txt/2107.14072.txt

What Does TERRA-REF’s High Resolution, Multi Sensor Plant Sensing Public
Domain Data Offer the Computer Vision Community?
David LeBauer
University of Arizona
[email protected] Burnette
University of Illinois
[email protected] Fahlgren
Donald Danforth Plant Science Center
[email protected]
Rob Kooper
University of Illinois
[email protected] McHenry
University of Illinois
[email protected] Stylianou
St. Louis University
[email protected]
Abstract
A core objective of the TERRA-REF project was to gen-
erate an open-access reference dataset for the evaluation
of sensing technologies to study plants under field condi-
tions. The TERRA-REF program deployed a suite of high-
resolution, cutting edge technology sensors on a gantry sys-
tem with the aim of scanning 1 hectare (104m) at around 1
mm2spatial resolution multiple times per week. The system
contains co-located sensors including a stereo-pair RGB
camera, a thermal imager, a laser scanner to capture 3D
structure, and two hyperspectral cameras covering wave-
lengths of 300-2500nm. This sensor data is provided along-
side over sixty types of traditional plant phenotype measure-
ments that can be used to train new machine learning mod-
els. Associated weather and environmental measurements,
information about agronomic management and experimen-
tal design, and the genomic sequences of hundreds of plant
varieties have been collected and are available alongside
the sensor and plant phenotype data.
Over the course of four years and ten growing seasons,
the TERRA-REF system generated over 1 PB of sensor data
and almost 45 million files. The subset that has been re-
leased to the public domain accounts for two seasons and
about half of the total data volume. This provides an un-
precedented opportunity for investigations far beyond the
core biological scope of the project.
The focus of this paper is to provide the Computer Vi-
sion and Machine Learning communities an overview of the
available data and some potential applications of this one
of a kind data.1. Introduction
In 2015, the Advanced Research Projects Agency for En-
ergy (ARPA-E) funded the TERRA-REF Phenotyping Plat-
form (Figure 1). The scientific aim was to transform plant
breeding by providing a reference dataset generated by de-
ploying a suite of co-located high-resolution sensors un-
der field conditions. The goal of these sensors was to use
proximate sensing from approximately 2m above the plant
canopy to quantify plant characteristics.
The study has evaluated diverse populations of sorghum,
wheat, and lettuce over the course of four years and ten
cropping cycles. Future releases of additional data will be
informed by user interests.
Figure 1. TERRA-REF field scanner at the University of Arizona’s
Maricopa Agricultural Center.
The TERRA-REF reference dataset can be used to char-
acterize phenotype-to-genotype associations, on a genomic
scale, that will enable knowledge-driven breeding and the
development of higher-yielding cultivars of sorghum and
wheat. The data is also being used to develop new algo-
rithms for machine learning, image analysis, genomics, and
optical sensor engineering. Beyond applications in plantbreeding, the resulting dataset provides opportunities for the
study and integration of diverse remote sensing modalities.
1.1. Types of Data
The TERRA-REF field scanner platform utilizes a sensor
suite of co-located instruments (Figure 2 and Table 1). The
TERRA-REF reference dataset includes several data types
(Figures 3 and 4, Table 2) including raw and processed
outputs from sensors, environmental sensor measurements,
manually measured and computationally derived pheno-
types, and raw and processed genomics datasets [16]. Ex-
tensive contextual measurements and metadata include sen-
sor information and extensive documentation for each of the
sensors, the field scanner, calibration targets, and the results
of sensor validation tests [16].
In addition to raw sensor data, the first release of
TERRA-REF data includes derived sensor data products in
enhanced formats including calibrated and georeferenced
images and point clouds (Table 2). Many of the data prod-
ucts are provided in formats that follow Open Geospatial
Consortium (OGC) standards and work with GIS software.
Figure 2. TERRA-REF field scanner sensor suite.
1.2. Sensors
Sensors available on the TERRA-REF field scanner in-
clude snapshot and line-scan imaging, multi-spectral radio-
metric, and environmental sensors. Table 1 and Figure 2)
provide a high level overview of the sensors deployed on
this system. Full documentation and metadata for each sen-
sor as well as the configuration and geometry of the sensor
box are provided as metadata alongside the TERRA-REF
data release.
2. Computer Vision and Machine Learning
Problems
There are a variety of questions that the TERRA-REF
dataset could be used to answer that are of high importance
to the agricultural and plant science communities, while
also posing extremely interesting and challenging computer
vision and machine learning problems. In this section, weconsider example research areas or topics within computer
vision and discuss the relevant agricultural and plant sci-
ence questions those research communities could help ad-
dress using the TERRA-REF data.
Measurement, Prediction and Causal Inference. The
TERRA-REF sensor data can be used to drive development
of vision-based algorithms for fundamental problems in
plant phenotyping, such as making measurements of plant
height, leaf length, flower counting, or estimating environ-
mental stress. Additional challenges include attempting to
predict end of season phenotypes, such as end of season
yield, from early season sensor data – an accurate predic-
tor of end of season yield from early season visual data, for
example, could help growers and breeders invest resources
only in the most promising of candidate crops. There are
additional opportunities to investigate the causal relation-
ship between genotypes or environmental conditions and
their expressed phenotypes, as the TERRA-REF dataset in-
cludes both comprehensive genetic information, as well as
high temporal resolution environmental information. The
TERRA-REF data contain over sixty hand measurements
that could be used to train models from one or more sensors.
In addition, there are opportunities to train models that pre-
dict plot-level phenotypes measured by an expensive sensor
with a less expensive sensor. Further, many events includ-
ing insect damage, heat stress, and plant lodging (falling)
could be labeled in new images.
Fine Grained Visual Categorization. The TERRA-REF
data is a rich source of visual sensor data collected from
crop species that are visually similar. Differentiating be-
tween data with low inter-class variance is an interesting
categorization challenge, requiring visual models that learn
the fine-grained differences between varieties of the same
crop.
Transfer Learning. There are a variety of interesting
transfer learning challenges of utmost importance to the
agricultural and plant science communities, including dis-
covering approaches that generalize across sensors, across
crops, or across environmental conditions. The TERRA-
REF data additionally presents an opportunity to help solve
the greenhouse-to-field gap, where models that perform
well in greenhouse conditions tend to not generalize to field
conditions; because the TERRA-REF data includes both
greenhouse and field data for the exact same varieties, re-
searchers in transfer learning could help build models that
bridge this gap.
Multi-sensor Integration. The TERRA-REF data in-
cludes data captured from a variety of visual sensors (de-
scribed in Section 1.2). These sensors have similar, but notTable 1. Summary of TERRA-REF sensor instruments.
Sensor Name Model Technical Specifications
Imaging Sensors
Stereo RGB Camera Allied Vision Prosilica GT3300C
Laser Scanner Custom Fraunhofer 3D Spatial Resolution: 0.3 to 0.9 mm
Thermal Infrared FLIR A615 Thermal Sensitivity: ≤50mK @ 30◦C
PS II Camera LemnaTec PS II Fluorescence Prototype Illumination 635nm x 4000 µmol/m2/s, Camera 50 fps
Multi-spectral Radiometers
Dedicated NDVI Multispectral Radiometer Skye Instruments SKR 1860D/A 650 nm, 800 nm ±5 nm; 1 down, 1 up
Dedicated PRI Multispectral Radiometer Skye Instruments SKR 1860ND/A 531nm +/- 3nm; PRI = Photochemical Reflectance Index
Active Reflectance Holland Scientific Crop Circle ACS-430 670 nm, 730 nm, 780 nm
Hyper-spectral Cameras
VNIR Hyperspectral Imager Headwall Inspector VNIR 380-1000 nm @ 2/3 nm resolution
SWIR Hyperspectral Imager Headwall Inspector SWIR 900-2500 nm @ 12 nm resolution
Environmental Sensors
Climate Sensors Thies Clima 4.9200.00.000
VNIR Spectroradiometer Ocean Optics STS-Vis Range: 337-824 nm @ 1/2 nm
VNIR+SWIR Spectroradiometer Spectral Evolution PSR+3500 Range 800-2500nm @3-8 nm; Installed 2018
PAR Sensor Quantum SQ–300 Spectral Range 410 to 655 nm
Table 2. Summary of the sensor data products included in the first release of TERRA-REF data.
Data Product Sensor Algorithm File Format Plot Clip Full Field
Environment Thies Clima envlog2netcdf netcdf NA NA
Thermal Image FLIR ir geotiff geotiff +
Point Cloud Fraunhofer Laser 3D laser3d las las +
Point Cloud Fraunhofer Laser 3D scanner3DTop ply
Images Time-Series PSII Camera ps2png png
Color Images RGB Stereo bin2tiff geotiff + +
Plant Mask RGB Stereo rgb mask geotiff x
identical, viewpoints from within the gantry box, may not
have captured data at the exact same time, and may have
captured different perspectives of the same part of the field
on different days. This presents interesting challenges in
terms of how to incorporate information across the various
sensors, and how to work with time-series data that is not
necessarily well-aligned or continuously captured.
Explainable Models. All too often in machine learning
research, datasets and models are built solely to drive the
development of machine learning algorithms. When build-
ing models to answer questions like “should I cut this plant
down because it won’t produce sufficient yield?” or “is this
plant under environmental stress?,” it is important not just
to have maximally accurate models but to also understand
why the models make the determinations that they make.
This makes the TERRA-REF data, and the biologically rel-
evant questions it supports, an excellent opportunity to drive
development of new approaches for explainable machine
learning, conveying the decisions made by algorithms to
non-machine learning experts.
Information Content. The TERRA-REF field scanner
and sensors represent a substantial investment, and it is still
not clear which sensors, sensor configurations, and spatialand temporal resolutions are useful to answer a particular
question. Presently, much less expensive sensors and sens-
ing platforms are available [11, 1]. What do we gain from
the 1mm spatial resolution on this platform relative to unoc-
cupied aerial systems (UAS) that are quickly approaching
1cm spatial resolution? Or, which subset of hyperspectral
wavelengths provide the most useful information? Can we
predict the useful parts of a hyperspectral image from RGB
images? Or get most of the information from a multispec-
tral camera with a half-dozen bands? At the outset, the team
recognized that this configuration would be oversampling
the plant subjects, but it wasn’t clear what the appropriate
resolutions or most useful sensors would be.
Overall Challenges. Within all of these topic areas in
computer vision and machine learning, the challenges that
must be addressed require addressing interesting questions
such as determining the most appropriate sensors and data
processing choices for specific questions, addressing dif-
ficult domain transfer issues, considering how to integrate
noisy side channels of information, such as genetic infor-
mation that may conflict or conflate with each other, or deal-
ing with nuisance parameters like environmental or weather
variations that simultaneously influence plant subjects and
the sensor data content.Figure 3. Example data from the TERRA-REF gantry system. (top left) RGB data (center top) RGB data with soil masked (top right)
close up of RGB data. 3D-scanner data (center row, right to left) depth, reflectance, surface normals, and point cloud data produced by
the 3D scanner. (bottom row, right to left) FLIR thermal image with transpiring leaves shown as cooler than soil; F v/Fmderived from
active fluorescence PSII sensor providing a measure of photosynthetic efficiency; the reflectance of light at 543nm wavelength measured
by the VNIR hyperspectral camera. Because these are two-dimensional representations of three-dimensional systems, all scale bars are
approximate.
3. Algorithm Development.
The process of converting raw sensor outputs into usable
data products required geometric, radiometric, and geospa-
tial calibration. In this regard, each sensor presented its own
challenges. Combining these steps into an automated com-
puting pipeline also represented a substantial effort that is
described by Burnette et al. [3].
Radiometric calibration was particularly challenging,
owing that many images contain both sunlit and shaded ar-
eas. In the case of hyperspectral images, the white sensor
box and scans spread out over multiple days confounded
an already challenging problem. Radiometric calibration
of images taken by the two hyperspectral cameras exempli-
fies these challenges, and a robust solution is described by
Sagan et al. [21] and implemented in [19]. Even process-
ing images from an RGB camera was challenging due to
fixed settings resulting in high variability in quality and ex-
posure, requiring the novel approach described by Li et al.
[18]. Herritt et al. [14, 13] demonstrate and provide soft-
ware used in analysis of a sequence of images that capture
plant fluorescence response to a pulse of light.Most of the algorithms used to generate data products
have not been published as papers but are made available
on GitHub ( https://github.com/terraref ); code
used to release the data publication in 2020 is available on
Zenodo [25, 15, 10, 6, 4, 19, 8, 7, 5, 9, 17].
Pipeline development continues to support ongoing use
of the field scanner as well as more general applica-
tions in plant sensing pipelines. Recent advances have
improved pipeline scalability and modularity by adopt-
ing workflow tools and making use of heterogeneous
computing environments. The TERRA-REF computing
pipeline has been adapted and extended for continuing
use with the Field Scanner with the new name ”Phy-
toOracle” and is available at https://github.com/
LyonsLab/PhytoOracle . Related work generalizing
the pipeline for other phenomics applications has been re-
leased under the name ”AgPipeline” https://github.
com/agpipeline with applications to aerial imaging de-
scribed by Schnaufer et al . [22]. All of these software
are made available with permissive open source licenses on
GitHub to enable access and community development.Figure 4. Summary of public sensor datasets from Seasons 4 and 6. Each dot represents the dates for which a particular data product is
available, and the size of the dot indicates the number of files available.
4. Uses to date.
TERRA-REF is being used in a variety of ways. For ex-
ample, hyperspectral images have been used to measure soil
moisture [2], but the potential to predict leaf chemical com-
position and biophysical traits related to photosynthesis and
water use are particularly promising based on prior work
[23, 24].
Plant Science and Computer Vision Research. A few
projects are developing curated datasets for specific ma-
chine learning challenges related to classification, objectrecognition, and prediction.
We currently know of at least three datasets curated
for CVPPA 2021. The Sorghum-100 dataset was cre-
ated to support development of algorithms that can clas-
sify sorghum varieties from RGB images from Ren et al.
[20]. Another set of RGB images curated for the Sorghum
Biomass Prediction Challenge on Kaggle was developed
with the goal of developing methods to predict end of sea-
son biomass from images taken of different sorghum geno-
types over the course of the growing season. Finally, RGB
images from the TERRA-REF field scanner in Maricopa
accounted for 250 of the 6000 1024x1024 pixel images inthe Global Wheat Head Dataset 2021 [12]. The goal of the
Global Wheat Challenge 2021 on AIcrowd is to develop an
algorithm that can identify wheat heads from a collection of
images from around the world that represent diverse fields
conditions, sensors, settings, varieties, and growth stages.
Most of the research applications to date have focused on
analysis of plot-level phenotypes and genomic data rather
than the full resolution sensor data.
5. Data Access
Public Domain Data. A curated subset of the TERRA-
REF data was released to the public domain in 2020 (Figure
4) [16]. These data are intended to be re-used and are acces-
sible as a combination of files and databases linked by spa-
tial, temporal, and genomic information. In addition to pro-
viding open access data, the entire computational pipeline is
open source, and we can assist academic users with access
to high-performance computing environments.
The total size of raw (Level 0) data generated by these
sensors is 60 TB. Combined, the Level 1 and Level 2 sen-
sor data products are 490 TB. This size could be substan-
tially reduced through compression and removal of dupli-
cate data. For example, the same images at the same resolu-
tion appear in the georeferenced Level 1 files, the full field
mosaics, and the plot-level clip.
Other Data Available. The complete TERRA-REF
dataset is not publicly available because of the effort and
cost of processing, reviewing, curating, describing, and
hosting the data. Instead, we focused on an initial public
release and plan to make new datasets available based on
need. Access to unpublished data can be requested from
the authors, and as data are curated they will be added to
subsequent versions of the public domain release ( https:
//terraref.org/data/access-data ).
In addition to hosting an archival copy of data on Dryad
[16], the documentation includes instructions for browsing
and accessing these data through a variety of online portals.
These portals provide access to web user interfaces as well
as databases, APIs, and R and Python clients. In some cases
it will be easier to access data through these portals using
web interfaces and software libraries.
The public domain data is archived on Dryad, with the
exception of the large sensor data files. The Dryad archive
provides a catalog of these files that can be accessed via
Globus or directly on the host computer at the National Cen-
ter for Supercomputing Applications.
6. Acknowledgements
The work presented herein was funded in part by the
Advanced Research Projects Agency-Energy (ARPA-E),U.S. Department of Energy, under Award Numbers DE-
AR0000598 and DE-AR0001101, and the National Science
Foundation, under Award Numbers 1835834 and 1835543.
References
[1] Jonathan A Atkinson, Robert J Jackson, Alison R Bentley,
Eric Ober, and Darren M Wells. Field Phenotyping for the
Future , pages 1–18. John Wiley & Sons, Ltd, Chichester,
UK, Nov. 2018. 3
[2] Ebrahim Babaeian, Paheding Sidike, Maria S New-
comb, Maitiniyazi Maimaitijiang, Scott A White, Jeffrey
Demieville, Richard W Ward, Morteza Sadeghi, David S
LeBauer, Scott B Jones, et al. A new optical remote sens-
ing technique for high-resolution mapping of soil moisture.
Frontiers in Big Data , 2:37, 2019. 5
[3] Maxwell Burnette, Rob Kooper, J D Maloney, Gareth S Ro-
hde, Jeffrey A Terstriep, Craig Willis, Noah Fahlgren, Todd
Mockler, Maria Newcomb, Vasit Sagan, Pedro Andrade-
Sanchez, Nadia Shakoor, Paheding Sidike, Rick Ward, and
David LeBauer. TERRA-REF data processing infrastructure.
InProceedings of the Practice and Experience on Advanced
Research Computing , page 27. ACM, July 2018. 4
[4] Max Burnette, David LeBauer, Solmaz Hajmohammadi,
ZongyangLi, Craig Willis, Wei Qin, Sidke Paheding, and JD
Maloney. terraref/extractors-multispectral: Season 6 Data
Publication (2019), Sept. 2019. 4
[5] Max Burnette, David LeBauer, Wei Qin, and Yan Liu.
terraref/extractors-metadata: Season 6 Data Publication
(2019), Sept. 2019. 4
[6] Max Burnette, David LeBauer, ZongyangLi, Wei Qin, Sol-
maz Hajmohammadi, Craig Willis, Sidke Paheding, and
Nick Heyek. terraref/extractors-stereo-rgb: Season 6 Data
Publication (2019), Sept. 2019. 4
[7] Max Burnette, Zongyang Li, Solmaz Hajmohammadi, David
LeBauer, Nick Heyek, and Craig Willis. terraref/extractors-
3dscanner: Season 6 Data Publication (2019), Sept. 2019.
4
[8] Max Burnette, Jerome Mao, David LeBauer, Charlie Zen-
der, and Harsh Agrawal. terraref/extractors-environmental:
Season 6 Data Publication (2019), Sept. 2019. 4
[9] Max Burnette, Craig Willis, Chris Schnaufer, David
LeBauer, Nick Heyek, Wei Qin, Solmaz Hajmohammadi,
and Kristina Riemer. terraref/terrautils: Season 6 Data Pub-
lication (2019), Sept. 2019. 4
[10] Max Burnette, Charlie Zender, JeromeMao, David LeBauer,
Rachel Shekar, Noah Fahlgren, Craig Willis, Henry Bu-
towsky, Xingchen Hong, ZongyangLi, Fengling Wang, Tin-
oDornbusch, JD Maloney, Wei Qin, Stuart Marshall, Abby
Stylianou, and Ting Li. terraref/computing-pipeline: Season
4 & 6 Data Publication (2019), Feb. 2020. 4
[11] Anna L Casto, Haley Schuhl, Jose C Tovar, Qi Wang, Re-
becca S Bart, Noah Fahlgren, and Malia A Gehan. Picturing
the future of food. Plant phenome j. , 4(1), Jan. 2021. 3
[12] Etienne David, Mario Serouart, Daniel Smith, Simon
Madec, Kaaviya Velumani, Shouyang Liu, Xu Wang, Fran-
cisco Pinto Espinosa, Shahameh Shafiee, Izzat S. A. Tahir,Hisashi Tsujimoto, Shuhei Nasuda, Bangyou Zheng, Norbert
Kichgessner, Helge Aasen, Andreas Hund, Pouria Sadhegi-
Tehran, Koichi Nagasawa, Goro Ishikawa, S ´ebastien Dan-
drifosse, Alexis Carlier, Benoit Mercatoris, Ken Kuroki,
Haozhou Wang, Masanori Ishii, Minhajul A. Badhon, Cur-
tis Pozniak, David Shaner LeBauer, Morten Lilimo, Jesse
Poland, Scott Chapman, Benoit de Solan, Fr ´ed´eric Baret, Ian
Stavness, and Wei Guo. Global wheat head dataset 2021:
more diversity to improve the benchmarking of wheat head
localization methods, 2021. 6
[13] Matthew T Herritt, Jacob C Long, Mike D Roybal, David C
Moller Jr, Todd C Mockler, Duke Pauli, and Alison L
Thompson. Flip: Fluorescence imaging pipeline for
field-based chlorophyll fluorescence images. SoftwareX ,
14:100685, 2021. 4
[14] Matthew T Herritt, Duke Pauli, Todd C Mockler, and Al-
ison L Thompson. Chlorophyll fluorescence imaging cap-
tures photochemical efficiency of grain sorghum (sorghum
bicolor) in a field setting. Plant methods , 16(1):1–13, 2020.
4
[15] David LeBauer, Nick Heyek, Rachel Shekar, Katrin Leinwe-
ber, JD Maloney, and Tino Dornbusch. terraref/reference-
data: Season 4 & 6 Data Publication (2019), Feb. 2020. 4
[16] David LeBauer, Burnette Maxwell, Jeffrey Demieville, Noah
Fahlgren, Andrew French, Roman Garnett, Zhenbin Hu,
Kimberly Huynh, Rob Kooper, Zongyang Li, Maitiniyazi
Maimaitijiang, Jerome Mao, Todd Mockler, Geoffrey Mor-
ris, Maria Newcomb, Michael Ottman, Philip Ozersky,
Sidike Paheding, Duke Pauli, Robert Pless, Wei Qin, Kristina
Riemer, Gareth Rohde, William Rooney, Vasit Sagan, Nadia
Shakoor, Abby Stylianou, Kelly Thorp, Richard Ward, Jef-
frey White, Craig Willis, and Charles Zender. Data from:
TERRA-REF, an open reference data set from high resolu-
tion genomics, phenomics, and imaging sensors, Aug. 2020.
2, 6
[17] David LeBauer, Craig Willis, Rachel Shekar, Max Burnette,
Ting Li, Scott Rohde, Yan Liu, JD Maloney, Noah Fahlgren,
Charlie Zender, Rob Kooper, Jerome Mao, Harsh Agrawal,
Xingchen Hong, Shannon Bradley, Samy Pess ´e, Katrin Lein-
weber, Justin Manzo, Jeff Terstriep, and Abby Stylianou. ter-
raref/documentation: Season 6 Data Publication (2019), Feb.
2020. 4
[18] Zongyang Li, Abby Stylianou, and Robert Pless. Learning to
correct for bad camera settings in large scale plant monitor-
ing. In Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR) Workshops ,
June 2019. 4
[19] Jerome Mao, Max Burnette, Henry Butowsky, Charlie Zen-
der, David LeBauer, and Sidke Paheding. terraref/extractors-
hyperspectral: Season 6 Data Publication (2019), Sept. 2019.
4
[20] Chao Ren, Justin Dulay, Gregory Rolwes, Duke Pauli, Na-
dia Shakoor, and Abby Stylianou. Multi-resolution out-
lier pooling for sorghum classification. In Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops , pages 2931–2939, June
2021. 5[21] Vasit Sagan, Maitiniyazi Maimaitijiang, Sidke Paheding,
Sourav Bhadra, Nichole Gosselin, Max Burnette, Jeffrey
Demieville, Sean Hartling, David LeBauer, Maria New-
comb, Duke Pauli, Kyle T. Peterson, Nadia Shakoor, Abby
Stylianous, Charlie Zender, and Todd C. Mockler. Data-
driven artificial intelligence for calibration of hyperspectral
big data. Transactions on Geoscience and Remote Sensing ,
2021 . 4
[22] Christophe Schnaufer, Julian L Pistorius, and David S
LeBauer. An open, scalable, and flexible framework for au-
tomated aerial measurement of field experiments. In Pro-
ceedings of SPIE , 2020. 4
[23] Shawn P Serbin, Dylan N Dillaway, Eric L Kruger, and
Philip A Townsend. Leaf optical properties reflect variation
in photosynthetic metabolism and its sensitivity to tempera-
ture. J. Exp. Bot. , 63(1):489–502, Jan. 2012. 5
[24] Shawn P Serbin, Aditya Singh, Ankur R Desai, Sean G
Dubois, Andrew D Jablonski, Clayton C Kingdon, Eric L
Kruger, and Philip A Townsend. Remotely estimating pho-
tosynthetic capacity, and its response to temperature, in veg-
etation canopies using imaging spectroscopy. Remote Sens.
Environ. , 167:78–87, Sept. 2015. 5
[25] Craig Willis, David LeBauer, Max Burnette, and Rachel
Shekar. terraref/sensor-metadata: Season 4 & 6 Data Pub-
lication (2019), Feb. 2020. 4