Dynamic prediction model for the net photosynthetic rate of tea plants based on Internet of Things and Deep-LSTM

Abstract: Tea plant has been cultivated as one of the perennial cash crops for years. The water condition of the tea plant has also posed an important influence on the growth and yield. Too much water can lead to the waste of water resources, while less water can produce water stress, indicating a negative impact on the development of plants. Since photosynthesis is one of the basic physiological processes, it is crucial to evaluate the water status of the tea plant using the net photosynthetic rate (Pn), further to provide the data supports for the water control. In this study, four water treatments of the tea plant were carried out to observe the Pn in the field for about five months. An environmental monitoring platform was also constructed for the tea plant to measure the soil moisture, air temperature, air humidity, canopy temperature, and photosynthetically active radiation (PAR) parameters using the Internet of Things (IoTs) and FreeRTOS operating system. The difference of canopy and air temperature (?T) upper and lower limit equations was applied to calculate the empirical crop water stress index (CWSI), in order to quantify the water stress degree of the tea plant. The Pn data of the tea plant during the experiment was measured every day to obtain the original one. A dynamic Pn prediction model of the tea plant was then established to name the deep long short-term memory (Deep-LSTM). The environmental parameters, canopy temperature, and CWSI of water treatments were integrated as the input features to predict the Pn under various water stress. A classic back propagation neural network (BPNN) model was also referred to evaluate the performance of the new model, compared with the root mean square error (RMSE) and determination coefficient (R2). The results showed that the environmental monitoring system of the tea plant performed better to collect the parameters. The linear fitting equation between the lower limit equation of ?T and vapor pressure deficit (VPD) was y=2.7-2.13x, where the R2 was 0.866. The mean CWSI values of four water treatments were 0.251, 0.437, 0.621, and 0.858 with 100%, 85%, 70%, and 55% soil field capacity (SFC), respectively. The special photosynthesis system was used to measure the Pn values under the four water stress in the long-term experiment. The statistical analysis of photosynthesis data showed that the mean Pn values of water treatments were 4.66, 4.15, 3.40, and 2.61 μmol/m2·s, respectively, indicating closely related to the water stress degree. The Pn values of the T2 group were only reduced by 10.9%, compared with the T1 group, while the water supply decreased by 15%. The water stress was used to provide the reference for water-saving irrigation. The Deep-LSTM model performed better to predict the Pn of water treatments, compared with the BPNN. The RMSEs of Deep-LSTM in water treatments were 0.304, 0.280, 0.157, and 0.160 μmol/m2·s, respectively, while the RMSEs of BPNN model were 0.980, 0.897, 0.633, 0.417 μmol/m2·s, respectively. The R2 values of the Deep-LSTM model were 0.846, 0.875, 0.893 and 0.954, respectively, and the R2 values of BPNN model were 0.516, 0.355, 0.315, and 0.432, respectively. It infers that the Deep-LSTM model presented better to accurately predict the Pn values of the tea plant. This finding can provide a reliable reference for the water-saving irrigation and water-stress strategy of the tea plant. The water stress was also quantified to construct the photosynthetic rate model, thereby rapidly estimating the Pn values of the tea plant with the IoTs system. Therefore, clear scheduling can be made for the deficit irrigation, particularly for the water-saving of the tea plant.