Greenhouse Heat Map Generation with Deep Neural Network Using Limited Number of Temperature Sensors

In recent years, there have been many attempts in smart agriculture to increase efficiency and profitability, especially in horticultural agriculture, where profitability is high. One of the measures to achieve this goal is to realize uniform quality by equalizing temperatures in greenhouses which have a huge influence on a process of growth. The most common method for measuring temperatures in greenhouses is the use of temperature sensors. However, to measure continuous temperature distribution by scattering temperature sensors, a large number of temperature sensors must be installed, a method that should be avoided because of its high cost. Therefore, the goal of this paper is to estimate the temperature of substances such as crops and soil in greenhouses, which are secondarily affected by the atmospheric temperature, at a low cost instead of measuring atmospheric temperatures in a costly way. Temperature sensors for substances must be directly attached to the target object to measure its surface temperature, which can lead to the deterioration in quality. In contrast, infrared array sensors can measure the surface temperature of materials from a distance. They have been increasingly used in recent years due to growing demand, and they can be used to measure the surface temperature of a wide range of objects in a greenhouse. However, infrared array sensors also have many operational problems, such as dirty lenses, and the measurement error is larger than that of temperature sensors. Therefore, this paper proposes a machine learning model that predicts continuous temperature distribution in the form of a 16 ×18 pixels heat map from a limited number of temperature sensors. Evaluation results show that our approach is useful in different greenhouse environments, including different airconditioning systems. In addition, the model is computationally inexpensive enough to run in practical fields with limited computational resources; therefore, it can be run on relatively inexpensive embedded terminals. As for the accuracy, the average error of the heat map obtained by the proposed model is as small as 0.28 [°C/pixel].