Abstract:
Conventional approach to seismic facies classification is full of subjectivity and is a labor intensive task which requires services of an experienced stratigraphic interpreter and takes time. In absence of any published workflow or software for automated translation of the interpretation done on few In-lines to others in the same region, I approached the problem of seismic facies classification as a segmentation problem wherein I attempted to classify different blocks in the image (here, different facies) and translate it onto other In-Lines. Convolutional Neural Networks (CNNs) is a supervised learning technique that can be employed directly to amplitude data in seismic. We employed a pre-trained VGG16 model architecture with a Fully Connected Layer to classify the seismic section, pixel-wise into one of the defined facies type. For the task of training the model, we used Netherlands F3 Seismic Data, which is pre-interpreted with 9 identified horizon types.
This work shows that the predicted results with a categorical cross-entropy training loss of approximately 0.13 for 100 epochs can be achieved beyond which it increases, indicating over-fitting. The model performs fairly well in identifying major facies blocks, but works only few times in clearly delineating the boundaries of the facies blocks
