Answers are grounded in this paper and cite their sources.
Distributed acoustic sensing (DAS) benchmarks can be overly optimistic when overlapping windows or samples from the same recording cross training and test partitions.We evaluate multi-representation feature extraction and ensemble feature selection on two public DAS datasets using leakage-controlled protocols.In the Cao dataset, 423 of 441 sessions span the released partitions; in Tomasov, every randomly partitioned test window overlaps a training window.Session-grouped and guarded blocked partitions eliminate these contamination sources.Our pipeline extracts 1,002 multi-representation features and benchmarks seven lightweight classifiers and six balancing strategies.Light Gradient Boosting Machine (LightGBM) achieves 98.90 macro-F1 on Cao and 81.98 on Tomasov for segment-level generalization within represented recordings, exceeding convolutional neural network (CNN) baselines by 8.01 and 16.43 points on identical partitions, and exceeding five architectures published from 2024 onward, including graph and transformer models, under the same partitions.However, leakage control reduces the openclose F1 of our representation from 0.72 to 0.044, a loss a spectro-spatial graph model does not incur on the same partition, and Cao accuracy falls from 99.06% to 86.81% when acquisition dates are held out.Thus, group-aware, cross-campaign evaluation is essential for credible DAS benchmarking.
No takes yet. Share an insight, caveat, or question.
Khan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: