NEURAL-NETWORK ANALYSIS FOR HIERARCHICAL PREDICTION OF SOIL HYDRAULIC-PROPERTIES

The solution of many field-scale Bow and transport problems requires e stimates of unsaturated soil hydraulic properties. The objective of th is study was to calibrate neural network models for prediction of wate r retention parameters and saturated hydraulic conductivity, K-s, from basic soil properties. Twelve neural network models were developed to predict water retention parameters using a data set of 1209 samples c ontaining sand, silt, and clay contents, bulk density, porosity, grave l content, and soil horizon as well as water retention data. A subset of 620 samples was used to develop 19 neural network models to predict K-s. Prediction of water retention parameters and K-s generally impro ved if more input data were used, In a more detailed investigation, fo ur models with the following levels of input data were selected: (i) s oil textural class, (ii) sand, silt, and clay contents, (iii) sand, si lt, and clay contents and bulk density, and (iv) the previous variable s and water content at a pressure head of 33 kPa, For water retention, the root mean square residuals decreased from 0.107 for the first to 0.060 m(3) m(-3) for the fourth model while the root mean square resid ual K-s decreased from 0.627 to 0.451 log(cm d(-1)). The neural networ k models performed better on our data set than four published pedotran sfer functions for water retention (by approximate to 0.01-0.05 m(3) m (-3)) and better than six published functions for K-s (by approximate to 0.1-0.9 order of magnitude). Use of the developed hierarchical neur al network models is attractive because of improved accuracy and becau se it permits a considerable degree of flexibility toward available in put data.