Redes neurais artificiais aplicadas na predição das umidades na capacidade de campo e no ponto de murcha permanente em solos do cerrado do centro goiano

Abstract

Obtaining moisture in the field capacity (CC) and at the permanent wilt point (PMP) areessential tools for irrigation management. Direct laboratory methods (tension table and pressurechamber) for obtaining CC and PMP are relatively expensive, requirea lot of time to obtaindata and are dependent on trained professionals to handle the devices. Thus, pedotransferfunctions (FPTs) obtained by multiple linear regressions (RLMs) or artificial neural networks(RNAs) have been shown as viable solutions in the prediction of these humidity. The presentework aimed to evaluate the use of RNAs and RLMs in the modeling of pedotransfer functionsfor the prediction of humidity in the field capacity and in the permanent wilt point in soils ofCentro Goiano (Axis BR-153). Soil samples were collected in 10 municipalities in the CentroGoiano Region (Axis BR-153) with 12 sampling points for each municipality and in two depths(0.0 -0.20 m and 0.20 -0.40 m). Laboratory analyzes were made for the physical atributes(sand, silt, clay, soil density, particle density, total porosity, macroporosity and microporosity),water (CC and PMP) and chemical (organic matter) of the soil. These attributes were used inthe study and development of RLMs and RNAs. The predictive variables for FPTs were: sand,silt, clay, soil density, particle density, total porosity, organic matter and depth. The predictedvariables were CC and PMP. The predicted variables were CC and PMP. After processing thedata and modeling the functions, the best models were chosen using the two proposedmethodologies. Model 2 of RLM was selected for variable CC with variables Micro and Ds aspredictors. Model 4 of RLM wasselected for PMP and has Micro, Pt, MO and Clay as predictorvariables. The models tested at the ANN were: Model1: Sand, Silt and Clay (CC); Model 2:Sand, Silt, Clay, Ds, Dp and MO (CC); Model 3: Sand, Silt, Clay, Ds, Dp, MO, Pt, Micro,Macro and Depth (CC); Model 4: Micro and Ds (CC); Model5: Sand, Silt and Clay (PMP);Model 6: Sand, Silt, Clay, Ds, Dp and MO (PMP); Model 7: Sand, Silt, Clay, Ds, Dp, MO, Pt,Micro, Macro and Depth (PMP) and Model 8: Micro, Pt, MO and Clay (PMP). Both RNA andRLM were able to estimate the CC variable in their chosen models with good accuracy. Allmodels generated by the ANN proved to be superior in predicting the PMP in relation to themodel of RLM. Models 2 (AIC=7,518; R²=0,890) of RLM and Model 4 (AIC=7,816;R²=0,856)ofRNA were indicated by the Akaike information criterion (AIC), for estimatingWC. Mo