Multinomial logistic regression with soil diagnostic features and land surface parameters for soil mapping of Latium (Central Italy)

One of the main challenges in traditional soil mapping is the identification of land components (LCs) - suitable combinations of morphology, lithology and land use - representing a fundamental step in the definition of soil typological units. LCs are traditionally used by pedologists to correlate different soils and to identify the relationships between soil and geography. The recognition of the various soil characteristics for LCs definition is usually performed considering a number of classes for each feature (e.g. slope classes), defined a priori at a national scale. Such classes tend by nature to generalize and to flatten the actual local variability. Moreover, when dealing with large areas, taking into account a very fragmented layout is very difficult. To reduce subjectivity in interpretation, the choice of features defining soil associated LCs should be performed in an automatic or semi-automatic way. Digital Soil Mapping techniques can help to overcome these limits. In this work we propose a procedure to define the soil LCs for the territory of Latium administrative region (Central Italy), starting from a dataset of about 1500 fully described and analysed soil profiles and associated land surface parameters. The main soil diagnostic characteristics - depth, internal drainage, topsoil and subsoil texture, gravel, organic carbon content, cation exchange capacity, calcium carbonate content - were used, together with auxiliary information derived from existing thematic maps (geology, land use, pedoclimate), and from Digital Elevation Model (geomorphometric parameters). Some indices derived from remotely sensed high resolution images were also introduced to improve the estimate in areas with uniform land characteristics (i.e. flat alluvial valleys and coastal plains). Classes for the measured characteristics were defined based on their frequency distribution and of the WRB classification diagnostic thresholds. The covariates were chosen by an ANOVA, and a Principal Component Analysis was performed to avoid multicollinearity effects. Resulting principal components were used as predictors in a Multinomial logistic regression to build a raster layer for each considered soil characteristic. From all these layers, a final map of LCs was derived by means of a clustering operation. The final map represents the basis to map Soil Typological Units for the whole Latium region.