Discussion and Future Perspectives

Discussion and Future Perspectives #

In this chapter, we review how we achieved physical reservoir computing (PRC) with plants, summarise the main achievements that led to this result and provide future perspectives on possible advancement studies at the end of this chapter.

Overview of the Main Results #

Since the conceptualisation of reservoir computing around the turn of the century and its transfer to physical systems, PRC has grown in scope and popularity as an alternative to conventional computing (Citation: & , ) Nakajima, K. & Fischer, I. (). Reservoir Computing: Theory, Physical Implementations, and Applications. Springer Singapore. https://doi.org/10.1007/978-981-13-1687-6 . Some of these new technologies promise to drastically reduce power consumption or improve performance for machine learning tasks (Citation: , & al., , , , , & (). Reservoir computing using dynamic memristors for temporal information processing. Nature Communications, 8(1). 2204. https://doi.org/10.1038/s41467-017-02337-y ; Citation: , ) (). Scaling up integrated photonic reservoirs towards low-power high-bandwidth computing  (dissertation). Ghent University Retrieved from http://hdl.handle.net/1854/LU-8666053 . Another subdomain of reservoir computing is focussing more on embodiment and outsourcing control to the body. This technique promises to reduce the complexity of the microcontroller and improve robustness in compliant robotics (Citation: , & al., , , , & (). Terrain Classification for a Quadruped Robot. https://doi.org/10.1109/ICMLA.2013.39 ; Citation: , (). Physical reservoir computing—an introductory perspective. Japanese Journal of Applied Physics, 59(6). 060501. https://doi.org/10.35848/1347-4065/ab8d4f ; Citation: , & al., ) , , , , , & (). Design and control of compliant tensegrity robots through simulation and hardware validation. Journal of The Royal Society Interface, 11(98). 20140520. https://doi.org/10.1098/rsif.2014.0520 .

In this work, we aim to introduce physical reservoir computing to an unexplored field: plant ecophysiology. While plants do not meet all the criteria for reservoir computing in a strict sense (Mapping the Physical Reservoir Framework to Plants), it can form the basis of a paradigm shift in plant ecophysiology. Instead of focussing on specific traits and their effect on physiology, a more holistic approach can yield interesting new insights into plant behaviour.

In a first study (chapter 5) where we investigated if plants can be considered living reservoirs, we used a hyperspectral camera to monitor the plant’s state. This study was unsuccessful; data from background materials and plants were equally good at predicting the physiological tasks. We suspect that the root causes were insufficient accuracy of the camera and limited spectral changes. Improved sensor technology might solve some of the issues, but specific sensory equipment appears more interesting for several reasons. Firstly, hyperspectral cameras are expensive sensors compared to conventional red green blue (RGB) cameras. Secondly, they produce vast amounts of data if one wants to study dynamic plant behaviour. Storage and processing of this data can be a challenge. Thirdly, hyperspectral data is complex. There are many bands available, and the spectral resolution might not be uniform. Moreover, extracting the relevant features from the data is an ongoing research topic. In summary, we advise avoiding hyperspectral cameras to study the dynamics or subtle variations of plants. However, based on many examples from literature, they can be very effective at capturing strong dynamics due to drought and diseases (Citation: , (). Plant Disease Detection by Imaging Sensors – Parallels and Specific Demands for Precision Agriculture and Plant Phenotyping. Plant Disease, 100(2). 241–251. https://doi.org/10.1094/PDIS-03-15-0340-FE ; Citation: , & al., ) , & (). Hyperspectral image analysis techniques for the detection and classification of the early onset of plant disease and stress. Plant Methods, 13(1). 80. https://doi.org/10.1186/s13007-017-0233-z .

Changing the sensory system to a contact-based system was an appropriate solution because contact sensors offer a more direct measurement and generate much less data, which in turn simplifies the analysis. However, most commercially available readout systems for phenotyping did not meet our requirements in terms of accuracy, flexibility and cost. Consequently, we designed a custom solution: Gloxinia. The design of this sensor platform is discussed in chapter 6. We successfully demonstrated its operation in that chapter and in chapter 7.

Using this custom sensory system, we demonstrate PRC-inspired computing with plants in chapter 7. We show how leaf thickness measurements of strawberry plants were used to assess ecophysiological, environmental and benchmark regression targets. Our results indicate that plants are unsuited for general-purpose computing, yet instead are highly relevant for plant-related tasks. Photosynthetic rate and transpiration rate are the two main biological tasks investigated.

Chapter 5 and chapter 6 illustrate the steps taken leading up to the results in chapter 7, demonstrating PRC with plants. Although the results in chapter 5 were inconclusive, they are highly relevant for the phenotyping community. Using our findings, we illustrate some of the limitations of current hyperspectral technologies. Moreover, we also illustrate that despite some tasks being performant, the plant was not the root cause of said performance. This highlights that it is always important to study the effect of the environment and sensory system. As a result, the PRC study from chapter 7 also included a control experiment to ensure computation arose due to plant dynamics and not the dynamics of the environment. PRC with plants is currently in the start-up and exploration phase. We suspect increased interest from the plant science community and generalisation to more plant species can transform PRC with plants to a wide field of research.

Gradual Improvements To the Experimental Setup #

The two main studies presented in chapter 5 and chapter 7 are portrayed without any of the obstacles and problems we experienced. In this section, we provide advice for future researchers working on PRC with plants.

Plants are living organisms and have clear preferences in terms of environmental conditions (Citation: , & al., , & (). Blossom Drop, Reduced Fruit Set, and Post-Pollination Disorders in Tomato. University of Florida, Institute of Food and Agricultural Sciences, Electronic Data Information Source. ; Citation: , & al., , , & (). Improving the CROPGRO-Tomato Model for Predicting Growth and Yield Response to Temperature. HortScience, 47(8). 1038–1049. https://doi.org/10.21273/HORTSCI.47.8.1038 ; Citation: , (). Development and dry matter distribution in glasshouse tomato: a quantitative approach. Wageningen University and Research. ; Citation: , & al., ) , & (). Interaction of temperature and co2 enrichment on soybean: growth and dry matter partitioning. Canadian Journal of Plant Science, 67(1). 59–67. https://doi.org/10.4141/cjps87-007 (see also Mapping the Physical Reservoir Framework to Plants). The variability they experience from germination also determines (partially) their ability to cope with variability in a more developed stage (Citation: , ) (). Plants and Microclimate: A Quantitative Approach to Environmental Plant Physiology (3). Cambridge University Press. https://doi.org/10.1017/CBO9780511845727 . We experienced major problems when transferring plants from the greenhouse to the growth chamber. Conditions inside the greenhouse are much more uniform, especially in terms of temperature and relative humidity. Since these two factors are modulated in the growth chamber for both studies (chapter 5 and chapter 7), this presented a major challenge to prevent the plants from experiencing stress. This stress was not due to extreme environmental conditions but due to slow acclimation. This was especially true for tomato plants. As a result, we switched growing plants from the greenhouse to another (large) growth chamber. In this growth chamber, plants were subject to similar differences as those during the experiments.

For strawberry and bean, this solved the problem of a stress build-up from the start to the end of the PRC experiments. However, tomato plants still experienced significant stress towards the end of each experiment. This is characterised by deformed and smaller leaves, as illustrated in figure 7.1.

A tomato plant at the start (a) and end of the experiment (b).
Figure 7.1: A tomato plant at the start (a) and end of the experiment (b).

Initially, we worked with short experiments when working with the hyperspectral camera. The idea was to first emulate PRC research on computing with a soft silicone arm (figure 4.1) (Citation: , & al., ) , , & (). Information processing via physical soft body. Scientific Reports, 5. 10487. https://doi.org/10.1038/srep10487 . Here, the input consists of various sinusoidal signals. This generates a semi-random signal that excites the reservoir. However, when attempting to do something similar with the environmental drivers (air temperature, relative humidity and light intensity), plants were often stressed at the end of the experiments and/or responses were divergent. Their internal circadian rhythm persisted because of internal signals (Citation: & , ) & (). Plant Physiology, Fifth Edition (Fifth edition). Sinauer Associates, Inc.. . It is thus important to either have a warm-up phase prior to the experiment where plants are acclimatised to such conditions or to change the experimental setup such that the circadian rhythm is preserved. We opted for the second option because it is more relevant from a biological perspective and for future applications.

Future Work and Applications #

From the results presented in chapters 5, chapter 6 and chapter 7, we can identify five future research lines: (a) studying the spectral changes of plants in greater detail; (b) further development of the Gloxinia platform towards more types of sensor technologies supported and more robust operation; (c) refining results of PRC with plants; (d) investigating a closed-loop setup; and (e) working towards applications.

Dynamic Plant Spectrum The research presented in chapter 5 features some of the limitations of hyperspectral cameras. Nevertheless, they are popular sensing devices in plant phenotyping. If this research wishes to transition towards capturing more subtle variation in plants, more research is needed to understand the timescales and variation one can expect in such setups. To that end, we envision a small-scale study where various plants are extensively monitored using high-resolution point or line sensors. Results from such experiments can be used to gain insight into the possible variation observable with state-of-the-art sensor technologies. Based on these results, large-scale experiments can be optimised to gain more insight into a plant’s phenotype and performance in agricultural systems.

Expansion of Gloxinia For research purposes, the Gloxinia platform is sufficiently capable. However, its robustness should be improved such that it can run unsupervised for extended periods of time. Currently, this is not possible because there are insufficient recovery mechanisms built-in when a sensor readout temporarily fails. Moreover, using the system requires extensive knowledge; if we want it to be employed by a wider audience in the plant science community, it should be more user-friendly. To this end, a large new iteration of the interconnectivity is required, for instance one could rely on USB-C type connectors to connect sensors to the measurement device. Moreover, an enclosure is also needed. If designed well, end-users should just have to plug-in the sensors and register them using a software tool. These improvements should make it possible for non-experienced users to use Gloxinia without learning the details of the system.

Refinement of PRC with Plants The results presented in chapter 7 are promising but can still be expanded further. Better sensor technology and calibration can likely reduce unwanted effects due to the sensor-environment interaction and improve signal extraction. Alternative sensor systems such as volatile organic compounds (Citation: , & al., ) , & (). Wearable Sensors for On-Leaf Monitoring of Volatile Organic Compounds Emissions from Plants. https://doi.org/10.1109/NEMS50311.2020.9265575 , biopotential (Citation: , & al., ) , , , , , , & (). A special pair of phytohormones controls excitability, slow closure, and external stomach formation in the Venus flytrap. Proceedings of the National Academy of Sciences, 108(37). 15492–15497. https://doi.org/10.1073/pnas.1112535108 , sap flow (Citation: & , ) & (). Sapflow+: a four-needle heat-pulse sap flow sensor enabling nonempirical sap flux density and water content measurements. New Phytologist, 196(1). 306–317. https://doi.org/10.1111/j.1469-8137.2012.04237.x or leaf length (Citation: , & al., ) , , , , , , , & (). Leaf elongation response to blue light is mediated by stomatal-induced variations in plant transpiration in Festuca arundinacea. Journal of Experimental Botany(eraa585). https://doi.org/10.1093/jxb/eraa585 might be better suited for certain tasks than leaf thickness. These refinements can be used to gain a better understanding of the types of tasks and sensors that match. Furthermore, we need to evaluate the generality of the framework among plant species. Therefore, the PRC framework should be applied to multiple plant species for specific regression or classification tasks to extend the results of chapter 7 to more species. Moreover, those results should also be expanded to detect stress. For instance, one could investigate the quantification of plant drought and heat stress in response to reduced water availability. To assess this stress, a quantifiable stress trait needs to be defined. For example, one could assess the biomass accumulation based on a digital twin and compare this with experimental data.

The experiments presented in chapter 7 are limited in scope for the quantification of the nonlinear and memory properties of plants due to the lack of additional theoretical analysis and benchmark experiments. We assumed matching timescales due to the choice of leaf thickness observation and weather-data based input pattern generation. However, additional test are needed that validate this match when we increase or decrease the timescale of the input. For instance, the light intensity variations can be increased and as the frequency of change increases there should be less information captured by the leaf thickness observation. Moreover, the plant-relevant tasks have limited time-dependence. Tasks with more extensive time dependencies should be evaluated. Moreover, the underlying processes that generate the memory and nonlinear properties of the plant reservoir should be be investigated too.

So far, we have focussed exclusively on the part of the plant that is above the soil. Yet, the reservoir also consists of plant roots. These roots are a large and vital part of the plant. While there is increased interest in studying the development and functioning of root systems, many aspects remain unknown due to the inaccessibility of roots. Mounting contact sensors requires digging up part of the roots. Consequently, one might unintentionally alter them and disturb their surroundings. Recent techniques such as electrical impedance spectroscopy, spectral induced polarisation and electrical resistivity tomography enable frequency-dependent characterisation of roots without the need to disturb them. Such techniques quantify the redistribution of water in the soil, which can be used as a proxy for root activity (Citation: , & al., , , , , , , , & (). Sensing the electrical properties of roots: A review. Vadose Zone Journal, 19(1). e20082. https://doi.org/https://doi.org/10.1002/vzj2.20082 ; Citation: & , ) & (). Root architecture and hydraulics converge for acclimation to changing water availability. Nature Plants, 6(7). 744–749. https://doi.org/10.1038/s41477-020-0684-5 .

In research setups, leaf thickness clips are good enough, but mounting leaf thickness clips in plants is labour intensive, and they cannot remain on the plant for months because they alter the leaf physiology. Consequently, for real applications, there is a need to transition away from contact-based technology towards contact-less technologies. To this end, image sensors are most attractive. However, care must be taken in selecting appropriate features and optimising the experimental design. Our work in chapter 5 showed that subtle spectral variations are challenging. An easily identifiable alternative could be plant movement. The leaves of beans, lettuce and Arabidopsis plants feature extensive movement (Citation: , & al., , , , & (). TRiP: Tracking Rhythms in Plants, an automated leaf movement analysis program for circadian period estimation. Plant Methods, 11(1). 33. https://doi.org/10.1186/s13007-015-0075-5 ; Citation: , & al., , , , & (). Leaf-Movement-Based Growth Prediction Model Using Optical Flow Analysis and Machine Learning in Plant Factory. Frontiers in Plant Science, 10. https://doi.org/10.3389/fpls.2019.00227 ; Citation: , & al., ) , & (). Possible Involvement of Phototropins in Leaf Movement of Kidney Bean in Response to Blue Light. Plant Physiology, 138(4). 1994–2004. https://doi.org/10.1104/pp.105.062026 . Monitoring leaf movement can be fully automated for large plant populations, thus enabling PRC to scale beyond plant research.

Closing the Loop The PRC framework includes the integration of output feedback loops. Within the framework, one is thus not limited to detection but can also realise control on the input. For example, Citation: , & al. () , & (). Morphological computation-based control of a modular, pneumatically driven, soft robotic arm. Advanced Robotics, 32(7). 375–385. https://doi.org/10.1080/01691864.2017.1402703 have shown that by exploiting the dynamics exhibited by a soft robot body, control can be reduced to simple linear regression. However, within the context of plants, we lack a supervisory training signal. This entails that we cannot rely on a classical supervised learning system: it is not possible to identify a correct growth path, multiple paths exist and exploring them all is not possible. Consequently, we need to rely on a global reward signal such as biomass increment, photosynthesis rate and stomatal conductance. Such a reword signal can be used to optimise the system without knowing the optimal values for each monitored variable. In neurobiology, reward-modulated Hebbian learning can alter the synapse weights driven by the correlation between a global reward signal, presynaptic activity, and the difference of the postsynaptic potential from its recent mean (Citation: & , ) & (). Operant matching is a generic outcome of synaptic plasticity based on the covariance between reward and neural activity. Proceedings of the National Academy of Sciences, 103(41). 15224–15229. https://doi.org/10.1073/pnas.0505220103 . Citation: , & al. ( , , & (). A Reward-Modulated Hebbian Learning Rule Can Explain Experimentally Observed Network Reorganization in a Brain Control Task. Journal of Neuroscience, 30(25). 8400–8410. https://doi.org/10.1523/JNEUROSCI.4284-09.2010 ; Citation: , & al. () , & (). Reward-Modulated Hebbian Plasticity as Leverage for Partially Embodied Control in Compliant Robotics. Frontiers in Neurorobotics, 9. https://doi.org/10.3389/fnbot.2015.00009 showed that a physical reservoir system can learn motor control tasks using an online reward-modulated Hebbian learning rule. Despite the lack of a supervised learning step, the system can learn motor control tasks with an instantaneous reward signal. Such a system can also be leveraged for plant-based control loops. As such, the plant dynamics can be exploited for controlling the environmental factors in greenhouses and indoor plant factories or even irrigation in the field. Figure 7.2 visualises a conventional control loop in figure 7.2a and a PRC-based controller in figure 7.2b. PRC shifts the conventional information processing for stress detection or growth control from an external computer to the plant.

Closed loop control using an external controller (i.e. a conventional setup) and a PRC controller.
Figure 7.2: Closed loop control using an external controller (i.e. a conventional setup) and a PRC controller. Instead of relying on an external controller that has to interpret the signal and select the optimal actuation, the sensors directly drive the actuation using the physical medium as a controller. The advantages of this setup are easier control and more timely response to deviation due to stress in the plants.

Applications In ecophysiological experiments, one typically measures only one or two plant variables to determine the plant’s physiological state (Citation: , & al., ) , , & (). Plant Phenomics, From Sensors to Knowledge. Current Biology, 27(15). R770–R783. https://doi.org/10.1016/j.cub.2017.05.055 . Whenever the parameters move outside predefined bounds, a limited number of environmental drivers (e.g., watering and temperature) are modulated to influence this state. This process completely ignores the complex dynamic interplay between the plant and the environment due to the inability to relate small changes in plant variables to environmental changes. Indeed, a plant’s responses are the result of current and previous environmental conditions. Since there is no one-to-one relationship, it is unclear what the cause of small changes is. However, with the PRC framework, we can interpret experiments in which multiple environmental factors continuously vary due to the characterisation of the plant’s state using multiple sensors at high time resolution. The sensor-agnostic framework facilitates a general approach to control problems. No single measurement will be linked to a single ecophysiological variable. Instead, PRC uses the aggregate of the various measurements to obtain a fuller picture of the plant’s physiological state. Consequently, PRC with plants provides an entirely new way of looking at plant responses at a much more integrated scale.

This integrated view offers the possibility to quantify and analyse plant responses from a completely new point of view. Plants and by extension ecological systems are all (non-stationary) nonlinear dynamic systems. These systems receive internal and environmental inputs and optimise their responses accordingly. This is highly analogous to how PRC works for stationary systems. PRC provides a means to quantify information processing by plants. If the right tasks are defined, this can yield the phenotyping community the ability to quantitatively study plant performance in a general way. As a result, we suspect that it might lead to the discovery of previously unknown relationships and traits due to this more holistic point of view.

Epilogue #

In this dissertation, we created a bridge between PRC and plant ecophysiology. We investigated the usage of hyperspectral cameras and (leaf) thickness sensors to monitor the plant’s state. By means of different regression tasks, we observed that plants are not fit for general-purpose computation but instead are prime candidates for plant-related computation such as characterisation of ecophysiological parameters. Moreover, we envision future applications in the field of plant ecophysiology, breeding and precision agriculture. Our aim is to inspire more researchers in the plant science community to investigate the computational properties of plants and leverage the advantages that the PRC framework brings to the field.