VOID FILL ACCURACY MEASUREMENT AND PREDICTION USING LINEAR REGRESSION

Le résumé fourni par la source

We present an innovative way to predict accuracy and associated error in void fill of digital surface elevation models. An answer to this question is desired: “How well does the filled data correspond with the truth values?” Typically, however, void filling is performed because this information is not known. It is often impractical due to time and cost to acquire truth data. This issue is typically ignored and treated as a best effort based only on visual appeal. However, by using statistical analysis, the behavior of the error given a particular output can be learned. An algorithm can learn from cases where some truth is available so that the behavior of the error can be predicted for cases where no truth data is available. For this process to succeed, each sample (void) must have a description that can be fed into a prediction model. This comes in the form of metrics computed from the void points and the surrounding non-void neighborhood. Any number of characteristic metrics can be computed for possible use in the final description collection, but hypothesis testing can then be used to determine a smaller subset of these metrics for use. We propose a methodology that predicts error produced by linear regression during the process of terrain void filling. We apply our approach to the filling of voids present in SRTM, IFSAR, LiDAR, or Seismic data. Also, while this methodology is independent of any specific algorithm, this discussion is centered on a non-image based void fill algorithm.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.