|Authors||Poggio, L., Gimona, A. and Brewer, M.J.|
|Publication details||Geoderma 209-210, 1-14.|
|Keywords||Digital Soil Mapping; variables selection; complex models; stepwise; variables tournament; organic soils|
Knowledge of soil properties with complete area coverage is needed for policy-making, land resource management, and monitoring environmental impacts. Remote sensing offers possibilities to support Digital Soil Mapping, especially in data-poor regions. The aim of this work was to test the potential of time-series of MODIS (Moderate Resolution Imaging Spectroradiometer) vegetation and drought indices to provide relevant information to model topsoil properties in a Boreal-Atlantic region (Scotland) focussing on differentiation between soils with high and soils with low organic matter contents. For each of the three considered years, 345 MODIS data sets were included in the exploratory analysis; 15 data products for 23 dates (bi-weekly) per year. Terrain parameters derived from Shuttle Radar Topography Mission were also included. A methodology was implemented to exploit fully the high number of covariates, to identify the band, index or product that best correlates with the soil property of interest. In particular the proposed approach i. relies on freely globally available data-sets; ii. uses statistical criteria to select the combination of covariates providing the highest predictive capability, among the data considered and available; iii. deals with both continuous (using Generalized Additive Models, GAM) and multinomial categorical (using Random Trees) types of variables; iv. takes into account fully the spatial autocorrelation of the data; v. provides estimates of the spatial uncertainty for each pixel; and vi. is computationally efficient when compared with methods such as forward stepwise. The models fitted show a fairly good agreement with existing data sets, presenting a consistent spatial pattern. The use of MODIS data as covariates increased the predictive capabilities of GAMs using only terrain parameters. The misclassification error for organic matter classes was between 25 and 35 %. The assessment provided of the spatial uncertainty of the modelled values can be used in further modelling and in the assessment of consequences of climate-change and trade-off in land use changes. This approach can contribute to improving our understanding and modelling of soil processes and function over large, and relatively sparsely sampled, areas of the world.