TECNICHE DI CARTOGRAFIA DIGITALE PER LA REDAZIONE DI MAPPE DI DETTAGLIO DEL SUOLO

The goal of this work was to produce high-quality thematic maps for an Apennine area, using digital soil mapping (DSM) techniques; various statistical methods were applied starting from 30 geomorphometric variables, obtained from the digital elevation model (DEM) with 10-m pixel. The aim was to: i) obtain cartographic products, with a low cost both in terms of time and money; and ii) verify their reliability. The study area, consisting of two adjacent valleys (Oltrepo Pavese, PV), included Val di Nizza (27 km2), Val Ardivestra (47 km2) and the "plaque" of Pizzocorno-Pietragavina (17 km2). With regard to soil data acquisition, two different soil surveys were carried out, one of which was planned to obtain a minimum representative sample of soil variability of the area. A total of 132 georeferenced soil profiles were opened and described, and 468 soil samples were collected and analyzed for the main chemical and physical soil properties. The thesis work was composed by 7 chapters. The first chapter (Study site) provided the study area description, focusing on main factors which affect pedogenetic processes, such as climate, vegetation, geomorphology and geology. Among these factors, geological characterization was carefully described due to its strong relationship with soil. Previous soil data and land use history were also considered and described. The second and the third chapters were relative to digital soil mapping and geomorfometric variables respectively. In the Digital Soil Mapping chapter the theoretical and practical aspects, needed for obtaining soil thematic maps using DSM techniques, were reported. The DSM methodology, its related problems and the different approaches used to represent the pedogenetic processes, were addressed. The two approaches adopted in this work (soil-landscape paradigm and geomorphometric assessment of topography) were described in detail. The third chapter (Geomorphometric variables) included the preparatory study for soil mapping. In this section the geomorphometric variables were calculated; various inference methods were tested, with different combinations of variables calculated with open source and/or proprietary software. Before statistical elaborations the characteristics of the geomorphometric variables used as predictors were studied: in particular, the trend was analyzed, as well as reciprocal correlations and collinearity. Particular interest has been directed to the outliers, considering the influence they can have on calculations. From the analyzes carried out emerged that the outliers are connected to the calculation of the variables themselves and there is some degree of correlation, which is not said to correspond to collinearity. This redundancy of statistical information, however, corresponds to a different interpretation of the physical morphology of the land, which can be considered an additional information value to be used in statistical elaborations. The fourth chapter (Soil Sampling Design) provided a guidance for soil sampling strategy. The aim was to reduce time and costs by providing a map of the representative sampling areas. The selected approach was simple from both conceptual and computational point of view. It was based on the soil-landscape paradigm and on landform segmentation. Firstly, a principal component analysis (PCA) was carried out on the overall set of geomorphometric variables; then on the base of PCA results, eight variables were selected and used as input variables of a neural network (Self-Organizing Feature Maps), resulting in the identification of eight different geomorphometric units. In order to increase the pedological detail, the map was cross-checked with the geological map obtaining a total of 25 land units. To assess the quality of the results, the distribution of the geomorphometric variables of the study area was compared with that of the sampling points: for 77% of cases the target was statistically achieved. The remaining three chapters are relative to map production of soil characteristics and qualities: soil depth, soil erosion and soil types. The Solum Thickness Map chapter was relative to production and validation methodologies of solum thickness. Two approaches were used: artificial neural network (ANN) and partial least square regression (PLSR). By adopting two methods based on the PCA results, 28 different sub-sets of geomorphometric variables were created; different ANN types were applied to each of them, obtaining a total of 84 models. With the PLSR approach, 18 models were created by varying the number of retained components. The results showed that: i) the ANN approach was better than PLSR approach; ii) increasing the number of geomorphometric variables increased the prediction perfomance. An external dataset was used for validation. The map with the best performance achieved a R2 values of 0.89 for the validation set and 0.88 for the test set. The Soil Erosion Map chapter included a description of the different erosion types existing in the study area and explained the approach adopted to produce the soil erosion map. Since soil erosion was present in different forms the following maps were produced: i) a general map of soil erosion; ii) a map of the areas subjected to calanchi dynamics; iii) a map of susceptibility to shallow landslides. The first two maps were obtained using discriminant analysis (DA), while the last using a standard neural network for pattern recognition; they used a training set consisting of field detection points and five different results were obtained for each map. The results were evaluated using confusion matrices: the best results showed a non-error rate of 81 and 87% respectively. For the map of susceptibility to shallow landslides, no comprehensive analysis was performed, but 12 models with different sub-sets of variables were constructed, based on the PCA results performed on 112 landslide trigger points. The best model correctly classified 80% of the validation set and 64% of the test set points. The Soil Types chapter was relative to the production of: i) a taxonomical soil map according to WRB classification; and ii) a typological soil map according to a classification based on soil texture and thickness. These maps were produced using DA and evaluated using confusion matrices. The first map included nine soil types at the second level of WRB classification; the best cartographic product had a non-error rate of 58%. The second map included twelve soil categories. The obtained map had a non-error rate of 55%, but some soil classes were characterized by high misclassification risk values.