Chimera

A multi-task recurrent convolutional neural network for forest classification and structural estimation

Tony Chang, Brandon P. Rasmussen, Brett G Dickson, Luke J. Zachmann

Research output: Contribution to journalArticle

Abstract

More consistent and current estimates of forest land cover type and forest structural metrics are needed to guide national policies on forest management, carbon sequestration, and ecosystem health. In recent years, the increased availability of high-resolution (<30m) imagery and advancements in machine learning algorithms have opened up a new opportunity to fuse multiple datasets of varying spatial, spectral, and temporal resolutions. Here, we present a new model, based on a deep learning architecture, that performs both classification and regression concurrently, thereby consolidating what was previously several independent tasks and models into one stream. The model, a multi-task recurrent convolutional neural network that we call the Chimera, integrates varying resolution, freely available aerial and satellite imagery, as well as relevant environmental factors (e.g., climate, terrain) to simultaneously classify five forest cover types ('conifer', 'deciduous', 'mixed', 'dead', 'none' (non-forest)) and to estimate four continuous forest structure metrics (above ground biomass, quadratic mean diameter, basal area, canopy cover). We demonstrate the performance of our approach by training an ensemble of Chimera models on 9967 georeferenced (true locations) Forest Inventory and Analysis field plots from the USDA Forest Service within California and Nevada. Classification diagnostics for the Chimera ensemble on an independent test set produces an overall average precision, recall, and F1-score of 0.92, 0.92, and 0.92. Class-wise F1-scores were high for 'none' (0.99) and 'conifer' (0.85) cover classes, and moderate for the 'mixed' (0.74) class samples. This demonstrates a strong ability to discriminate locations with and without trees. Regression diagnostics on the test set indicate very high accuracy for ensembled estimates of above ground biomass (R 2 = 0.84, RMSE = 37.28 Mg/ha), quadraticmean diameter (R 2 = 0.81, RMSE = 3.74 inches), basal area (R 2 = 0.87, RMSE = 25.88 ft2/ac), and canopy cover (R 2 = 0.89, RMSE = 8.01 percent). Comparative analysis of the Chimera ensemble versus support vector machine and random forest approaches demonstrates increased performance over both methods. Future implementations of the Chimera ensemble on a distributed computing platform could provide continuous, annual estimates of forest structure for other forested landscapes at regional or national scales.

Original languageEnglish (US)
Article number768
JournalRemote Sensing
Volume11
Issue number7
DOIs
StatePublished - Apr 1 2019

Fingerprint

chimera
aboveground biomass
basal area
coniferous tree
canopy
ecosystem health
forest inventory
forest cover
carbon sequestration
satellite imagery
forest management
land cover
imagery
environmental factor
learning
climate

Keywords

  • Forest classification
  • Forest structure
  • High resolution imagery
  • Multi-task learning
  • NAIP
  • Recurrent convolutional neural networks
  • Remote sensing

ASJC Scopus subject areas

  • Earth and Planetary Sciences(all)

Cite this

Chimera : A multi-task recurrent convolutional neural network for forest classification and structural estimation. / Chang, Tony; Rasmussen, Brandon P.; Dickson, Brett G; Zachmann, Luke J.

In: Remote Sensing, Vol. 11, No. 7, 768, 01.04.2019.

Research output: Contribution to journalArticle

@article{a4851b7fd8ff4fad9a4bb1b998c54f0a,
title = "Chimera: A multi-task recurrent convolutional neural network for forest classification and structural estimation",
abstract = "More consistent and current estimates of forest land cover type and forest structural metrics are needed to guide national policies on forest management, carbon sequestration, and ecosystem health. In recent years, the increased availability of high-resolution (<30m) imagery and advancements in machine learning algorithms have opened up a new opportunity to fuse multiple datasets of varying spatial, spectral, and temporal resolutions. Here, we present a new model, based on a deep learning architecture, that performs both classification and regression concurrently, thereby consolidating what was previously several independent tasks and models into one stream. The model, a multi-task recurrent convolutional neural network that we call the Chimera, integrates varying resolution, freely available aerial and satellite imagery, as well as relevant environmental factors (e.g., climate, terrain) to simultaneously classify five forest cover types ('conifer', 'deciduous', 'mixed', 'dead', 'none' (non-forest)) and to estimate four continuous forest structure metrics (above ground biomass, quadratic mean diameter, basal area, canopy cover). We demonstrate the performance of our approach by training an ensemble of Chimera models on 9967 georeferenced (true locations) Forest Inventory and Analysis field plots from the USDA Forest Service within California and Nevada. Classification diagnostics for the Chimera ensemble on an independent test set produces an overall average precision, recall, and F1-score of 0.92, 0.92, and 0.92. Class-wise F1-scores were high for 'none' (0.99) and 'conifer' (0.85) cover classes, and moderate for the 'mixed' (0.74) class samples. This demonstrates a strong ability to discriminate locations with and without trees. Regression diagnostics on the test set indicate very high accuracy for ensembled estimates of above ground biomass (R 2 = 0.84, RMSE = 37.28 Mg/ha), quadraticmean diameter (R 2 = 0.81, RMSE = 3.74 inches), basal area (R 2 = 0.87, RMSE = 25.88 ft2/ac), and canopy cover (R 2 = 0.89, RMSE = 8.01 percent). Comparative analysis of the Chimera ensemble versus support vector machine and random forest approaches demonstrates increased performance over both methods. Future implementations of the Chimera ensemble on a distributed computing platform could provide continuous, annual estimates of forest structure for other forested landscapes at regional or national scales.",
keywords = "Forest classification, Forest structure, High resolution imagery, Multi-task learning, NAIP, Recurrent convolutional neural networks, Remote sensing",
author = "Tony Chang and Rasmussen, {Brandon P.} and Dickson, {Brett G} and Zachmann, {Luke J.}",
year = "2019",
month = "4",
day = "1",
doi = "10.3390/rs11070768",
language = "English (US)",
volume = "11",
journal = "Remote Sensing",
issn = "2072-4292",
publisher = "Multidisciplinary Digital Publishing Institute (MDPI)",
number = "7",

}

TY - JOUR

T1 - Chimera

T2 - A multi-task recurrent convolutional neural network for forest classification and structural estimation

AU - Chang, Tony

AU - Rasmussen, Brandon P.

AU - Dickson, Brett G

AU - Zachmann, Luke J.

PY - 2019/4/1

Y1 - 2019/4/1

N2 - More consistent and current estimates of forest land cover type and forest structural metrics are needed to guide national policies on forest management, carbon sequestration, and ecosystem health. In recent years, the increased availability of high-resolution (<30m) imagery and advancements in machine learning algorithms have opened up a new opportunity to fuse multiple datasets of varying spatial, spectral, and temporal resolutions. Here, we present a new model, based on a deep learning architecture, that performs both classification and regression concurrently, thereby consolidating what was previously several independent tasks and models into one stream. The model, a multi-task recurrent convolutional neural network that we call the Chimera, integrates varying resolution, freely available aerial and satellite imagery, as well as relevant environmental factors (e.g., climate, terrain) to simultaneously classify five forest cover types ('conifer', 'deciduous', 'mixed', 'dead', 'none' (non-forest)) and to estimate four continuous forest structure metrics (above ground biomass, quadratic mean diameter, basal area, canopy cover). We demonstrate the performance of our approach by training an ensemble of Chimera models on 9967 georeferenced (true locations) Forest Inventory and Analysis field plots from the USDA Forest Service within California and Nevada. Classification diagnostics for the Chimera ensemble on an independent test set produces an overall average precision, recall, and F1-score of 0.92, 0.92, and 0.92. Class-wise F1-scores were high for 'none' (0.99) and 'conifer' (0.85) cover classes, and moderate for the 'mixed' (0.74) class samples. This demonstrates a strong ability to discriminate locations with and without trees. Regression diagnostics on the test set indicate very high accuracy for ensembled estimates of above ground biomass (R 2 = 0.84, RMSE = 37.28 Mg/ha), quadraticmean diameter (R 2 = 0.81, RMSE = 3.74 inches), basal area (R 2 = 0.87, RMSE = 25.88 ft2/ac), and canopy cover (R 2 = 0.89, RMSE = 8.01 percent). Comparative analysis of the Chimera ensemble versus support vector machine and random forest approaches demonstrates increased performance over both methods. Future implementations of the Chimera ensemble on a distributed computing platform could provide continuous, annual estimates of forest structure for other forested landscapes at regional or national scales.

AB - More consistent and current estimates of forest land cover type and forest structural metrics are needed to guide national policies on forest management, carbon sequestration, and ecosystem health. In recent years, the increased availability of high-resolution (<30m) imagery and advancements in machine learning algorithms have opened up a new opportunity to fuse multiple datasets of varying spatial, spectral, and temporal resolutions. Here, we present a new model, based on a deep learning architecture, that performs both classification and regression concurrently, thereby consolidating what was previously several independent tasks and models into one stream. The model, a multi-task recurrent convolutional neural network that we call the Chimera, integrates varying resolution, freely available aerial and satellite imagery, as well as relevant environmental factors (e.g., climate, terrain) to simultaneously classify five forest cover types ('conifer', 'deciduous', 'mixed', 'dead', 'none' (non-forest)) and to estimate four continuous forest structure metrics (above ground biomass, quadratic mean diameter, basal area, canopy cover). We demonstrate the performance of our approach by training an ensemble of Chimera models on 9967 georeferenced (true locations) Forest Inventory and Analysis field plots from the USDA Forest Service within California and Nevada. Classification diagnostics for the Chimera ensemble on an independent test set produces an overall average precision, recall, and F1-score of 0.92, 0.92, and 0.92. Class-wise F1-scores were high for 'none' (0.99) and 'conifer' (0.85) cover classes, and moderate for the 'mixed' (0.74) class samples. This demonstrates a strong ability to discriminate locations with and without trees. Regression diagnostics on the test set indicate very high accuracy for ensembled estimates of above ground biomass (R 2 = 0.84, RMSE = 37.28 Mg/ha), quadraticmean diameter (R 2 = 0.81, RMSE = 3.74 inches), basal area (R 2 = 0.87, RMSE = 25.88 ft2/ac), and canopy cover (R 2 = 0.89, RMSE = 8.01 percent). Comparative analysis of the Chimera ensemble versus support vector machine and random forest approaches demonstrates increased performance over both methods. Future implementations of the Chimera ensemble on a distributed computing platform could provide continuous, annual estimates of forest structure for other forested landscapes at regional or national scales.

KW - Forest classification

KW - Forest structure

KW - High resolution imagery

KW - Multi-task learning

KW - NAIP

KW - Recurrent convolutional neural networks

KW - Remote sensing

UR - http://www.scopus.com/inward/record.url?scp=85064007601&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85064007601&partnerID=8YFLogxK

U2 - 10.3390/rs11070768

DO - 10.3390/rs11070768

M3 - Article

VL - 11

JO - Remote Sensing

JF - Remote Sensing

SN - 2072-4292

IS - 7

M1 - 768

ER -