Improving Satellite Image Forest Cover Classification with Field Data Using Direct Sequential Cosimulation

Abstract

Land surface cover classification is assessed using Direct Sequential Co-Simulation, combining field observations with classified remote sensing data. Local co-regionalisation models are applied to account for local differences in both, field data availability and distribution, and the correlation between these hard data and the classified satellite images as soft data. The suggested methodology is based on two criteria: influence of the field observations dependent on field data availability and proportional to field data proximity; and, influence of the soft data dependent on their local correlation to the hard data. The method is applied to a study of four economically important forest tree species on the Setúbal peninsula. Local correlations between field observations (hard data) and satellite image classification results (soft data) are computed and interpolated for the whole study area. Direct Sequential Co-Simulation is performed conditioned to the local correlation estimates, yielding estimates and uncertainties for forest cover proportions. Cover-probabilities are combined into one forest cover classification map, constrained to reproducing the global proportions for the different classes. Direct Sequential Co-Simulation results show more contiguous forest covers — i.e. more spatial contiguity — than the classified satellite image. In comparison to the field data used for calibration during satellite image classification, the proposed simulation method improved forest cover estimations for species with good local correlation between hard and soft data and worsened those for species with poor local correlations.