Source-linked AI summary
TERRA: Task-Oriented Edge AI for Regional Fire Detection with Identity-Free Reporting
Mohammad Kazemi, Julie A. McCann
TL;DR
Regional fire monitoring seeks evidence of fire by region, although conventional individually addressed reports can add addressing and access coordination. TERRA combines local task-oriented reporting with identity-free, region-tagged aggregation; with nine reports and perfect recovery, TERRA-Feature has slightly higher AUPRC, while TERRA-Score uses fewer task bits and is less sensitive to false-message contamination.
Problem
Method
TERRA combines task-oriented local inference, identity-free region-tagged reports, and permutation-invariant regional aggregation using either quantized local fire probabilities or learned feature vectors.
Source: §1 · p. 1Results
With nine reports and perfect recovery, TERRA-Feature achieves mean AUPRC 0.990 using 24 task bits, while TERRA-Score achieves 0.987 using 6 bits; both attain MDR 1.81% at FAR ≤1%.
Source: §5 · p. 8Takeaways & Limitations
TERRA-Feature has a small AUPRC advantage under reliable recovery, whereas TERRA-Score is less sensitive to false decoded messages over much of the evaluated region.
Source: §5 · p. 8The evaluation uses synchronized multisensor observations from one controlled indoor fire-test environment, not multiple independently instrumented regions or an outdoor wildfire deployment.
Source: §4.4 · p. 8
Abstract
from arXiv · showhide
Distributed fire-monitoring systems may involve many sensing devices reporting over a shared uplink, even though the operational objective is often to identify regions exhibiting evidence of fire rather than to recover device-specific measurements. We introduce TERRA (Task-Oriented Edge Representation and Regional Aggregation), an edge-AI framework that combines task-oriented local inference, identity-free region-tagged reporting, and permutation-invariant regional evidence aggregation. TERRA supports two reporting modes: TERRA-Score, in which each device communicates a quantized local fire probability, and TERRA-Feature, in which it communicates a compact learned representation for receiver-side aggregation using a permutation-invariant DeepSets model. Using all 15 numerical sensing modalities of the distributed indoor EN54 fire-alarm dataset as a proof-of-concept testbed, we evaluate regional inference for an episode-level fire-scenario target under repetition-disjoint three-fold cross-validation. We study representation size, task-payload quantization, report availability, and robustness to missed and false recovered messages. With nine reports and perfect message recovery, TERRA-Feature achieves mean AUPRC 0.990 using a 24-bit task payload, while TERRA-Score achieves 0.987 using only 6 bits; both attain an MDR of 1.81% at FAR $\leq$1%. TERRA-Feature achieves slightly higher AUPRC under reliable recovery, whereas the more compact TERRA-Score is less sensitive to increasing false-message contamination.
1 Introduction
TERRA addresses regional fire detection when recovering device-specific reports can require unnecessary addressing and access coordination. It combines task-oriented edge reporting with identity-free regional aggregation, framing a tradeoff between compact local decisions and richer features for receiver-side inference.
1.1 Related Work and Positioning
- Edge intelligence and task-oriented communication transmit local decisions or intermediate representations; TERRA applies these ideas to region-wise event decisions rather than sensor-record reconstruction.TERRA recovers discrete task reports before regional aggregation, accommodating region tags and both scalar scores and learned feature vectors.Source: §1.1 · pp. 1–2
1.2 Contributions
- With nine reports and perfect recovery, TERRA-Feature achieves mean AUPRC 0.990 using 24 bits, while TERRA-Score achieves 0.987 using 6 bits; both attain MDR 1.81% at FAR ≤1%.The 30-bit Handcrafted-Stats benchmark achieves AUPRC 0.971 and MDR 11.42%.Source: §1.2 · p. 2
1. Introduction
- TERRA combines compact task reports, identity-free transmission with explicit region tags, and permutation-invariant fusion into regional decisions.TERRA-Score communicates completed local decisions; TERRA-Feature communicates intermediate learned representations.Source: §1 · p. 1, §1.2 · p. 2
1.2 Contributions
- TERRA-Feature performs best under reliable recovery, whereas the more compact TERRA-Score is less sensitive to increasing false-message contamination.The two modes expose a communication–inference tradeoff between compact completed decisions and more flexible receiver-side inference.Source: §1 · p. 1, §1.2 · p. 2
2 TERRA System Model and Framework
TERRA targets regional fire inference rather than device-specific measurement recovery, combining local edge inference, identity-free region-tagged reporting, and region-wise evidence fusion. Devices report either completed local scores or learned representations for receiver-side aggregation.
2 TERRA System Model and Framework
- The receiver produces a region-wise event map without requiring persistent transmitter identities or device-specific measurements.Each device belongs to a geographic region with a binary event state.Source: §2 · p. 2, §2.1 · p. 2
2.1 Spatial Sensing and Edge Reporting Modes
- TERRA-Score reports a compact local fire probability, while TERRA-Feature sends an intermediate representation for receiver-side inference.A device uses one reporting mode at a time; Feature carries a larger task payload.Source: §2.1 · p. 2
2.2 Task-Oriented Messages and Identity-Free Access
- Each message combines a quantized task payload, a region tag, and a fresh nonce, without a persistent device identifier.The nonce distinguishes otherwise identical messages, and URA recovers messages without resolving transmitter identities.Source: §2.2 · pp. 2–3
2.3 Region-Wise Task Aggregation
- TERRA-Score averages reconstructed local probabilities, whereas TERRA-Feature uses DeepSets to transform and pool reports into a regional fire probability.The aggregation is invariant to recovered-report ordering; a region with no reports receives a default probability of zero in the experiments.Source: §2.3 · p. 3
2.4 Message-Recovery Errors
- Missed recoveries remove genuine regional evidence, while false recoveries add decoded messages assigned to regions by their recovered tags.A decoding error may appear as both a missed genuine message and a false decoded message.Source: §2.4 · p. 4
3 Experimental Methodology
TERRA’s experimental methodology evaluates frozen edge models and regional receivers on repetition-disjoint EN54 episodes, then tests report availability, payload quantization, and message-recovery sensitivity. The study uses fixed windowing and controlled indoor data as a proof of concept.
3 Experimental Methodology
- Models and receivers are frozen before report-availability, quantization, and message-recovery experiments, which do not include communication-error adaptation or quantization-aware fine-tuning.Training and receiver selection use development data under repetition-level cross-validation.Source: §3 · p. 4
3.1 Dataset, Windowing, and Cross-Validation
- 304,920 measurements from nine synchronized EN54 nodes cover 15 numerical modalities; episodes are divided into 12-sample windows and evaluated as controlled episode-level fire scenarios.The indoor dataset is a proof-of-concept testbed, not direct validation of outdoor wildfire deployment; windows are fixed at approximately two minutes and are not optimized.Source: §3.1 · p. 4
- Three repetition-disjoint folds hold out one repetition of each fire scenario, with non-fire episodes assigned so every episode appears in testing once.Training-duration selection uses inner train–validation partitions; windows from an episode never cross between development and test sets.Source: §3.1 · p. 4
3.2 Edge Models, Regional Receiver, and Representation Selection
- The shared sensor encoder maps engineered temporal features from each 15 × 12 window to a four-dimensional embedding, then supplies either a local probability or that embedding for reporting.The dimension-selection rule chooses d=4; both reporting modes use this same sensor model, isolating the difference between transmitting a completed inference and an intermediate representation.Source: §3.2 · pp. 4–5
- TERRA-Feature aggregates recovered reports with a permutation-invariant DeepSets receiver, while Handcrafted-Stats provides a nonlearned temporal-statistics benchmark using the same receiver architecture and training protocol.The learned receiver transforms each report independently, masked-mean-pools available reports, and maps the resulting regional representation to a fire probability.Source: §3.2 · p. 5
3.3 Report Availability and Task-Payload Quantization
- With perfect message recovery, report availability is tested by retaining K ∈ {1, 2, 3, 5, 7, 9} from each synchronized nine-report regional example, using ten independently generated, identically matched subset realizations per K, without missed or false decoded messages.Source: §3.3 · p. 5
- Score probabilities use endpoint-preserving uniform quantization, whereas feature and handcrafted components use clipped, standardized mid-rise quantization without range optimization or receiver fine-tuning.The evaluated Score widths are 1, 2, 3, 4, 6, and 8 bits; feature and handcrafted components use those same per-component widths.Source: §3.3 · p. 5
3.4 Message-Recovery Sensitivity Study and Metrics
- Message-recovery sensitivity is evaluated over nine-node target examples, a synthetic 16-tag address space, and varied list-level missed-detection and false-alarm probabilities.False messages receive uniformly random valid region tags and valid task words; these distributions are modeling choices, and metrics are repetition-disjoint engineering estimates rather than formal significance tests.Source: §3.4 · pp. 5–6
4 Results and Discussion
With nine reports and perfect recovery, TERRA-Feature slightly exceeds TERRA-Score in AUPRC, while Score uses fewer task bits; under false-message contamination, Score is less sensitive. The evaluation also examines representation compactness, report availability, and communication limits.
4.1 Compact Edge Representations
- Selecting d = 4 retains mean regional AUPRC 0.990 and MDR 1.81% at FAR ≤1%, while reducing feature dimension by approximately 89% relative to d = 35.The best observed AUPRC is 0.993 at d = 35; subsequent comparisons use the same d = 4 sensor-side model for both TERRA modes.Source: §4.1 · p. 6
4.2 Payload Efficiency and Report Availability
- With nine reports and perfect recovery, TERRA-Feature reaches AUPRC 0.9896 versus 0.9868 for TERRA-Score; both have MDR 1.81% at FAR ≤1%, using 24 versus 6 task bits.The selected operating points exclude the region tag and nonce.Source: Table 1 · p. 6, §4.2 · p. 6
- The 30-bit Handcrafted-Stats report uses one bit per mean and standard-deviation component and has mean AUPRC 0.9713, versus approximately 0.946 unquantized.The improvement is interpreted as an empirical regularization effect, not evidence that one-bit quantization is intrinsically preferable.Source: §4.2 · p. 6, §4.1 · p. 6
- With fixed payloads, the learned modes gain most from the first few reports: at five reports, AUPRC is 0.983 for Score and 0.985 for Feature, with comparatively modest gains beyond K = 5.Feature slightly exceeds Score at all evaluated report counts; Handcrafted-Stats performs best with one report but saturates lower as reports are added.Source: Figure 2 · p. 7, §4.2 · p. 7
4.3 Robustness to Message-Recovery Errors
- False-message contamination separates the learned modes: at P_FA = 0.6 with P_MD = 0, TERRA-Score has AUPRC 0.9729 versus 0.9398 for TERRA-Feature.Their MDRs at that setting are 5.34% and 16.56%, respectively; the paper presents a possible explanation but does not establish it causally.Source: §4.3 · p. 7
- Across 90 sampled error settings, TERRA-Score has the highest mean AUPRC at 53 points, TERRA-Feature at 35, and Handcrafted-Stats at two, rather than uniformly dominating.At (P_MD, P_FA) = (0.5, 0.3), AUPRCs are 0.9605, 0.9414, and 0.9230, respectively.Source: §4.3 · p. 7, Figure 3 · p. 8
4.4 Communication Interpretation and Limitations
- The communication results use controlled list-level errors and a synthetic 16-region tag space on EN54 indoor data, not multiple instrumented regions or an outdoor wildfire deployment.Task-bit counts exclude region tags and nonces, so the study does not infer total message lengths or channel-use, energy, or latency reductions.Source: §4.4 · p. 8
5 Conclusion
TERRA combines compact local representations, identity-free region-tagged reporting, and permutation-invariant regional aggregation. On the indoor EN54 testbed, TERRA-Feature has a small AUPRC advantage under reliable recovery, while TERRA-Score is more compact and less sensitive to false decoded messages.
5 Conclusion
- A four-dimensional TERRA-Feature report achieves mean AUPRC 0.990 using 24 task bits, while TERRA-Score achieves 0.987 using only 6 bits; both attain an MDR of 1.81% at FAR ≤1%.The results concern the episode-level fire-scenario target on the distributed indoor EN54 testbed.Source: §5 · p. 8
- Under message-recovery errors, TERRA-Feature retains a small advantage when errors are dominated by missed reports, whereas TERRA-Score is less sensitive to false decoded messages over much of the evaluated region.Future work will map these application-level results to concrete low-power access schemes and evaluate larger outdoor deployments and richer regional inference tasks.Source: §5 · p. 8