Deep Learning-Supported Kriging for Interpolation of High-Resolution Indoor REMs

In future communications systems, precise location information of users is a declared target. To improve the radio systems, spatial channel knowledge with the same local accuracy in form of precise Radio Environment Maps (REMs) is beneficial. Constructing REMs with channel measurements is not only costly but often not feasible for specific regions of interest. Consequently, it is necessary to construct REMs based on a limited number of observations. Kriging is typically used for interpolation in the literature. The solely distance-dependent semi-variogram inherently assumes an isotropic environment. However, radio environments, especially indoor, are not isotropic and modeling the directionality of the spatial correlation is not possible by means of a simple variogram function. That is why we propose to enhance the Kriging spatial interpolation by exchanging the semi-variogram model by a Deep Neural Network (DNN) to better describe the anisotropic channel correlations in real-world environments. Ordinary Kriging and our proposed approach are compared for different sampling resolutions and sampling methodologies, namely random and regular. Our proposed method improves the average accuracy and more importantly further increases the confidence in the provided predictions. Higher confidence in the prediction is a way to unlock the usage of such techniques for future communication networks.