Automatic first break picking with structured random forests

In the seismic wave exploration field, first break picking is an important underlying task. With the emergence of massive seismic data, it is urgent to develop automatic and accurate picking algorithms to relieve the huge workload of manual picking. Although many traditional and machine learning based techniques have been developed, it is still a challenging task to obtain satisfactory picking results in practice because of complex subsurface structures and low signal-to-noise ratio. Ensemble learning has been demonstrated to be a powerful tool to produce good prediction results in many fields. Its core idea is to create multiple models with some special techniques and then to combine the prediction results of each model with a fusion strategy. Structured random forests , a type of ensemble learning method, have been proven to be quite effective in implementing structured learning tasks. Based on the observation that adjacent traces can provide useful information to pick the first arrival time for a specific trace, we propose in this paper a novel structured random forest based first break picking framework. In particular, a multi-scale normalization technique is presented to make full use of the amplitude information at different scales. To capture local features of first breaks, an enhanced feature map based on the short-term/long-term average ratio method is also computed. By extracting patches from two channel feature maps, we construct a structured random forest to predict the locations of the first breaks. On the basis of the probability score map produced by the structured random forest, an effective post-processing strategy is proposed to further deal with the detected results. By conducting experiments with synthetic and field data, the proposed method is shown to be effective in identifying first breaks. It is significantly superior to the short-term/long-term average ratio method and support vector machine in terms of picking accuracy. With the advantage of parallel computing, the computational cost of structured random forest is acceptable, that is, it is much lower than the support vector machine while being higher than the short-term/long-term average ratio method. In addition, the experiments also confirm that the multi-scale normalization plays an important role in improving the picking performance of the structured random forest.