Predictive Modelling and Computer Vision Based Decision Support To Optimize Resource Use in Vertical Farms

Abstract

Vertical Farm (VF) control systems can be designed so that they automatically adjustenvironmental conditions to minimize operational costs. To make this possible, the traditional environmental control sensors for measuring the aerial and root zone environments must be supplemented with sensors that integrate crop specific information, such as crop biomass (or a proxy for biomass) in real time. The system can then use a dynamic growth model to make predictions about how the crop should grow along with the sensors which measure how biomass is developing in reality. Together, these sensors and model allow a control system to create a feedback loop which can correct for deviation in the estimated biomass. Environmental conditions such as air temperature, carbon dioxide (CO2) concentration, and light intensity must then be monitored and can be varied to optimize the final harvest mass, product quality, and harvest time. More importantly to costs, these inputs can be interchanged in a dynamic control strategy to optimize resource use and minimize costs. For example, in many cases increased CO2 can take the place of increased light in order to improve yield at a lower expense. Also, periods where costs are predicted to be more expensive, such as hotter times of day or peak-energy hours, can be avoided, and any observed or predicted loss of growth due reduced inputs at peak times can be made up for at later times. This research evaluated the viability of using computer vision to measure Top Projected Canopy Area (TPCA) as a proxy for biomass and its suitability as a feedback mechanism, and it also validated and modified, for the vertical farm setting, a crop-biomass-prediction model originally developed for greenhouse use. Together, these tools can be used in a feedback based control system to form a decision support tool for optimizing resource use efficiency in VFs. The model was validated against experimental results from several cultivars of lettuce grown in a VF. With these tools, various hypothetical scenarios were conceived to explore the model's ability to predict and potentially co-optimize CO2 concentration vs light intensity and to evaluate its potential for saving electrical energy and associated expenses. Future research should build upon this foundation to develop a real-time Reinforcement Learning (or other machine learning) based approach to decide upon the best sequence of future environmental conditions needed, on a moment by moment basis, in order to optimize resource use for each growing crop. The results of this study showed that for lettuce (and possibly leafy greens in general) while TPCA is a good indicator of crop maturity (prior to canopy closure), it can only be an accurate predictor of crop biomass after a calibration function is applied across the spectrum of possible setpoints of light intensity and CO2 concentration. This is because the rate of leaf expansion varies slightly with environmental conditions, especially light intensity and light spectrum, and, thus, plants’ TPCA expands at different rates for different environments. Further research can characterize the difference in expansion at various intensities and spectrums. The predictive model used in this research was improved and validated for 3 cultivars of lettuce at values of light intensity ranging from 11.6 to 14.4 mol m-2 day-1 DLI and CO2 levels ranging from 400 to 900 ppm for air temperatures between 19-23 °C. More importantly for dynamic control, the model was validated across a simple set of time varying setpoint conditions for CO2. This validation experiment also demonstrated that the efficiency of CO2 utilization is higher for more developed crops. This indicates a potential for conserving CO2 in early development since it is less effective during that time. Moreover, the predictive model evaluated growth outcomes for various environmental setpoints and indicated potential for co-optimizing resources both by providing a way to find optimal growing temperatures and by calculating the degree to which CO2 may be substituted for light without affecting yield. This also provides a means of quantifying the potential energy and, thus, cost savings which may be expected under various hypothetical sets of environmental conditions. It also provides a strategic decision support tool for growers to improve crop management and for researchers to trial various experimental designs before growing. It was also demonstrated that the effect of large changes in CO2 on the expansion of the canopy can be detected nearly instantaneously when measured by TPCA. This ability to capture changes in environment validated the use of TPCA as a part of feedback mechanism for dynamic control systems in the time prior to canopy closure, at which time the TPCA signal saturates due to the crowding of leaves. The results also suggest that further improvements can be made in remote biomass estimation with the adoption of additional computer-vision-based feedback metrics such as plant height.