Função de pedotransferência para estimar o fósforo remanescente em solos, utilizando rede neural artificial

Abstract

The remaining phosphorus consists of the P concentration that remains in solution after shaking for 1 hour a soil sample with 0.01 mol L-1 CaCl2 containing 60 mg L-1 P. The remaining phosphorus values can be used as suitable indicators of the soil capacity of anion sorption due to be more dependable on the soil mineralogy than on the soil clay content. In Brazil, the remaining phosphorus is used as an ancillary variable in the official guidelines for determining fertilizer and amender requirements of agricultural soils of the Minas Gerais state. The main goal of this research was to develop a pedotransfer function (PTF) capable of providing fairly accurate estimates of remaining phosphorus values of representative soils of the São Paulo state from often-determined soil chemical properties and/or from other ones of easier determination. In this context, two pedotransfer functions were developed by using artificial neural networks (ANN) and multiple regression analysis (MRA) applied to a database formed by values of soil chemical and physical properties derived from soil surveys previously carried out in different locations of the São Paulo state. The multi-layer feedforward neural networks approach was used for the development of the ANN-based PTF being its topology determined from successive experiments. The simultaneous criteria adopted for choosing the best neural network were the performance during the training stage, measured by the mean squared error, and its capacity of providing accurate Prem values, which was evaluated by using a validation database in which statistical comparisons were done between the measured and estimated Prem values. The topology of the network that provided the most accurate estimates of the remaining phosphorus was [3 14 1], i.e., three neurons at the input layer, fourteen at a unique hidden layer and one neuron at the output layer; further development features included the use of the sigmoid logistic model as activation function, the input of data normalized in the [0;1] interval and the use of the resilient backpropagation learning algorithm. The three variables at the input layer were the soil pH value measured in 1 mol L-1 NaF (pH NaF), the sum of exchangeable bases (SB) and the soil content of exchangeable aluminum (Al3+), being the two last ones usually determined in soil test laboratories whereas the pH NaF determination is easier than the remaining phosphorus one. The MRA-based PTF was developed considering the same input variables of the ANN-based one, i.e., pH NaF, SB and Al3+. The comparisons performed with a same validation database showed that the pedotransfer function developed from artificial neural networks provided more accurate estimates of remaining phosphorus values. Despite the database used for the PTF development not be so comprehensive for the establishment of a definitive pedotransfer function for estimating remaining phosphorus values, the results of the present research indicate as promising the development of a massive database from chemical results often obtained by the Brazilian laboratories dedicated to soil fertility evaluation and that includes Prem and pH NaF values. This database will allow the development of a comprehensive ANN-based pedotransfer function capable of not only calculating suitable Prem values for practical applications but also reducing the expenses related to the analyses of a great number of soil samples.