Effect of Neural Network Architecture When Predicting Soil Water Content at Field Capacity and Permanent Wilting Point.

Measuring soil water content is time consuming and expensive, so indirect methods have been developed to predict the water contents. Among these methods, Pedotransfer Functions (PTFs) have proved to be effective for predicting soil hydraulic properties such as water content at field capacity and wilting point. In the last decades, the use of Artificial Neural Networks (ANNs) has been used to build more accurate PTFs for predicting the water contents. Typically, a neural network is built by using a computer program like Matlab which include specific toolbox for this purpose. Because this process requires a series of decisions, the objective of the study was to compare the performance of different networks based on the number of input and hidden layers, with a large set of Chilean soils. By default, Matlab build the network by using 10 hidden layers to get one output, and uses 70% of the data for training, 15% for validation, and the rest for testing the network. There are also three algorithms to train the network, which vary in memory use and time for processing. In this study a series of combination among these options were evaluated in search of improving the R², root mean square error (RMSE) and Nash-Sutcliffe coefficient of the water content estimates at field capacity and wilting point. The results showed that ANNs perform better when developed with 5 inputs and 14 or 15 hidden layers, and the quality of estimates is not highly affected when using 4 inputs instead of 5. These two finding can be crucial when using the ANNs techniques with a reduced soil database or when the computer memory and time for processing is limited. This paper provides a set of recommendations and analysis for building the most appropriate ANNs when predicting soil water contents.