Full text loading...
-
Channelized systems detection using deep learning with fast labelling strategy
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, Second EAGE Digitalization Conference and Exhibition, Mar 2022, Volume 2022, p.1 - 5
Abstract
Channelized systems characterization is key for geological context understanding. Moving the point of view from seismic vertical profiles to attribute in surface layers facilitates their detection and becomes more suitable for deep learning techniques. However, the channelized systems depict so complex delimitations than an exact labeling is not realistic in reasonable time. We assumed then a coarse labeling that may be done fast. In that case, some pixels are labelled as channel while they are not. We proposed to automatically modify the label values from deterministic 0 or 1 to probabilistic 0 to 1 with a fixed 1 values area around the channels’ skeletons and a smooth decreasing to 0 values. Then, classical accuracy metrics like Intersection-Over-Union are not relevant in coarse labeling context. Indeed, they act at pixel level without distinction between pixel level of importance. We derived a formula proposed for 1D structures to 2D channelized systems detection. The idea is to compute the distance map where pixel values are their distance to the areas of interest in both label and prediction images. Then, the accuracy metric uses a Gaussian attenuation with a tolerance on the pixel distance to allow geophysical uncertainties.