Estudio de la distribución espacial y cartografía digital de algunas propiedades físicas, químicas e hidrodinámicas de suelos de la cuenca del Segura

Resum

ABSTRACT This PhD Thesis combines statistical techniques and the use of Geographic Information Systems to model and map some water storage capacity features of the soil arable layer in the Region of Murcia. The first objective involves regional scale models to predict and map the amount of several soil components (organic carbon, total or equivalent calcium carbonate and texture fractions). The second objective is to obtain models to predict the same soil properties at a local scale, in order to compare them with the models obtained at the regional scale and to analyse the differences between them. The third objective is to validate pedotransfer functions from the literature for the estimation of hydrological properties of the soil (bulk density, soil water content at field capacity and permanent wilting point), and to compare their accuracy with new pedotransfer functions calibrated with the soil data available for the Region of Murcia. The information sources used for the prediction of the physical-chemical soil properties were three soil data surveys; the Digital Terrain Model (DTM) at a 25m resolution; climatic data from 625 weather stations; the CORINE Land Cover map from 1990, and the map of soil classes digitalized by the Region of Murcia Autonomous Community (CARM). These data were analysed using the statistical software R and the geographic information systems GRASS and SAGA, obtaining 181 variables that were subsequently standarized. Afterward, those showing a higher collinearity were removed using the Variance Inflation actor method. In order to obtain the predictive models, a regression-kriging method was selected, in which the trend was adjusted using different parametric and non-parametric techniques, and the residuals were interpolated using ordinary kriging. For each selected soil component, 11 models were tested using different packages of R. A 5-Fold-Cross-Validation method was used, in which the first 4 blocks worked as calibration set, and the last fifth block worked as validation set. The goodness of fit of each model was measured through the correlation between measured and predicted values and the value of RMSE. The models performed at the regional scale are useful to estimate the mean values of the selected soil properties, although they are not able to explain all the spatial variability of the selected properties. In all of them, the Random Forest method with 5-Fold-Cross-Validation was the technique that obtained better results, including mainly climatic predictors, soil classes and covers, and some morphometric features. The highest goodness of fit and the best spatial structure of the residuals were obtained for Soil Organic Carbon and total calcium carbonate contents. On the other hand, The models for the texture fractions obtained lower correlations, higher error values and also a worst spatial structure of the residuals. The accuracy of the models at the local scale was no better than at the regional scale. Regarding the pedotransfer functions obtained from the literature, the best results for the three hydrological properties were obtained with the Barahona and Santos (1981) equation. The use of the available data to calibrate new pedotransfer functions did not improve these results. Notwithstanding the error propagated from the maps of physical-chemical soil parameters, it can be considered that the maps of the hydrological soil properties provide a good estimation of these edaphic attributes.