehensive forest information retrieval for research, governmental bodies, and industry. This study is part of the AIROBEST project on Artificial Intelligence for Retrieval of Forest Biomass & Structure, funded by the Finnish Academy (grant 317387) as part of the programme on Novel Applications of Artificial Intelligence in Physical Sciences and Engineering Research (AIPSE).

UR - https://lps19.esa.int/NikalWebsitePortal/living-planet-symposium-2019/lps19/Agenda/AgendaItemDetail?id=3c5e6eef-3661-4b8c-8d02-81795eb0cb43

M3 - Conference abstract in proceedings

BT - ESA Living Planet Symposium 2019

PB - European Space Agency ESA

CY - Milano

ER -

Molinier M, Astola H, Mõttus M. Combining Deep Learning Methods for Retrieval of Forest Biomass & Structure. In ESA Living Planet Symposium 2019. Milano: European Space Agency ESA. 2019