《Modeling water retention curves of sandy soils using neural networks》.pdf

WATER RESOURCES RESEARCH, VOL. 32, NO. 10, PAGES 3033-3040, OCTOBER 1996 Modeling water retention curves of sandy soils using neural networks Marcel G. Schaap and Willem Bouten Landscapeand EnvironmentalResearchGroup, Universityof Amsterdam,Amsterdam,Netherlands Abstract. We usedneuralnetworks(NNs) to modelthe dryingwater retentioncurve (WRC) of 204 sandysoilsamplesfrom particle-sizedistribution(PSD), soilorganicmatter content(SOM), and bulk density(BD). Neural networkscanrelatemultiplemodelinput data to multiplemodeloutputdatawithoutthe needof an a priori modelconcept.In this way a highperformanceblack-boxmodelis created,which isvery usefulin a data explorationeffortto assessthe maximumobtainablepredictionaccuracy.We useda series of NN modelswith an increasingparametrizationof input and outputvariablesto get a better interpretabilityof modelresults.In the first two modelswe usedthe nine PSD fractions,BD, and SOM as input,while we predictedthe nine pointsof the water retentioncurve.These NNs had 12 input and 9 outputvariables,predictingWRCs with an averageroot-mean-squareresidual(RMSR)watercontentof0.020cm3cm-3.Aftera few intermediarymodelswith increasingparametrizationof PSD andWRC using(adapted) van Genuchten[1980]equationswe arrivedat a final NN modelthat usedsixinput variablesto predictthreevan Genuchten[1980]parametersresultingin a RMSR of 0.024 cm3 cm-3. We foundsaturatedandresidualwatercontentsto be unrelatedto the PSD, BD, or SOM, thereforethe saturatedwatercontentwasconsideredto be an independent inputvariable,while the residualwater contentwassetto zero. Sensitivityanalyses showedthat the PSD had a major influenceon the shapeof the WRC, while BD and SOM were lessimportant.On the basisof thesesensitivityanalyseswe establishedmore explicitequationsthat demonstratedsimilarityrelationsbetweenPSD and WRC and incorporatedeffectsof SOM and BD in an empiricalway. Despitethe fact that we