USING SIMILARITY AND NEURAL NETWORK APPROACHES TO INTERPOLATE SOIL PARTICLE-SIZE DISTRIBUTIONS

Authors: NEMES, ATTILA - HUNGARIAN ACADEMY OF SCI; Pachepsky, Yakov; Rawls, Walter; WOSTEN, HENK - ALTERRA, NETHERLANDS; ZEILIGUER, ANATOLE - MOSCOW STATE UNIVERSITY

Submitted to: World Congress of Soil Science
Publication Type: Proceedings
Publication Acceptance Date: 4/1/2002
Publication Date: 8/1/2002
Citation: Nemes, A., Pachepsky, Y.A., Rawls, W.J., Wosten, H., Zeiliguer, A. 2002. Using similarity and neural network approaches to interpolate soil particle-size distributions. World Congress of Soil Science. p.943-1-943-10.

Interpretive Summary: Soil texture is defined by proportion of different textural fractions, i.e. particles having specific ranges of sizes. Historically, textural fractions have been defined by different ranges of sizes in different countries. Recently, conversion between different national systems of textural fractions became of interest because of use soil data for global change modeling and estimation. Such use is based mostly on development soil hydraulic pedotransfer functions to transfer soil textural information into soil hydraulic properties, which are otherwise costly to measure. Many advances in PTF development were based on the USDA or FAO texture classification system. A texture conversion procedure is needed to use those PTFs in other countries. Our objective was to develop procedures for such conversion. Two such procedures were developed and tested using an extensive world database on texture of 119,000 soil samples that we accumulated. One procedure is based on a search for soils in the database that are the most similar to the soil in question. Another procedure is based on using artificial neural networks to uncover the complex relationships between the proportions of textural fractions and to use them for prediction purposes. Both procedures gave similar results in terms of the prediction accuracy, and both substantially outperformed the log-linear interpolation that had long been a standard method to use. To improve the accuracy of the estimates, it was important to use data from different regions of the world both in training neural network and in using soils with similar textures. The proposed new techniques offer a solution for expanding the worldwide applicability of pedotransfer functions.

Technical Abstract: Soil hydraulic pedotransfer functions (PTFs) transfer simple-to-measure soil survey information into soil hydraulic characteristics, that are otherwise costly to measure. Soil texture is among the key inputs for most current PTFs. Most PTFs are developed using texture data compatible to the USDA or FAO particle size system. However, many laboratories around the world use different particle size systems. The purpose of this work was to develop and compare methods to convert such data into USDA or FAO system. A total of about 119.000 measured soil particle-size distribution (PSD) from the NRCS, BIS (The Netherlands) and HYPRES databases were used to develop and test two methods for the interpolation of the PSD curve. The 'similarity method ' does not rely on mathematical interpolation but involves searching in a sufficiently large external reference data set for a number of soils that have a particle-size distribution similar to the PSD Dof the soil in question. The neural network method was used as an interpolator to predict intermediate points on the PSD curve from measured PSD points. The accuracy of each method was tested with independent data. Both methods outperformed the log-linear interpolation which had long been a standard method to use. The two methods showed comparable prediction accuracy. Soil texture affected the prediction error. Classifying soils using estimated data by textural class was successful in 50% to 90% of cases. The proposed new techniques offer a solution for expanding applicability of pedotransfer functions in various regions of the world.