Model-free spatial Interpolation and error prediction for survey data acquired by mobile platforms

International audienceThe paper proposes a new randomized Cross Validation (CV) criterion specially designed for use with data acquired over non-uniformly scattered designs, like the linear transect surveys typical in environmental observation. Numerical results illustrate the impact of randomized cross-validation in real environmental datasets showing that it leads to interpolated fields with smaller error at a much lower computational load. Randomized CV enables a robust parameterization of interpolation algorithms, in a manner completely driven by the data and free of any modelling assumptions. The new method proposed here resorts to tools and concepts from Computational Geometry, in particular the Yao graph determined by the set of sampled sites. The method randomly chooses the hold-out sets such that they reflect, statistically, the geometry of the design with respect to the unobserved points of the area where the observations are to be extrapolated, minimizing biases due to the particular geometry of the designs