Early Drought Plant Stress Detection
with Bi-Directional Long-Term Memory Networks
Haohan Li, Zhaozheng Yin, Paul Manley, Joel G. Burken, Nadia Shakoor, Noah Fahlgren, and Todd Mockler
Abstract
Early drought stress detection is a promising strategy that
enables us to move from a reactive to a more proactive ap-
proach to manage drought risks and impacts. In this work,
we apply for the first time the Bidirectional Long Short-Term
Memory (
BLSTM
) networks to
RGB
images for accurate drought
plant stress detection in the early stage. In addition, an
optimal data collection strategy (
ODCS
) is investigated to use
less time and manpower for the purpose of accurate early
drought stress condition detection. The proposed method is
validated on two independently collected
RGB
image data-
sets. In both datasets, the
BLSTM
method achieves competi-
tive classification performances compared to three other
deep learning methods. By using the proposed
ODCS
, our
method can use only
⅔
of the entire dataset to achieve 74.6
percent F-score for the patch sequence classification and
72.0 percent F-score for the image sequence classification.
Introduction
On a global basis, drought, in conjunction with high tempera-
ture and radiation, poses the most important environmental
constraint to plant survival and to crop productivity [1].
Agriculture is the major victim of drought in many regions
of the world. Because the usable water supply in the world
is limiting, the future food demand for rapidly increasing
population pressures will be further aggravating the effects of
drought [27], which calls for attention to advance research to
improve the breeding strategies of drought tolerant plants and
early drought stress detection approaches.
Related Work
The mechanism of drought tolerance in plants has been dis-
cussed at the molecular level [11]. There are three main mecha-
nisms in drought plants which reduce the crop yield: (i) re-
duced canopy absorption of photosynthetically active radiation,
(ii) decreased radiation-use efficiency, and (iii) reduced harvest
index [7]. However, the reproducibility of drought stress treat-
ments to these mechanisms is cumbersome, which has hindered
both traditional breeding efforts and modern genetic approaches
in the improvement of drought tolerance of crop plants [34].
In addition, the mechanistic basis underlying drought toler-
ance is complex as it is mainly contributed by related traits
that are mostly determined by polygenic inheritance [21]. In
recent years, by measuring the structural and functional status
of plants, phenomic approaches may overcome the limited pre-
dictability. However, the lack of high throughput phenomic data
has been labeled as the “phenomic bottleneck” [20].
In the past years, as imaging systems and image analysis
techniques are developing, hyperspectral cameras have been
widely used in plant science research, such as monitoring
the growing condition of crops. In hyperspectral imaging,
the measured radiative properties of plant leaves or canopies
can be used to determine structural and physiological traits
of vegetation [16][30], for instance, a low reflectivity in the
visible part of the spectrum can be used to characterize the
spectral reflectance as a strong absorption by photosynthetic
pigments, whereas a high reflectivity in the near infrared is
produced by a high scattering of light by the mesophyll tis-
sues of the leaf. In addition, the reflectivity in the shortwave
infrared part of the spectrum is determined by the water,
protein, cellulose, and lignin content of plant tissues [19].
However, the spectral reflectance is a combination of multiple
physiological traits. Despite several laboratory studies that
have shown a relationship between the amount of water in
the leaf and the reflectance intensity in the short infrared part
of the spectrum, the determination of the water content pres-
ents some difficulties, due to the large reflectance variation
among leaves with the same water status. The most challeng-
ing issue in estimating the water content using the spectral
reflectance information is the decoupling of the contributions
of water content and other physiological traits.
Remote sensing has been successfully used in the precision
agriculture for providing the timely crop condition informa-
tion during growing seasons. In the optical region, the vegeta-
tion indices (
VIs
) have been used to detect crop conditions,
such as the water content and the nitrogen status. Most ap-
proaches are aiming at quantifying plant traits by calculating
VIs
that quantify specific plant structural changes [9]. Although
VIs
have been widely used to detect multiple crop growing
stresses in the advanced stage, such as the leaf nitrogen and
the chlorophyll content [29][10], the crop biomass [28] and the
vegetation moisture content [35], the use of
VIs
for the early
drought stress detection is still challenging, because different
crop stresses have similar
VI
computations in the early stage
[21]. Furthermore, the high cost of the hyperspectral camera
system and its further maintenance is limiting the develop-
ment of drought stress detection approaches based on the
hyperspectral image analysis for the consumer applications.
Motivation and Contribution
From the computer vision perspective, the image analysis
based drought stress detection can be defined as the classifica-
tion of images containing drought plants or not. Most previous
Haohan Li and Zhaosheng Yin are with the Department
of Computer Science, Missouri University of Science and
Technology, Engineering Research Lab Building, Rolla,
Missouri 65409 (
).
Paul Manley and Joel G. Burken are with the Department
of Civil, Architectural, and Environmental Engineering,
Missouri University of Science and Technology, Butler-
Carlton Hall, Rolla, Missouri 65409.
Nadia Shakoor, Noah Fahigren, and Todd Mockler are with
the Donald Danforth Plant Science Center, 975 North Warson
Road, St. Louis, Missouri 63132.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 7, July 2018, pp. 459–468.
0099-1112/18/459–468
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.7.459
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2018
459
