Seronet- a Robust Algorithm to Fit Neural Network Models for Pedotransfer Functions.
Carlos Alberto Alves Varella and Marcos Bacis Ceddia. UFRRJ, Campus universitário - BR 465/KM 7 - Intituto de Tecnologia, SEROPÉDICA, Brazil
Artificial Neural Networks were used to develop an algorithm to fit pedotransfer functions (PTFs) for estimation water retention at 10 kPa from soil texture and field position. Pedotransfer functions can be defined as predictive functions of certain soil properties from other easily, routinely, or cheaply-measured properties. An advantage of using the neural network approach is that no statistical assumptions need to be assumed allowing fitting models based only in training sets. The Neural Networks were trained with "backpropagation" and each weight and bias djusted according to Levenberg-Marquardt. The algorithm was tested on a data set of 125 points collected in an agro ecological farm located at Seropédica-RJ/BR. Therefore, the main objetive of this work was to develop an algorithm that can be used to estimate soil field capacity for precision agriculture. The algorithm perform the following steps: compares observed with predicted data, plot graphics, and presents a methodology for validation the selected model. The model fitted presents no tendency, witch can be observed by the parameters from cross-validation (Figure1). Figure1. Results from validation test.