Modelagem da retenção de água em solos sob clima subtropical usando funções de pedotransferência e redes neurais artificiais

Abstract

Pedotransfer functions (PTFs) emerge as an alternative to estimate the soil water content from soil attributes that are easier available in soil databases. The artificial neural networks (ANNs) are a tool that enables PTFs generation with better estimates. Therefore, the objective of this study was to explore the tools that allow the assessment and generation of PTFs to estimate the soil water content on a soil watershed. To achieve this, the study area selected was the Pelotas River Watershed (PRW), located in the southern Rio Grande do Sul - Brazil. In the PRW were established 100 sample points along a 25-kmspatial transect in which were collected soil samples from the surface soil (0-20 cm) to determine soil and topographical attributes. Thus, the following attributes were obtained: clay, silt, and sand content, bulk density, particle soil density, macroporosity, microporosity, soil organic content, soil water content in the potentials 0, -1, -6, -10, -33, - 100 e -1,500 kPa, elevation, aspect, soil slope, plan curvature, profile curvature, and curvature. The first study evaluated the PTFs transferability developed in other climate conditions for estimating the soil water content in the matric potentials -33 and 1,500 kPa. The bests PTFs were chosen according to the performance to estimate the variation rates of actual evapotranspiration, obtaining that the PTF developed by Vereecken et al. (1989) had the better performance. The second study carried out the generation of a pseudocontinuous PTF (PC-PTF) through the ANN technique to estimate the soil water content in several matric potentials. A soil database was generated to achieve this, joining the PRW data and the 105-point grid collected in the Ellert Creek/Canguçu-RS Watershed (ECW), which is a headwater of the PRW. The random search technique was used to explore the hyper-parameter effects of the ANN, to find the input attributes selection, and to obtain one model for improving soil water content estimations. The ANN's best input attributes were soil bulk density, soil organic carbon content, microporosity, clay, curvature, and plane curvature. Thus, the model with best performance to estimate the soil water content for the set of all matric potentials had hyper-parameters set: seed = 98.186; learning rate = 0.2; number of epochs = 5,000; and moment = 0,2. This model obtained performance statistical metrics r² = 0.73; ME = 0.052 cm3 cm-3 ; and RMSE = 0,065 cm3 cm-3.