ASSESSING DIGITAL SOIL MAPPING APPROACHES IN AN AGRICULTURAL LOWLAND AREA (LOMBARDY REGION, ITALY)

This PhD thesis delves into Digital Soil Mapping (DSM) field with a particular emphasis on its application in lowland areas. Comprising four distinct studies, this research endeavor unravels the intricacies of soil mapping accuracy, spatial resolution, machine learning models, and the transferability of DSM models. Lowland regions, often overlooked in DSM studies, come into sharp focus as the ecological significance of these areas for agriculture, urbanization, and environmental resilience is underscored. The systematic review in the first study reveals an escalating interest in DSM for lowlands, indicating a burgeoning appreciation for its potential, driven by advancements in high-resolution Digital Elevation Models (DEMs) and accessible remote sensing data. This study underscores the importance of considering diverse environmental covariates and choosing appropriate DSM approaches, setting the stage for further investigations. The second study employs a range of machine learning models to predict and map soil properties in an agricultural lowland area of Lombardy region, Italy. Insights gleaned from this study lay the groundwork for the application of linear and nonlinear models as well as ensemble machine learning models and highlight the significance of terrain attributes in soil property prediction. In the third study, machine learning techniques, combined with residual kriging, were leveraged to predict the spatial distribution of Soil Organic Carbon (SOC) in an agricultural lowland area of Lombardy region, Italy. The findings elucidate the potential of machine learning with residual kriging in predicting SOC and underscore the importance of terrain attributes in the spatial distribution of SOC, offering tangible implications for soil management. The fourth study ventures into model transferability in DSM, shedding light on the impact of DEM spatial resolution. This critical exploration underscores the need for a thoughtful consideration of spatial resolution in DSM applications and advocates for caution when transferring models to varying resolutions. Recommendations arising from these studies include the integration of additional data sources, advanced machine learning techniques, and the development of improved methods for model transferability. This PhD thesis collectively contributes to advancing the field of Digital Soil Mapping. Its findings have direct implications for sustainable land management, precision agriculture, and environmental impact assessment. The comprehensive insights offered pave the way for future research aimed at enhancing soil mapping accuracy and soil health in lowland areas and beyond.