Assessment of Soil Quality under Different Soil Management Strategies: Combined Use of Statistical Approaches to Select the Most Informative Soil Physico-Chemical Indicators

Assessment of soil quality under different management practices is crucial for sustainable agricultural production and natural resource use. In this study, different statistical methods (principal component analysis, PCA; stepwise discriminant analysis, SDA; partial least squares regression with VIP statistics, PLSR) were applied to identify the variables that most discriminated soil status under minimum tillage and no-tillage. Data collected in 2015 from a long-term field experiment on durum wheat (Triticum durum Desf.) were used and twenty soil indicators (chemical, physical and biological) were quantified for the upper soil layer (0–0.20 m). The long-term iteration of different management strategies affected soil quality, showing greater bulk density, relative field capacity (RFC), organic and extractable carbon contents (TOC and TEC) and exchangeable potassium under no-tillage. PCA and SDA confirmed these results and underlined also the role of available phosphorous and organic carbon fractions as variables that most discriminated the treatments investigated. PLSR, including information on plant response (grain yield and protein content), selected, as the most important variables, plant nutrients, soil physical quality indicators, pH and exchangeable cations. The research showed the effectiveness of combining variable selection methods to summarize information deriving from multivariate datasets and improving the understanding of the system investigated. The statistical approaches compared provided different results in terms of variables selected and the ranking of the selected variables. The combined use of the three methods allowed the selection of a smaller number of variables (TOC, TEC, Olsen P, water extractable nitrogen, RFC, macroporosity, air capacity), which were able to provide a clear discrimination between the treatments compared, as shown by the PCA carried out on the reduced dataset. The presence of a response variable in PLSR considerably drove the feature selection process.