SeisBench DAS: A machine learning framework for Distributed Acoustic Sensing

Summary

This software report presents SeisBench DAS, an extension of the SeisBench machine learning library for distributed acoustic sensing (DAS) data. Motivated by the lack of standardization among deep learning approaches for DAS, the authors define a benchmark dataset format (parquet metadata plus HDF5 records with per-channel annotations and optional chunking), a unified model API with annotate/classify entry points and memory-constant callbacks, and training pipelines with window-selection and probabilistic-labelling augmentations. A processing engine built on xdas virtual datasets supports larger-than-memory ingestion, chunked asynchronous inference with prefetching, cross-chunk causal IIR filtering, an optional PyTorch F-K filter, and smooth splicing of overlapping outputs; a wrapper also lets classical three-component models such as PhaseNet and EQTransformer run directly on DAS records. Throughput tests on a 3.1 GB record (36,718 channels, 225 s at 100 Hz) span 0.8–74 MB/s across models and configurations; the authors conclude that on a single Nvidia A40 only the wrapped classical models or DeepSubDAS at low overlap approach real time, with a roughly 13× CPU/GPU gap for DeepSubDAS. Main limitations: evaluation is throughput-only with no picking-accuracy results, the benchmark record's provenance is unstated, the initially shipped catalog is thin (one dataset, one DAS-native model), and the performance section contains an internal overlap-ordering inconsistency.