An improved method of tomato photosynthetic rate prediction based on WSN in greenhouse

In order to improve the efficiency of CO2 fertilizer and promote high quality and yield, it is necessary to precisely control CO2 fertilizer by wireless sensor network based on a model of photosynthetic rate prediction in greenhouse. An experiment was carried out on tomato plants in greenhouse for photosynthetic rate prediction modeling combined rough set and BP neural network. In data acquiring phase, plants growth information and greenhouse environmental information that may have influences on photosynthetic rate, including plant height, stem diameter, the number of leaves and chlorophyll content of functional leaves, air temperature, air humidity, light intensity, CO2 concentration and soil moisture, which were measured. And LI-6400XT photosynthetic rate instrument was used for obtaining net photosynthetic rate of functional leaf. After preliminary processing, 135 sets of data were obtained. And twelve of them were used for model test of neural network, while the others were used for modeling. All of the data were normalized before modeling. Two models were built to predict photosynthetic rate based on BP neural network. One had total nine input parameters. The other had six input parameters, chlorophyll content, air temperature, air humidity, light intensity, CO2 concentration, and soil moisture, which were reducted from original nine based on attributes reduction theory of rough set. Both two models have one output parameter, the net photosynthetic rate of single leaf. The genetic algorithm was adopted to reduct attributes. Since continuous data cannot be processed by rough set, the K-mean cluster method was used to discretize the data of nine input parameters before attributes reduction. The prediction results of two models showed that the model with six input parameters had a mean absolute error of 0.6958, an average relative error of 7.28%, a root-mean-square error of 0.7428, and a correlation coefficient of 0.9964, while the other model respectively had 0.4026, 4.53%, 0.3245 and 0.9965, which proved that the model with minimum attributes had higher prediction accuracy. On the other hand, the number of iterations was used to represent the neural network train speed. The result showed that the model with six input parameters had an iteration of 544, while the other had 1038. Hence, the reduction model was applied to controlling CO2 concentration. The net photosynthetic rates at different CO2 concentrations were predicted at a certain condition. The results had the same curve trend with theory analysis, and a high prediction accuracy, which proved that the model was useful for CO2 concentration control.