Moisture and temperature prediction in tillage layer based on deep reinforcement learning

【Objective】 To accurately predict the water and temperature of the arable layer using the correlation between soil near surface air temperature and humidity and soil internal parameters, and serve for the realization of fine agricultural planting management.【Method】 Aiming at the actual needs of soil tillage layer moisture and temperature prediction in training set acquisition and model verification, an internet of things data acquisition system based on embedded system and narrow band internet of things (NB-IoT) wireless communication technology was designed. A model combination strategy was explored based on the deep Q network (DQN) deep reinforcement learning algorithm. Based on the weighted combination of long short-term memory (LSTM), gated recurrent unit (GRU) and Bi-directional long-short term memory (Bi-LSTM), the DQN-L-G-B combination prediction model was obtained.【Result】 The data acquisition system achieved long-term stable and reliable collection of time series environmental data with equal intervals, and provided accurate training set and verification set data for soil moisture, temperature time series prediction based on deep learning. Compared with models such as LSTM, Bi-LSTM, GRU and L-G-B, the DQN-L-G-B combined model not only lowered the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) in the prediction of moisture and temperature on the tillage layer of the two soil types (loam and sand), but also increased R2 by about 0.1%. 【Conclusion】 Through the internet of things data acquisition system and the DQN-L-G-B combined model, the accurate prediction of soil moisture and temperature in the cultivated layer based on soil near surface air temperature and humidity can be effectively completed.