Visual Mamba-Inspired Directionally Gated State-Space Backtracking for Chemical Gas Source Localization

Rapidly pinpointing the origin of accidental chemical gas releases is essential for effective response. Prior vision pipelines—such as 3D CNNs, CNN–LSTMs, and Transformer-based ViViT models—can improve accuracy but often scale poorly as the temporal window grows or winds meander. We cast recursive backtracking of concentration fields as a finite-horizon, multi-step spatiotemporal sequence modelling problem and introduce Recursive Backtracking with Visual Mamba (RBVM), a Visual Mamba-inspired, directionally gated state-space backbone. Each block applies causal, depthwise sweeps along H±, W±, and T± and then fuses them via a learned upwind gate; a lightweight MLP follows. Pre-norm LayerNorm and small LayerScale on both branches, together with a layer-indexed, depth-weighted DropPath, yield stable stacking at our chosen depth, while a 3D-Conv stem and head keep the model compact. Computation and parameter growth scale linearly with the sequence extent and the number of directions. Across a synthetic diffusion corpus and a held-out NBC_RAMS field set, RBVM consistently improves Exact and hit ≤1 over strong 3D CNN, CNN–LSTM, and ViViT baselines, while using fewer parameters. Finally, we show that, without retraining, a physics-motivated two-peak subtraction on the oldest reconstructed frame enables zero-shot dual-source localization. We believe RBVM provides a compact, linear-time, directionally causal backbone for inverse inference on transported fields—useful not only for gas–release source localization in CBRN response but more broadly for spatiotemporal backtracking tasks in environmental monitoring and urban analytics.