CalStack-SP: Calibration-First Class-Wise Stacking with Selective Prediction for Φ-OTDR DAS event classification

Distributed acoustic sensing (DAS) based on phase-sensitive optical time-domain reflectometry (Φ-OTDR) enables large-scale monitoring of pipelines, transportation corridors, and perimeter security. However, field DAS measurements are often affected by low and heterogeneous signal-to-noise ratios (SNR), environmental noise, and temporal drift, which may lead to overconfident but unreliable recognition outputs. To support reliable threshold-based alarming, this paper proposes Calibration-First Class-Wise Stacking with Selective Prediction (CalStack-SP), a confidence-aware measurement interpretation framework tailored to Φ-OTDR DAS event classification. CalStack-SP first applies branch-wise temperature scaling to align predicted confidence with empirical correctness, thereby producing uncertainty-calibrated sensing outputs. It then performs class-wise simplex-constrained convex stacking on the probability simplex to fuse heterogeneous learners in a stable and interpretable manner. A lightweight top-2 margin gate further enables selective prediction by accepting high-confidence segments and deferring ambiguous measurements to the calibrated ensemble, supporting risk-aware trade-offs among coverage, missed detections, and false-alarm control. To address Φ-OTDR distributed sensing constraints, DAS-specific feature stabilization and training augmentations, including SNR-targeted noise injection, temporal shifts, and class-conditional mixup, are incorporated. On a public nine-class Φ-OTDR benchmark, CalStack-SP achieves 95.06% accuracy and 95.51% Macro-F1, with a Brier score of 0.0832 and an ECE of 0.0081. On an open metro-tunnel intrusion dataset, it reaches 98.91% accuracy and 97.21% Macro-F1, suggesting favorable cross-scenario effectiveness with low additional inference cost.