Name: NELSON DE CARVALHO SANTOS
Publication date: 06/02/2026
Examining board:
Name |
Role |
|---|---|
| KATIA VANESSA BICALHO | Presidente |
| MARISTELA GOMES DA SILVA | Examinador Externo |
| SILVIO ROMERO DE MELO FERREIRA | Examinador Externo |
| WILIAN HIROSHI HISATUGU | Coorientador |
Summary: Soil water content (v) and suction are variables that influence the hydraulic and mechanical
behavior of unsaturated soils (US). However, direct measurements of these quantities may be
costly and operationally limited, motivating the use of indirect methods, such as Time Domain
Reflectometry (TDR), for estimating v, and the filter paper method (FPM), for estimating
matric or total suction. In this context, this dissertation aimed to analyze and compare the
performance of artificial neural networks (ANNs) and regression equations in the calibration
of TDR for estimating soil water content, as well as to evaluate the performance of bilinear
and exponential calibration curves of the FPM, based on experimental data available in the
literature, in estimating matric suction in different US. For TDR, a database was compiled from
the literature aiming at calibration through regression equations and ANNs, considering as input
variables the apparent dielectric constant (Ka), dry bulk density (BD), organic matter content
(OM), and clay content (% clay). For the FPM, bilinear and exponential calibrations reported in
the literature were evaluated, considering initially air-dried filter paper and the wetting contact
path. The results indicated, for TDR, superior performance of ANNs compared to regression
equations, with higher coefficients of determination (R2) and lower errors, expressed by the root
mean square error (RMSE) and mean absolute error (MAE), in addition to better generalization
capacity when applied to external datasets. It was observed that the exclusive use of Ka is
insufficient to adequately represent the variability of v, and the best-performing architectures
were those combining Ka with BD, OM, and clay content, with emphasis on ANN12-6, which
presented the best overall performance and behavior close to the ideal 1:1 condition for different
soil types. For the FPM, bilinear calibrations generally remained within normative tolerances;
however, they showed greater variability and performance loss under high suction regimes,
especially in soils with higher fine fractions. Among exponential calibrations, better overall
performance was observed, with estimates predominantly within confidence intervals. Overall,
it is concluded that the appropriate selection of calibration approaches and input variables is
essential to reduce uncertainties in estimating v and suction in unsaturated soils, highlighting
the superior performance of ANN-based approaches and exponential FPM calibrations.
Keywords: Unsaturated soils; TDR; Filter paper method; Volumetric water content; Matric
suction; Calibration; Artificial neural networks.
