IEEE Robotics & Automation Magazine - December 2023 - 66

Figure 14. However, and despite those incorrect
state detections, all the networks were able
to properly detect the stem; this feature can be
exploited to facilitate network fine-tuning with
additional datasets, such as the second dataset
part considered herein.
In an effort to assess the effect of fine-tunEVEN
MODEST
ing on network performance, we retrained the
tuned YOLO v7 network using a small sample
of the newly obtained data, that is, 1,180 images,
with a ratio of 50% for each wet/dry state.
Owing to the fact of training on a small dataset,
the model was trained on a Tesla P100 GPU
for only 30 epochs, at an initial learning rate of
0.01. An SGD optimizer with a momentum of
0.937 and a weight decay of 0.0005 was used
for optimization. The obtained results (the
final entry of Table 8) demonstrate a considerable
improvement in detection since the accuracy increased
from 0.75 to 0.98, suggesting that the structure can now detect
the stem in difficult scenarios. The recall increased modestly,
to 0.57, yet the fine-tuned network was able to provide over
50% more accurate detections. These results demonstrate that
more extended fine-tuning, but not a full network retraining,
may, in fact, be sufficient for the methods developed herein to
generalize to other tree crop species as well; extending this
dataset is part of ongoing work.
DISCUSSION AND OUTLOOK
CONTRIBUTIONS AND KEY FINDINGS
In this work, we have presented what, to the best of our
knowledge, is the first robot-assisted system for SWP measurements.
The value of this system lies in its capability to
autonomously collect samples at scale by automating timeconsuming
(e.g., leaf collection) or error-prone processes
(e.g., identification of the status of the leaf xylem). These are
core building blocks toward enabling complete robot autonomy
in physical specimen sampling and transport in the field.
Notably, accurate SWP measurement is instrumental to
implement precision agriculture practices to conserve water.
Our work demonstrated how it is possible to employ learningbased
machine vision to help determine the dry/wet status of
TABLE 8. The inference results on the dataset
evaluating the generalization capacity.
NETWORK
YOLO v5 base
YOLO v6 base
YOLO v7 base
YOLO v5 tuned
YOLO v6 tuned
YOLO v7 tuned
YOLO v7 fine-tuned
PRECISION
0.38
0.43
0.59
0.38
0.56
0.75
0.98
IMPROVEMENTS
IN IRRIGATION
PRACTICES
CAN HAVE
HUGE IMPACTS,
ESPECIALLY IN
THE SEMIARID
SOUTHWESTERN
UNITED STATES.
the xylem. Results from testing with the static
benchtop pressure chamber demonstrated that
we can provide a stable, crisp, and high-resolution
and high-refresh-rate video feed that is annotated
in real time and automatically determines
the wetness status of the xylem. This alone can
be a major aid to reduce false positive xylem
status determination as well as human errors
and reading variability when compared to the
" look-through-the-hole " approach that is currently
the only way to manually observe the
wetness status of the xylem. Further, we believe
that the produced datasets can stimulate further
research on robotics, machine perception, and
agronomy and lend themselves to help train
human operators in determining SWP via the
pressure chamber method.
DIRECTIONS FOR FUTURE WORK
The project outcomes achieved thus far enable various opportunities
for future work. One key direction involves enhancing
the detection of xylem water expression by giving heavier
consideration to smoothing the transitional phase. Further, we
seek to extend the dataset and evaluate the models' performance
in other tree crops, such as olive and almond trees.
Preliminary findings from evaluating the generalization
capacity indicated that fine-tuning the networks, or fully
retraining with additional data, can help expand the method
to other tree crops. Two critical factors affecting the generalization
capacity (especially when using the manual pump-up
chamber) are ambient light variations and induced motion.
The former can be addressed by possibly adding a cover
around the exposed stem. The latter is a challenge inherent to
the pump-up chamber, which operates by manual upward/
downward strokes (like a pump), but image stabilization techniques
can be employed. Nonetheless, we would recommend
the use of the static benchtop chamber if possible (at the
expense of higher acquisition and operating costs) since it can
lead to consistent and high-quality footage.
During the leaf cutting experiments, we observed that
RECALL
0.34
0.32
0.47
0.34
0.35
0.49
0.57
66 IEEE ROBOTICS & AUTOMATION MAGAZINE DECEMBER 2023
some successful cuts were not accurate enough to be used for
SWP analysis (i.e., leaving a long enough stem to be able to
secure the leaf to the pressure chamber's gasket). One main
issue was that the front face of the end effector may push other
interconnected leaves and/or branches away; hence, the linear
approach may not always suffice. Ongoing work aims to
improve the end-effector design to a more minimal assembly
and, in addition, enhance the perception capabilities of the end
effector to determine when the stem is aligned with the cutting
plane to trigger the cutting action. Also relevant is extending
the vision pipeline to determine and localize leaves bagged
in reflective foil bags (which allows leaves collected in SWP
analysis to reach an internal equilibrium, thus providing more
accurate information regarding the whole tree's water stress).
Another aspect worth exploring is whether it is possible to
mount the pressure chamber directly on the robot to eliminate
IEEE Robotics & Automation Magazine - December 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2023
