Prediction of CO2 emission from greenhouse to atmosphere with artificial neural networks and deep learning neural networks

The research aimed to model CO2 flux from soil to atmosphere in greenhouse conditions, using multiple linear regression (MLR) artificial neural networks (ANN), and deep learning neural networks (DLNN). Following the purpose, crop species, soil temperature, soil moisture content, photosynthetic active radiation (PAR), and soil oxygen exchange were considered as input parameters and CO2 flux as an output parameter. Levenberg-Marquardt learning function and logarithmic symmetric sigmoid transfer function were utilized in both ANN and DLNN. The optimal number of hidden layer neurons was determined through empirical observation, the model which produces the least mean absolute error value was chosen in each structure. Thus, ANN utilized 8 neurons, while DLNN utilized 14 neurons in the first hidden layer and 10 neurons in the second hidden layer. According to the result, CO2 flux from soil to atmosphere was modeled using MLR with an accuracy of 95.63%, ANN with an accuracy of 95.56% and DLNN with an accuracy of 98.29%. Sensitivity analyses were conducted for both models to determine the pro rota efficiency of the input parameters on CO2 flux. In the research, it was concluded that CO2 flux from soil to atmosphere can be modeled in high accuracy, and deep artificial neural networks can have higher efficiency in similar works.