Uso de diferentes metodologias na geração de funções de pedotransferencia para a retenção de água em solos do Rio Grande do Sul

• Autor(a) principal: Soares, Fátima Cibéle
• Data de Publicação: 2013
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do UFSM
• Texto Completo: http://repositorio.ufsm.br/handle/1/3601
• topic: Pedofunções; Umidade do solo; Potencial matricial; Inteligência artificial; Pedofunctions; Soil moisture; Potential matrix; Artificial intelligence; CNPQ::CIENCIAS AGRARIAS::ENGENHARIA AGRICOLA

Studies on the dynamics of water in the soil-plant-atmosphere such as water availability cultures infiltration drainage and movement of solutes into the soil, require knowledge of the relation between the water content in soil matric potential and represented by retention curve water. However, its implementation is laborious, requires considerable time and cost. An alternative is your estimate through statistical equations called pedotransfer functions (PTFs). The aim of this study was to generate PTFs for the different soil classes in the state of Rio Grande do Sul, through prediction methodologies. To develop the work we used data available in the literature, with values of hydro-physical characteristics and mineralogical characteristics of soils of the State, to estimate values of soil unit under different stresses. In possession of the database was conducted subdivision thereof, in different textural classes identified in the state in an attempt to improve the predictive ability of pedofunctions, forming more homogeneous subsets. The development of PTFs was from two modeling methods: (i) multiple linear regression (MLR) and (ii) artificial neural networks (ANNs). For the development of PTFs first methodology was used the "stepwise" (SAS, 1997). The PTFs generated from ANNs were implemented through the multilayer perceptron with backpropagation algorithm and Levenberg-Marquardt optimization. Each network is trained by varying the number of neurons in the input layer and the number of neurons in the hidden layer. The output variable was water content in soil matric potentials of 0, -6, -10, -33, -100, -500 and -1500 kPa. For each architecture, the network was trained several times, picking up training at the end of the architecture with lower mean relative error and lower variance in relation to the validation data. The efficiency of PTFs were analyzed graphically by the ratio 1:1 between data versus the observed and estimated by means of the following statistical indicators: correlation coefficient (r); concordance index Wilmont (c); coefficient of determination (R2) and performance index (id). The results showed that the more homogeneous is the data of the variables that compose the PTFs, the greater the precision in estimating the water retention in the soil, for the same. The network architecture consists of 4 inputs, showed high accuracy in the estimation of variables. The PTFs developed by ANNs outperformed the predictive ability of the standard method (MLR). Thus, the estimate of the retention curve of soil water by means of ANNs trained by classes textures, presents itself as a subsidy techniques adopted in irrigated agriculture.