An Energy-Efficient PAR-Based Horticultural Lighting System for Greenhouse Cultivation of Lettuce

This paper presents an intelligent horticulture lighting and monitoring system to achieve energy-efficient supplemental lighting while maintaining the light quality and intensity at desired levels in the photosynthesis spectrum. Energy-efficiency is achieved through delivering only the required net light intensity, consisting of sunlight and supplemental LED light, using an intelligent controller that does not depend on the lighting system model. To this end, an online neural-network learning control system is developed, comprised of low-cost light sensors for measuring the photosynthetic photon flux density (PPFD), dimmable LED light fixtures, cameras, and internet-of-things (IoT)-enabled firmware used for crop monitoring and performance evaluation. Experiments performed in a research greenhouse facility on the lettuce crop are presented which indicate that the system can deliver the desired Daily Light Integrals (DLIs) to the plants in the presence of changing daylight conditions. The proposed method can thus deliver the exact amount of light to a specific crop based on the required light recipes during different growth phases. The control performance is further compared with a conventional on-off time-scheduling method in terms of plant health, growth, and energy requirements. The experiments indicate that the proposed solution can reduce energy consumption per unit dry mass of lettuce by 28% when compared to existing time-scheduling methods.


I. INTRODUCTION
Supplemental lighting is used to increase the daily light exposure of crops in greenhouses located in northern latitudes. To improve plant growth, yield, and quality while minimizing electricity consumption of artificial lighting, it is desirable to utilize the freely available sunlight as much as possible [1]. Thus, there is a great incentive to incorporate innovative technologies into today's greenhouse lighting automation systems. The emergence of horticultural light-emitting diodes The associate editor coordinating the review of this manuscript and approving it for publication was Kashif Sharif .
(LEDs) has facilitated this integration when compared to traditional horticulture light sources such as metal halide (MH) and high pressure sodium (HPS). Horticultural LEDs can further deliver dimmable light in the photosynthetic spectrum and can readily be incorporated into digital control systems, allowing for lighting schemes to be tailored to the plants' needs. Considering the plants' sensitivity to light, the photosynthetically active radiation (PAR) flux density is defined as the light energy in the 400-700nm spectrum, referred to as the photosynthesis band, measured in W m 2 s [2]. However, photosynthesis is a quantum process that is more dependent on the received number of photons than their energy [3].
In relation to plant growth and morphology, PAR is measured in terms of the flux of photons per unit area, or Photosynthetic Photon Flux Density (PPFD), expressed in µmol m 2 s . Horticultural lighting has traditionally concentrated on issues such as yield quality, quantity, and experimental crop-specific light recipes involving light intensity, spectrum, and photoperiod. In this respect, intelligent methods for precise delivery of the required light recipe, while minimizing energy consumption, has received increasing attention in recent years. A discussion of pertinent literature and existing challenges is presented in the following along with the contributions of the present work.
Horticultural LED lighting and its effects on crop growth have been investigated by several authors [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]. Research on LED crop lighting dates back to the 1980's, with red-only lights used in the international space station, to high-density multi-spectral LED chip-on-board devices in recent years [4], [5]. An early work by Albright et al. [7] showed that dry mass accumulation is proportional to the daily light integral (DLI) and a consistent DLI is central to quality crop production. Later studies have shown that plant growth and development (e.g., flower-bud initiation, inter-node length, branching, leaf area) and crop production value (yield, vitamins, pigments) are directly affected by lighting [8], [9], [10], [11]. In [12], the authors provided an overview of the effects of LED lighting on the growth, yield, and nutritional quality of green leafy vegetables, fruits, and ornamental plants. In [13], the authors discussed how spectral quality of LEDs can dramatically affect crop anatomy, morphology, nutrient uptake, and pathogen development. In [14], an adjustable spectrum LED was presented to match the plant's spectrum based on the relative quantum efficiency [2]. In [15], the authors provided a review of research activities related to LED lighting and highlighted issues such as plant cultivar-dependent recipes and cost analysis.
The studies in [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], and [29] demonstrate how dynamic lighting and optimization methods can help to increase the energy savings achievable by LED systems. In [17], the effects of dynamic regulation of supplemental lighting intensity on electricity consumption and fresh weight buildup were examined. The study in [18], analyzed the growth of an ornamental crop under dynamic regulated LED lighting. The data from a quantum PAR sensor was used to incrementally increase/decrease the LED light intensity to within a 10% tolerance. The results indicated a 21% energy reduction. However, no difference in crop quality, or time to anthesis, was observed and the method is open-loop and not robust to parametric variations and disturbances. In [20], the authors examined lettuce cultivation using red and blue LEDs. The plant's growth was represented in a feed-forward network using an image segmentation method based on K-means clustering. The paper reported 40-52% reduction of energy and an increase the leaf area by up to 6%; however, it was done in a simulated scenario with no experimental results. In [21] and [22], the researchers investigated spectrum controlled and targeted LED lighting to achieve energy savings. Through targeted plant lighting and spectral optimization, a 50% reduction in energy usage per unit of dry biomass was achieved. However, the comparison for energy savings was with traditional inefficient horticultural lights such as HPS and MH technologies.
In [27], a herbal remedy plant was grown under natural light and under LED fixtures with varying light conditions. Simulation of multiple spectra showed that the method could replicate the intended spectrum with minimum fitness errors. However, the study did not report energy saving amounts. In [28], the LED light intensity was optimized over time for different electric energy and lettuce crop prices, but was restricted to simulations in a plant factory with no daylight harvesting. In [29], the authors presented existing energysaving methods using LEDs in plant factories. The plant canopy image was captured during different growth stages and appropriate LEDs above the plant were turned on to avoid light wastage. At the same time, the on-time and brightness levels of LEDs were calculated to produce the required light profile. This study focused on spatial configuration of the LEDs for energy saving but used an on-off method for intensity control which has limitations in terms of energy-savings and light quality as presented in this paper. In [20], the authors investigated lettuce cultivation in an indoor farming setup exposed to red and blue LED lights using an image segmentation method, based on K-means clustering, to identify and optimize the lighting schedule. However, this work did not have an energy harvesting focus through daylighting.
The above works have presented interesting results related to yield increase and energy saving potential of dynamic LED lighting schemes. However, they mostly suffer from certain issues in their sensing and control schemes. For instance, quantum PAR sensors are very costly to deploy in large numbers in a greenhouse environment. Meanwhile, primitive light sensors only allow for color ratio control and not PAR control. Furthermore, the control and optimization algorithms used in previous works are mostly very simple and highly dependent on the system model. In a horticultural system, both the light environment and daylight contribution change with plant growth and seasonal and weather conditions. Thus, model-based methods would not have acceptable performance since the environmental parameters change with time, leading to variations in the system model and hence deteriorated closed-loop control performance. Another drawback of model-based approaches is that they are not scalable since the system model changes with system size. Also, there is little evidence of research that utilizes both dynamic lighting for energy-efficiency and independent set-point light intensity regulation and control of color ratios.
The contribution of present work is addressing some of the above issues in horticultural lighting through utilizing machine learning-based light intensity sensing, using newly introduced low-cost spectral light sensors, combined with a learning neural network controller that does not depend VOLUME 11, 2023 on the mathematical model of the lighting system. To this end, a comparative study of machine learning methods for mapping the sensors readings to PPFDs is presented and experimentally tested in an IoT-based feedback control system. The controller can operate with minimal tuning to deliver canopy-level desired light recipes. The proposed system achieves cross-color-channel and cross-light-source illumination capability, leading to further improvement energy consumption and yield when compared to on-off time scheduling methods. A by-product of providing accurate light intensity in each spectrum would be to avoid plant health issues such as tip-burn due to excessive light exposure.
The rest of this paper is organized as follows. Section II formulates the lighting control problem. Section III introduces the proposed solution including the IoT control system platform, estimation of PAR measurements using low-cost spectral light sensors, and the learning neural controller. Section IV presents the experimental setup used to evaluate performance of the proposed system. Section V presents the results of conducted experiments in a greenhouse facility. Conclusions and future work are presented in section VI.

II. PROBLEM FORMULATION
Reducing the cost of electricity for supplemental lighting to promote photosynthesis is desirable from an economic perspective. Supplemental lighting in the blue and red spectrum is the preferred method to maximize photosynthetic activity and minimize energy usage. Since the red:blue ratio of light received by plants would directly affect plant growth, dimmable blue and red light fixtures have been developed to achieve desirable light recipes ( [30], [31]).
Using a two channel LED fixture to supplement light in the blue and red spectrum, it is necessary to have PAR sensors that can measure these colors separately, instead of the full photosynthetically active spectrum. Thus, we define two parameters for the photosynthesic photon flux density (PPFD) in the red and blue spectra, called PPFD r and PPFD b , respectively. These parameters are defined as the number of photons in the 600-700nm and 400-500nm, respectively, that reach a surface area per unit of time (measured in µmol m 2 s ). The Daily Light Integral (DLI) is often used in plant growth and morphology. Thus, let us define the corresponding DLI parameters as DLI r and DLI b , which are the daily flux of photons in the 400-500nm and 600-700nm range per unit area, respectively (measured in mol m 2 d ). The conventional PAR measures of in the full 400-700nm photosynthesis spectrum are denoted by PPFD full and DLI full throughout the rest of this paper.
The system model, consisting of n light fixtures with two color channels per fixture to control the PPFD levels of red and blue at n target points can be represented by a linear, static, time-invariant MIMO system in the following form where y(t) is the 2n × 1 output vector representing spectrum-specific PPFD readings at points of interest and is defined as in which i = 1, . . . , n indicates the number of each sensor placed at a target point, and the subscripts r, b stand for red and blue spectral readings, respectively.
The 2n × 1 control input vector, u(t), represents the dimming commands (0-100%) for the red and blue channels of LED fixtures as follows where i = 1, . . . , n indicates the number of each two-channel light fixture, and the subscripts r, b stand for red and blue channel input commands of each fixture, respectively. Matrix T (2n × 2n) is a full-rank system matrix and y L (2n × 1) is the vector of daylight contribution, modeled as an additive output disturbance. The system matrix T represents a cascade connection of three subsystems, namely, LED fixtures, LED drivers, and mapping between the emitted flux by light fixtures and light intensity at sensor locations from direct and reflected light rays [32].

III. SYSTEM COMPONENTS
In this section we present the proposed system which utilizes data from multiple light sensors positioned at crop levels to estimate the PAR received at the plant canopy and provides supplemental lighting through dimmable, multi-channel horticultural LED lights.

A. IoT PLATFORM
An IoT-enabled supplemental lighting control platform was designed and implemented consisting of the following components: • Light Sensing: The multi-channel spectral light sensor AS-7341 (from ams-OSRAM) was used to sense light intensity. The spectral response of this sensor is defined by 11 individual channels covering 350-1000nm, with 8 channels centered in the visible spectrum (VIS), plus one near-infrared (NIR), and a clear channel. The SS-110 spectroradiometer (from Apogee Instruments) was used as a reference module for PAR data collection and sensor calibration. The SS-110 continuously measures lighting within the wavelength range of 340-820nm. It can be utilized for measurement of spectral output (energy flux density, photon flux density, or illuminance) of different radiation sources.
• Supplemental Horticultural LED Lights: Wireless controlled LED fixtures (Q400 from QuantoTech Solutions, Ltd.) allow dimmable red and blue light color intensities. The LED unit delivers a maximum PPFD output of 480 µmol/s and consist of 7W blue LED (460nm), 14W red LED (660nm) and 2W UV LED (385nm). The wireless-enabled LED fixtures can be controlled through a local WiFi network using a RESTful API architecture. The WiFi module on the light fixture joins the local network with a self-defined IP address assigned automatically and allows for a controller connected to the same network to call for connection and transmit control commands.
• Image Sensor: A low-cost Raspberry Pi Camera Module v2 was used with an 8 mega-pixel camera based around the Sony IMX219 image sensor, which is capable of producing 3280 × 2464 pixel static images and a 640 × 480p90 video. With physical size of 25mm × 23mm × 9mm and optical size of 1/4 ′′ , the sensor is attached to, and controlled by, a Raspberry Pi unit by way of a dedicated Camera Serial Interface port (CSI).
• Controller: A Raspberry Pi 3 B+ was used with a BCM2837B0 System-on-Chip (SoC) and 1.4 GHz ARM Cortex-A53 processor running a Debian-based Linux distribution called Raspbian OS. This module is responsible for collecting sensor lighting data, using the machine learning trained PAR model, executing the control algorithm, and sending dimming commands to LED fixtures in real time. It is also in charge of setting image sensor configurations, timing and triggering captures, pre-processing of images, and authenticating and uploading the data to a cloud storage service for further analysis.
• Cloud storage: The Google service was used to store the image data along with their meta-data. The Raspberry Pi node device contains the code for using the Google Drive API to authenticate the Google account credentials and upload the data.
• Processing and Analytics: A desktop PC was used as the main processing unit for machine learning model fitting and image processing tasks to allow for powerful and cost-effective image processing and machine learning computations, and to verify performance of the lighting control scheme. The data was accessed on the cloud server, read into corresponding algorithms, and processed. The results were sent back to the Raspberry Pi for model/algorithm updates, uploading to the cloud, presentation on the PC, and feeding the ThingSpeak platform (from Mathworks, Inc).

B. LIGHT SENSOR CALIBRATION
Conversion of raw incident light data using the multi-spectral AS-7341 sensor to PPFD parameters was conducted using multiple machine learning algorithms including multi-linear regression, neural network, decision tree, and random forest. The details are not presented here due to space limitations. A very large data-set was created by collecting synchronized lighting data from AS-7341 (from amsOSRAM) along with the SS-110 spectroradiometer (from Apogee Instruments) throughout a year in SFU Surrey Campus, British Columbia, Canada (geo-location: 49.276765 -122.917957). Different weather conditions, geometrical configurations and supplemental lighting levels were considered in the data-set. Various linear and nonlinear regression models were used to fit the data. Although nonlinear regression models such as random forest model obtained more accurate results, the trade-off between simplicity and accuracy makes a multiple linear regression model the best choice for real-time control applications such as ours.

C. SUPPLEMENTAL LIGHTING CONTROLLER
As established in [31], a learning neural controller can harvest daylight and regulate light levels while eliminating the need for system identification, i.e., obtaining the system matrix T in (1). The adaptive nature of the control scheme also results in compensation of any uncertainty, or changes in the system model T, by the controller. Furthermore, the proposed method does not require any knowledge of the daylight term y L . The proposed neural network controller input presented in [31], is as follows where N ∈ R (2n+1)×(2n+1) represents the weight matrix for the neurons of a single layer neural network with the same number of nodes as outputs. The terms y da = y d 1 T is an augmented desired output vector with y d ∈ R (2n+1)×1 , representing the desired output with the term 1 being a unit firing threshold bias. Note that each neuron uses an identity linear activation function. The proposed update law for evolution of weights is as followṡ where ε is the output error defined as and k, η, ∥.∥ are positive scalar design parameters the Euclidian norm, respectively. As discussed in [31], the closed-loop system with the proposed neural network controller is uniformly ultimately bounded if the following conditions hold where ∥.∥ F denotes the Ferobinious norm and σ T represents the smallest singular value of T . Moreover, the ultimate bound on the output error can be made arbitrarily small using the design parameters k and η as long as conditions (7), (8), (9) are satisfied. The PPFD-based control system flowchart is provided in Fig. 1, in which the error signal to the neural network controller is obtained using the desired and actual PPFDs for red and blue channels. The desired PPFDs are obtained based on a target crop-dependent light recipe including daily light integral (DLI), desired red to blue color ratio (R:B), photo-period, and a light profile (e.g., a constant daily value based on the crop development stage). The light sensor uses trained machine-learning coefficients to obtain PPFD in the red and blue channels and feeds them to the error block. The controller then obtains the dimming commands and applies them to the LED red and blue channels.

IV. EXPERIMENT DESIGN
A comparison experiment was designed to evaluate the performance of the proposed lighting control system and analyze its effects on plant health and growth.

A. EXPERIMENTAL SETUP
A greenhouse compartment as shown in Fig. 2 was utilized in the Institute for Sustainable Horticulture (ISH), Kwantlen Polytechnic University (KPU), Langley, BC, Canada (geolocation: 49.109619242053256, -122.64535537470125). Two systems were installed side by side, each consisting of two light sensors, two LED fixtures, and an independent controller: System 1 (Fig. 2 on the right), ran a time-scheduling  on-off controller, and System 2 ( Fig. 2 on the left) ran the proposed neural network controller.

B. IRRIGATION AND NUTRIENTS
The two systems were built side-by-side on the same grow tray and shared the same hydroponic nutrition and water delivery system to keep nutritional conditions consistent. A drain to waste hydroponic system was utilized to provide water and nutrients to plants. The nutrient solution recipe, in parts per million (ppm), is shown in Table 1. Electrical Conductivity (EC) and pH of the nutrient solution were fixed at 2.0 mS/cm and 5.8, respectively. The temperature of nutrients was kept at approximately 19 • C. The ambient temperature of the greenhouse varied in the range 18-20 • C. The utilized irrigation schedule varied from a one-minute pulse twice a week, to two pulses of 1.5 minutes, every other day, using irrigation drippers with a capacity of 2 liters per hour.

C. CROP SELECTION
A green mini romaine organic lettuce variety, called Dragoon (LATIN NAME: Lactuca sativa), was selected. This variety is found, based on the producer's trials [33], to be a great performer in a climate-controlled greenhouse environment and suitable to be grown successfully using hydroponic growing methods, or other soil-less growing systems. The number of days to maturity for this variety is 43 for transplants, and 47 days for direct seeding, but can be sown every 2-3 weeks for a continuous supply of either full heads or baby leaves.

D. LIGHTING PLAN
An accurate lighting plan was created for the neural network system (System 2) to deliver the required DLI to the plant canopy according to a precise light recipe (red to blue PPFD ratio and photoperiod). Based on the required DLI, photoperiod, and red:blue ratio, and assuming a uniform daily light profile (constant PPFD throughout the day), the PPFD set-points for red and blue spectral ranges were obtained. Table 2, shows the obtained PPFD set-points based on provision of 15.12 mol m 2 d daily light integral through supplying constant red and blue PPFDs to the plants during a 12 hour period from 7AM to 7PM. The selected target DLI was designed to promote healthy plant growth while preventing health issues such as tipburn according to previous studies [34], [35]. A red to blue ratio of 4:3 was selected as the most suitable spectra for the selected crop based on previous works [36], [37]. The light plan for the time scheduled system (System 1) was selected to represent the traditional on-off method used by growers for comparison of energy consumption and plant health properties. Both the red and blue channels of LED fixtures were turned on to full power (7W blue channel (460nm), and 14W red channel (660nm)) with a 12 hour photoperiod between 7AM and 7PM. In this case, the light source red/blue ratio was 2:1.

E. LIGHTING DATA COLLECTION
Two light sensors corresponding to the two LED fixtures were utilized to collect lighting data in real time. The sensor mount heights were selected to provide sufficiently accurate measurement of light received by the plants but were not shaded by the canopy as the plants grew. The SS-110 spectroradiometer (from Apogee Instruments) was placed close to one of the sensors in System 2, which collected 340-820nm spectral data continuously to provide a reference for PPFD measurements by the machine-learning calibrated sensors.

F. IMAGE DATA COLLECTION
An image sensor was located on top of each plant canopy beside the LED fixtures to collect image data. Periodic top view images were captured four times a day at 8AM, 14PM, 18PM, and 23PM under white light conditions. The system automatically turned the red and blue channels of the fixtures off and the white light on, captured images, saved them to the database, and resumed providing supplemental lighting. The image data was later processed using the PlantCV open source software [38] to extract color and geometrical features and monitor plant health and growth.

A. PPFD TRACKING RESPONSE
The tracking responses of the proposed neural-network system when compared to the time-scheduling system     plant health and growth was improved due to the uniformity and consistency of light profile received throughout each day. In Tables 3 and 4, a comparative analysis of Systems 1 and 2 for the sample day of the growth cycle is presented in terms of error between the desired and measured PPFD at plant canopy along with the energy savings for red, blue, and full PPFD, respectively. Fig. 3.a, shows the tracking response for the average PPFD r of sensors in System 2 between 7AM and 7PM. Fig. 3.b displays the corresponding average dimming percentage for LED fixtures of System 2. It can be observed from Fig. 3.a, 3.b that, using the proposed controller, the 200 µmol m 2 s set-point for red PPFD is tracked with 0.3% error. To achieve this set-point, the red channels of LED fixtures were turned ON to a maximum value of 86% of their full power, resulting in an average of 32% energy savings on the red channel supplemental lighting for this day. Figs. 3.c, 3.d, illustrate the average PPFD r of the sensors and the red channel dimming percentage of LED fixtures for System 1. In this case, the panels are always 100% ON during the photoperiod with no savings in energy. Additionally, the red PPFD tracking error is at a very high value of 50%, resulting in the plant canopy to receive excessive red light which can result in plant health issues due to excessive light such as tipburn. Fig. 4.a demonstrates the tracking response for the average PPFD b of sensors in System 2 throughout the day while Fig. 4.b represents the average dimming percentage of its LEDs. Referring to Figs. 4.a, 4.b, the 150 µmol m 2 s set-point for blue PPFD is tracked with 0.4% error. To achieve this level of precision in the blue spectral range, the blue channels of LED fixtures were turned ON to a max of 87% of full power to save an average of 33% in energy on the blue channel supplemental lighting for this day. Figs. 4.c, 4.d, indicate the average PPFD b of sensors and the dimming percentages of blue channel for System 1. In this case, the panels were always ON at 100% power during the photoperiod, resulting in no energy savings. Additionally, the blue PPFD tracking error were very high at 49%, exposing the plant canopy to an excessive amount of blue light, which might cause plant health problems.   was regulated with 14% error. In comparison, the full PPFD tracking error for System 1 was a staggering 43%.

B. DLI TRACKING RESPONSE
A comparison of the received DLIs at plant canopy levels and the corresponding LED fixture dimming levels (representing the energy used) for Systems 1 and 2 is depicted in Figs. 6, 7, and 8 for red, blue, and full spectral ranges, respectively. The DLI is calculated from daily PPFD sensor readings by adding up all the collected samples as follows where PPFD r/b/f (t) denotes the red, blue or full PPFD sample collected at time t, assuming that one sample is collected every second. In Table 5, a comparison of mean absolute error (MAE), mean absolute percentage error (MAPE), mean squared error (MSE) and root mean squared error (RMSE) of achieved DLIs in Systems 1 and 2 is presented. Table 6 presents the average, minimum, and maximum of energy savings. Examining the MAPE column in Table 5 reveals that the proposed neural network controller can reduce the mean absolute percentage error from 61% to 14% for red, from 40% to 16% for blue, and from 44% to 19% for full PPFD, respectively. From Table 6, it can be concluded that this error reduction is achieved with an average energy savings of 32% and 27% in the red and blue channels, respectively. It is worthwhile noting that the above percentages have increased to 80% and 55%, on certain days, depending on the specific daylight and weather conditions. C. PLANT HEALTH AND YIELD COMPARISON Fig. 9 shows a general comparison of plants in both systems. Figs. 10a, 10b depict a comparison of some plant samples from a side and top view. Tipburn incidence is observed in plant samples from System 1 due to excessive supplemental lighting. Tipburn is a serious problem in lettuce production under artificial illumination, which occurs as a result of decreasing calcium concentrations in the leaves [39], [40].   This problem is avoided in System 2 by producing a uniform PPFD at plant canopy level.
About 40% of the crop were selected randomly as samples to conduct comparative measurements of fresh and dry  weights. The selected plant samples were cut from the stem and fresh weighs were recorded immediately. To record the dry weights, the YAMATO DX 402 Drying Oven was used and the samples were dried for 48 hours at 105 • C. From each plant sample, a leaf sample was collected, weighed, and scanned.
The results of weight measurements between the two systems are shown in Table 7, according to which the total fresh VOLUME 11, 2023  weight for plant samples of System 1 was 23% higher than that of samples of System 2. For the collected leaf samples, however, samples of System 2 have 26% higher fresh weight than System 1.
Fresh weight is not a reliable indicator for comparing yield after applying any treatments. Dry weight on the other hand, provides a precise measurement of biomass, eliminating fluctuations caused by water content because the drying process eliminates water from the plant. The plant total biomass can be directly related to our plant performance in response to factors such as photosynthetic capacity, nutrition, and environmental conditions. This is why dry weight is the best option to record weight when evaluating treatments in terms of yield or quality [41], [42].
Referring to Table 7, the dry weight of plant samples of System 1 are just 14% higher than those of System 2. Leaf samples of System 2 have 4% higher dry weights than System 1. Comparing the dry weight results for the two systems, dry weight of plant samples in System 2 was decreased by 14%, whereas the dry weight of leaf samples of this system was increased by 4%. This result is consistent with the improvements observed in plant quality and energy savings, compared to a non-significant weight increase.

D. PLANT GROWTH INDICATORS THROUGH VISUAL MONITORING
Plant area and plant leaf area are shown to have a direct relationship with plant health and yield. In an attempt to quantify this relationship, periodic top view images were captured at different times of day under white light conditions. The PlantCV open source software was used to process these images and extract some features indicative of plant growth. These indicators only work until the plant canopies start to overlap (November 25th-day 35 for NN system and November 24th-day 34 for TS system). Figures 11 and 12 illustrate the results of shape analysis performed on collected plant images of Systems 1 and 2, respectively. The shape analysis is a tool used to extract shape characteristics of a plant image, including height, object area, convex hull, convex hull area, and perimeter.
In Fig. 13, a comparison of convex hull measurements for plants in Systems 1 and 2 is shown. This feature indicates the plant growth process. Some inaccuracies in the color-based plant segmentation method occurred as a result of the green  hue of the wool blocks and algae growth on the block surfaces, resulting in erroneous measurement peaks throughout the cycle. Plants 6 and 7 were the only ones that are in full view in both systems throughout the cycle. Comparing the corresponding figures, a slightly faster growth and higher plant area is observed for System 1, which is in agreement with the findings in section VI.C.

E. ENERGY SAVINGS PER UNIT OF DRY BIOMASS
The energy usage for the full growth cycle with D days for each of Systems 1 and 2 denoted by E c,j (j = 1, 2) can be calculated as follows where the daily energy usage E d,j (j = 1, 2) can be formulated as follows   in which E d,r,j and E d,b,j are the red and blue channel energy usage of system j in day d, respectively; T p is the photoperiod; u r,j,i and u b,j,i are the red and blue LED channel commands (dimming percentages), respectively; and P r = 14W and P b = 7W are the full power usage of each color channel of the light fixtures, respectively. Based on formulations in (11), (12), (12), and (13), the energy usage for each system throughout the full 45-day grow cycle is calculated as follows E c,1 = E c,r,1 + E c,b,1 = 17 + 9 = 26 KWh (15) E c,2 = E c,r,2 + E c,b,2 = 11 + 6 = 17 KWh (16) which shows that during the entire cycle, energy savings of 40% in the red spectrum and 35% in the blue spectrum were achieved, respectively, which corresponds to a net 38% electric energy saving. As discussed in section VI.C, despite the significant energy savings achieved by System 2, the dry weight of lettuce grown in System 1 was 14% percent higher than that of System 2. To make meaningful comparisons between the two systems, parameters such as dry biomass produced per kWh and energy consumption per unit dry mass are presented in Table 8. These results demonstrate that 40% more dry biomass was produced per kWh and energy consumption per unit dry mass of lettuce was reduced by 28% using the proposed neural-network controller compared to the time scheduling technique.

VI. CONCLUSION
Experimental results demonstrate that the proposed approach can reduce energy consumption per unit dry mass of lettuce when compared to the conventional time scheduling technique and ameliorate certain plant health issues related to excessive light intensity. The proposed system demonstrates improvements that can be achieved through controlling and maintaining a desired lighting profile during the crop growth phases but challenges remain. For instance, strategic placement of light sensors throughout the plant canopy must be addressed due to sensor shading by plants' leaves. Combining image data with sensor data to identify the correct amount of light that each plant receives could be one approach to address this challenge.
For many years, red and blue color channels have been the spectra of choice for horticultural LEDs due to their photosynthetic efficiency and well understood effects on the plant growth process. However, recent studies have highlighted the importance of lesser-understood light spectra, including UV, green, and far-red lights. Thus, LED technologies that permit independent manipulation of narrow spectral components can be utilized to manipulate the light wavelengths to achieve desired plant properties.
Finally, supplemental lighting for photosynthesis in greenhouses can impact spatial distributions of temperature, relative humidity, water vapor pressure deficit, air movement, and CO 2 concentration. The study of these relationships through establishment of an ecosystem comprised of various environmental sensors and actuators is a potential area of future research. For each specific case, a multi-objective optimization problem can be defined based on plant requirements, resource availability, and cost, to optimize resource use efficiency, and plant quality and quantity.