Modelling of some soil physical quality indicators using hybrid algorithm principal component analysis - artificial neural network

One of the important issues in the analysis of soils is to evaluate their features. In estimation of the hardly available properties, it seems the using of Data mining is appropriate. Therefore, the modelling of some soil quality indicators, using some of the early features of soil which have been proved by some researchers, have been considered. For this purpose, 140 disturbed and 140 undisturbed soil samples were collected from Jiroft, southern Kerman, Iran. Some physical and chemical properties of soil, for example, sand, silt and clay percentage, organic matter (OM), calcium carbonate (CaCO3), electrical conductivity at saturation (ECe), porosity (F), and bulk density (BD) were measured using standard methods. Some soil physical property indicators, including plant available water (PAW), relative field capacity (RFC), air capacity (AC) and saturated hydraulic conductivity (Ks) were also calculated. Using the hybrid algorithm of principle component analysis-artificial neural network (PCA-ANN), the calculated indicators were predicted by the easily available properties. The results showed that PCA-ANN had an acceptable accuracy in the modelling of soilphysical quality. The coefficient of determination (R2) of training and testing data for PAW, RFC and AC were 0.82 and 0.81, 0.90 and 0.79, 0.99 and 0.99, respectively. The optimization of Ks did not have the desired results. In other words, the R2 values of the training and testing data for this indicator were equal to 0.25 and 0.13, respectively.