The paper proposes a novel approach for modelling multi-scale pattern of stromatolite deposit geometry. Spatial modelling of microbial carbonates is challenging due to the multi-scale structure of their porosity and permeability patterns. Conventional geostatistical methods are limited by stationarity assumptions, while object based approach may result in over-complicated representation of trends and bio-diversity. Therefore, it becomes difficult to adequately represent uncertainty across possible spatial patterns. The proposed kernel learning method is a data driven non-linear predictor that is based on blending spatial features derived from an outcrop analogue. Multiple kernel learning algorithm is capable to learn the relevant features from data and propagate them into prediction. Uncertainty of the predicted pattern is reflected by variation in the impact of individual input features. Application of the algorithm is demonstrated using a outcrop stromatolite analogue from Brazil.
|Number of pages||5|
|Publication status||Published - Jun 2014|
|Event||76th EAGE Conference and Exhibition 2014 - Amsterdam, Netherlands|
Duration: 16 Jun 2014 → 19 Jun 2014
|Conference||76th EAGE Conference and Exhibition 2014|
|Period||16/06/14 → 19/06/14|