Evaluation of Feedforward Artificial Neural Networks (ANN) to Adjust Soil Moisture Estimates Derived From Time Domain Reflectometry (TDR) Measurements Using Soil Temperature and Gravimetric Data

Arturo Mendez-Patino, Jose Antonio Gutierrez-Gnecchi, Enrique Reyes-Archundia, Adriana del Carmen Tellez Anguiano

Soil temperature is one of the soil characteristics that greatly influences the accuracy of Time Domain Reflectometry (TDR) measurements for estimating soil moisture content. The authors examine the performance of two feedforward Artificial Neural Networks (ANN) configurations, commonly used for data regression analysis, to adjust TDR soil moisture estimates using soil temperature and gravimetric data. The data used for this study was obtained during a period of six weeks (October-November 2017) in three adjacent test sites in the Purepecha Plateau (Michoacán, México) managed under different tillage practices: at rest, reduced tillage and intensive tillage respectively. 10 TDR measurements per week were obtained from each test site. 60 Soil samples from each measurement site were also collected simultaneously, to determine the soil moisture content by the gravimetric method, and the soil temperature at 20 cm depth. 24 different configurations of ANNs were tested. The best result was obtained using a feedforward ANN with 11 tanh-sigmoid neurons in the input (hidden) layer. In addition, the authors also analyze the effect of different tillage practices on the soil moisture data. The results corroborate that tillage practices influence the soil moisture measurements and thus the best ANN results are obtained when the data used for training the ANNs is derived from sites managed under the same tillage practice.