MAPPING SOIL ORGANIC CARBON DYNAMICS OVER THE LAST DECADES IN MEDITERRANEAN AGRO-ECOSYSTEMS WITH LEGACY DATA

Summary Soil organic carbon (SOC) represents the biggest carbon pool of the biosphere, bigger than the living plant pool. In agriculture, SOC is of pivotal importance for sustainable soil management and is a main soil fertility indicator. As soils are responsible for food production and the provision of various ecosystem services, there is a sturdy interest in understanding how land use and management affect natural plant and crop growth, and ecosystem resilience and functioning. These processes require time and soil sustainability is to be evaluated in a long-term economic perspective by policy makers with the aim of maintaining adequate, and likely improved, conditions of the soil and the whole farm for the future. Thus, long-term actions for crop sustainability could also admit little short-time yield reduction if yield potential, stability and environmental health are maintained at the long-time. Food production and ecosystem services provision depend on the maintenance, or increase, of SOC in agricultural soil, since SOC act as a short-term nutrient reservoir, increase water holding capacity and soil infiltration rate, reduce soil compaction, and favour soil resilience against pollutants. These effects should be taken into account at both a narrow and broad geographical breadth. When aiming to manage SOC at broad geographical extent, a detailed knowledge of SOC distribution and likely change in time is required. However, such a knowledge relies on correct sampling method and modelling procedures that in turn depend on the environmental variability of the area under study. Mediterranean areas are frequently variable as an harbour, the area has been subjected to a high share of soil and above-ground biodiversity and experienced long cultivation history and intensification since the last century, which increased their fragility. In this environment, the acquisition of reliable information on SOC can require a highly dense sampling, which can also negatively affect some relict environment. In addition, sampling can imply a high cost for field work and laboratory analyses. The aim of my Ph.D. work was thus to investigate the main factors related to SOC spatial distribution in agricultural land under various pedoclimatic conditions in semiarid Mediterranean areas, using a legacy soil database (1968-2008) of SOC and soil bulk density. The dissertation is structured in six chapters: the first one is a general introduction where the rationale of the dissertation is explained, and the research questions are stated. The second chapter is a novel approach to systematically collecting literature from international peer-review issues, namely systematic map. The third one is an analysis of the legacy soil database, which intends to make the database ready to be used for the SOC assessment and for the digital soil mapping. The fourth chapter touches an issue dealing with SOC stock mapping with the boosted regression tree and a set of covariates to produce local SOC benchmarks to be compared with European and Global SOC maps. The fifth chapter fits in the same modelling frame and it is addressed at the SOC dynamics using the most widespread legacy sampling campaign. A high number of available spatial data were collected and computed and used to calibrate the SOC models. At this stage, due to the ungridded structure of the data, a machine learning based model has been used (Boosted Regression Trees). The last chapter is a comparison of models (geostatistical, machine learning and linear), and shows useful information about the way that the error is reported by each algorithm. Soil maps are not just produced for the sake of creating attractive geographical visualizations: they have a very precise task to fulfil, i.e. provide accurate and reliable information on soil properties that decision makers can use to plan interventions of any kind. The use of the Regression Kriging and Boosted Regression Trees models, which resulted in the best prediction performance in terms of R2 and RMSE, highlighted the SOC dependence on environmental factors, and the prediction of the agricultural land covers. All land cover groups were studied in the preliminary stage of this study (chapter 2), while only the cropland identified with the legacy data was the candidate for the development of the final models which lead to the detection of a positive SOC trend. The last chapter aimed at the comparison between geostatistical, machine learning and linear models to predict SOC in agricultural lands, and an improvement in local uncertainty estimation. The outstanding result was that SOC at the monitoring sites were accurately simulated, being in full agreement with observed data. Once more, actual data will be available and the model will be calibrated and validated, a model of SOC potential sequestration regional scale can be produced. The results of this dissertation has led to a clear and shared vision in the community regarding the selection of the estimation methods for SOC prediction needs to be based on careful considerations. It is good practice to test algorithms already used in literature for similar purposes, but it may be counterproductive to only look at an algorithm because it is new and never used before in a particular field. This sometimes happens in science where methods are selected only because fashionable and not based on real and tested experiments. In the dissertation the origin of the data was sometimes know and sometimes it has been data driven based. In particular, sampling design was based on geostatistics only in the 2008 campaign and it may well be that looking at very advanced methods like deep-learning could be interesting, but still less accurate than the geostatistical kriging based algorithms, which can also provide robust and well tested uncertainty estimations. In summary, even though we have now access to advanced algorithms it does not mean that we need to use them blindly without fully considering what we are trying to achieve with our working hypothesis and research question.