PaperlayerView exact revision on arXiv

Source-linked AI summary

MethaneFuse: Learning from Multi-Sensor Satellite Observations for Methane Plume Detection

Yuyao Wang, Juliana Y. Leung, Di Niu

arXiv:2609.09762v1cs.CVcs.LG

TL;DR

Methane plume detection is limited by incomplete, unevenly paired satellite observations. MethaneUnion organizes naturally available multi-sensor observations, and MethaneFuse learns from available sensor subsets, achieving stronger detection across sensor-availability settings and evaluation scales.

  • Problem

    Methane plume cases are captured by different subsets of public satellites, while many detectors rely on single-sensor inputs and leave reported cases unusable.

  • Method

    MethaneUnion matches Carbon Mapper plume reports with naturally available S2, L8/9, EMIT, and S5P observations, while MethaneFuse learns sensor-subset representations with lightweight sensor-aware adaptation.

  • Results

    MethaneFuse improves over the strongest baseline by 5.65 F1 and 8.30 AUROC points while reducing false positives by 8.19 points in the main 480 m evaluation.

  • Takeaways & Limitations

    MethaneFuse remains effective across sensor-availability conditions and evaluation crops from 120 m to 960 m, including when S2 is unavailable.

  • Takeaways & Limitations

    MethaneUnion inherits spatial, temporal, and reporting biases from Carbon Mapper records and matched acquisitions, while its plume masks provide approximate rather than exact boundaries.

Abstract

from arXiv · show

Methane plume detection from satellite imagery is constrained by incomplete observations: public satellites provide complementary spatial, spectral, and atmospheric evidence, but real plume cases rarely contain fully paired multi-sensor measurements because of revisit schedules, cloud coverage, acquisition quality, and the transient nature of emissions. Most learning-based detectors rely on single-sensor inputs, especially Sentinel-2 (S2), leaving many reported plume cases unusable. We construct MethaneUnion, a temporal multi-sensor dataset built from Carbon Mapper plume reports and matched S2, Landsat 8/9 (L8/9), EMIT, and Sentinel-5P (S5P) observations. Built on MethaneUnion, MethaneFuse learns from heterogeneous satellite observations under partial sensor availability without requiring complete four-sensor measurements. MethaneUnion expands usable coverage from 3,211 valid S2-matched plume cases to 8,981 reported plume cases with multi-sensor observations. At the representative 480 m setting, MethaneFuse achieves 84.87 F1 and 93.62 AUROC, improving over the strongest baseline by 5.65 F1 and 8.30 AUROC points while reducing false positives by 8.19 points. Sensor-availability experiments show that MethaneFuse improves detection when S2 is available and transfers plume knowledge to L8/9, EMIT, and S5P when S2 is unavailable. These results demonstrate the value of learning from incomplete heterogeneous sensor observations for practical methane plume detection.

I. INTRODUCTION

MethaneUnion addresses the scarcity of complete multi-sensor plume observations, and MethaneFuse learns from whichever satellite subsets are available. The framework expands usable cases and improves detection across sensors, scales, and tasks.

  • Dataset and motivation: 8,981 reported plume cases with multisensor observations expand usable coverage beyond the 3,211 cases matched to valid Sentinel-2 observations.MethaneUnion uses Carbon Mapper reports matched with S2, L8/9, EMIT, and S5P observations.
  • Method: MethaneFuse learns from available sensor subsets without requiring complete four-sensor measurements for every target location and time.Stage 1 keeps each sensor in native spatial–spectral form, masks unavailable sensors, and learns masked sensor-set fusion.
  • Sensor availability: Detection improves when S2 is present and transfers to L8/9, EMIT, and S5P combinations when S2 is absent.Reported gains reach 14.3 F1 points for S2+S5P and 15.3 F1 points in S2-absent groups.
  • Scale and segmentation: Performance remains stronger than the strongest heuristic fusion baseline across 120–960 m evaluation crops, while MethaneFuse-Seg improves fused IoU+ for segmentation.The segmentation result uses only sensors that can provide meaningful plume-mask predictions.

II. RELATED WORK

Prior methane detectors and generic Earth-observation models are largely sensor-specific or optimized for transferable geospatial semantics. MethaneFuse instead treats methane evidence as sensor-dependent and combines the available heterogeneous cues through methane-supervised learning and sensor-aware adaptation.

  • Sensor complementarity: Methane monitoring combines coarse atmospheric screening, multispectral SWIR plume analysis, and hyperspectral retrieval or mapping in complementary roles.Operational systems often use coarse detections to guide higher-resolution follow-up.
  • Prior methane learning: Existing methane detectors remain largely sensor-specific, while recent multi-sensor datasets mainly unify separate sensor-specific collections rather than model temporal partial observations.They do not explicitly represent each plume case through the heterogeneous observations available at the same target time.
  • Generic EO models: Generic Earth-observation models support heterogeneous inputs, but their strongest transfer settings involve persistent semantics rather than transient, observation-conditioned methane plumes.The distinction motivates methane-specific representation learning.
  • Methane-specific fusion: Methane evidence may appear as SWIR absorption, hyperspectral structure, multispectral anomalies, or coarse atmospheric enhancement depending on sensor and scale.MethaneFuse preserves and aggregates this union of available cues instead of learning only their common intersection.
  • Adaptation: Sensor-aware LoRA expert adapters address sensor-dependent false positives while keeping the shared methane representation frozen during downstream adaptation.Clouds, surface materials, spectral noise, retrieval artifacts, and resolution affect sensors differently.

A. MethaneUnion Dataset Construction

MethaneUnion organizes Carbon Mapper plume reports with naturally available observations from four satellite sources into partial, temporally contextualized sensor sets. The construction preserves event-level separation and supports learning despite missing sensors.

  • Event-centered construction: MethaneUnion represents each Carbon Mapper event using naturally available S2, L8/9, EMIT, and S5P observations rather than requiring a complete sensor set.Each event is anchored by timestamp, geolocation, and plume mask.
  • Partial observations: The binary availability indicator determines which sensors participate in tokenization, encoding, fusion, and inference.Unavailable sensors are excluded rather than imputed into a complete four-sensor input.
  • Temporal context: Each available sensor contributes an event-day observation, a seasonal-history observation, and a long-term-background observation.The implementation sets T = 3 and stacks temporal observations along the channel dimension.
  • Sensor preprocessing: EMIT preprocessing reduces its original 285 bands to a compact methane-relevant representation while preserving absorption structure near 2.3 µm.The observations are projected through WorldView-3 response functions, percentile-aligned, and reprojected to the native 60 m grid.
  • Data partitioning: Splitting occurs before crops, queries, or masks are constructed so all derived samples from one Carbon Mapper report remain in the same split.This prevents evaluation views derived from a training plume report from crossing the split boundary.

B. Sensor-Native Representation Learning

Stage 1 learns methane-aware representations from each available sensor in its native spatial–spectral form, then fuses only the observed sensor subset. The dataset pipeline separates representation-learning crops from query-level data while preserving partial availability.

  • Native representation learning: Fixed-pixel crops preserve each sensor’s native spatial context rather than forcing observations into a shared geographic footprint.S2, L8/9, and EMIT use 32×32-pixel crops, while S5P uses 3×3 pixels.
  • Native representation learning: Each available observation is projected into a shared ViT token space while retaining sensor-specific spatial, spectral, temporal, and channel structure.Metadata encodes sensor and channel semantics before common transformer processing.
  • Sensor encoding: The shared encoder processes available sensors independently, and cross-sensor interaction occurs by aggregating their sensor-level CLS representations.This preserves sensor-native encoding before fusion.
  • Sensor fusion: Masked attention pooling uses an availability mask so the same model can fuse arbitrary observed sensor subsets.Unavailable sensors do not enter the fusion operation.
  • Dataset construction: MethaneUnion separates event-level data into Drep for sensor-native representation learning and Dqry for query-level classification and segmentation.Event-level splitting occurs before crop and query construction.
  • Training objective: The Stage 1 objective combines fused methane supervision with auxiliary sensor-level supervision to learn from incomplete sensor sets.The fused head predicts from the available sensor set while auxiliary heads regularize sensor-specific representations.

C. Sensor-Aware Adaptation for Query-Level Detection

Stage 2 adapts the frozen Stage 1 representation to scale-controlled query classification and segmentation using lightweight, sensor-aware components. CLS-conditioned LoRA routing and masked fusion support incomplete sensor sets, while segmentation uses only sensors with meaningful local mask supervision.

  • Query construction: Stage 2 adapts Stage 1 representations to query-level classification and segmentation using event-indexed, timestamped, center-located, scale-controlled queries.Each query records the source event, timestamp, center, footprint, and available sensor observations.
  • Query construction: Queries use only naturally available sensor observations, with crop sizes determined by each sensor’s ground sample distance.S5P remains a coarse contextual CH4 view because its footprint is much larger than the localization footprints.
  • Sensor-aware adaptation: Stage 2 freezes shared weights and inserts trainable LoRA expert adapters into query and value projections of each ViT block.The adapters provide lightweight input-conditioned adaptation while the base projection remains frozen.
  • Sensor-aware adaptation: The current CLS representation softly routes each input through LoRA experts rather than assigning experts permanently to sensors.Expert selection reflects the sensor-native input, scene content, and transformer state.
  • Query classification: For classification, adapted sensor-level CLS features receive sensor identity embeddings before masked attention pooling across available sensors.This preserves sensor identity during partial-observation fusion.
  • Plume segmentation: Segmentation predicts independent probability maps for S2, L8/9, and EMIT, then applies late fusion after geospatial alignment.S5P contributes coarse atmospheric context to classification but is excluded from dense plume-mask supervision.

A. Experimental Setup and Baselines

The experiments evaluate methane detection under realistic event-level splits, incomplete sensor availability, and controlled query footprints. Comparisons include independent sensor predictors with heuristic fusion and fine-tuned Earth observation foundation models.

  • Labels and sampling: Training uses balanced positive and negative query sampling with approximately 50% positive samples.Negative queries come from the same event-centered acquisition but at least 5 km from the reported plume and with zero mask overlap.
  • Evaluation protocols: The primary chronological split assigns plume reports through May 16, 2025 to training and later reports to testing before constructing derived samples.A geo-clustered split separately tests spatial robustness by holding out entire macro-regions.
  • Evaluation protocols: The default query-level metadata contains approximately 625k samples, including about 498k training and 126k testing samples.These generated samples correspond to an approximately 8/2 partition.
  • Metrics: Classification is measured with F1, accuracy, false positive rate, recall, and AUROC, while segmentation uses IoU+.FPR and AUROC are emphasized because sensor-specific background artifacts affect methane classification.
  • Baselines: Baselines include independently trained sensor predictors combined by majority vote, average score fusion, or logical OR, alongside fine-tuned EO foundation models.The foundation-model comparisons include SatMAE, AnySat, and Panopticon.

B. Main Query-Level Classification Results

At the 480 m footprint, MethaneFuse outperforms heuristic fusion and generic EO transfer under naturally incomplete sensor availability. Its F1 advantage persists across the evaluated query footprints.

  • 480 m classification: MethaneFuse improves over average score fusion by 5.65 F1 points and 8.30 AUROC points while reducing FPR by 8.19 points.Average score fusion is the strongest heuristic baseline, with 79.22 F1 and 85.32 AUROC.
  • 480 m classification: MethaneFuse achieves the strongest performance across all reported 480 m classification metrics under partial sensor availability.The comparison includes heuristic score fusion and fine-tuned EO foundation models.
  • Baseline comparison: Panopticon-FT reaches 77.65 F1 and 83.28 AUROC, while SatMAE-FT and AnySat-FT perform lower in this methane-specific setting.These results are weaker than the strongest heuristic and MethaneFuse results.
  • Footprint sensitivity: MethaneFuse remains above heuristic fusion baselines at 120 m, 360 m, 480 m, and 960 m query footprints.The result indicates that performance is not tied to a single query scale.

C. Cross-Sensor Knowledge Transfer

MethaneFuse transfers methane-relevant knowledge across sensor availability regimes, improving detection for single-sensor and multi-sensor inputs while providing stronger ranking quality and regime-dependent false-positive control.

  • MethaneFuse improves F1 across most sensor-availability regimes, including single-sensor, S2-present, and S2-absent settings.The transfer evaluation compares MethaneFuse with independently trained per-sensor ViT predictors and average-score fusion.
  • MethaneFuse raises EMIT single-sensor F1 from 73.25 to 83.91, the largest gain among the four individual sensor streams.
  • MethaneFuse improves AUROC for all availability-conditioned groups, indicating more consistent plume-background separation across sensor subsets.
  • MethaneFuse reduces false positives for several groups, especially EMIT and multiple S2-containing combinations, but not uniformly across regimes.

D. Geographic Generalization under Geo-Cluster Splits

The geo-cluster protocol tests transfer to spatially held-out macro-regions, where geographic source and background context differ from training regions and performance is lower than under chronological splitting.

  • Geographic Generalization under Geo-Cluster Splits: The geo-cluster split assigns entire 275 km macro-regions to training or testing, reducing geographic neighborhood overlap and imposing stricter domain shift.
  • Geographic Generalization under Geo-Cluster Splits: MethaneFuse achieves the highest F1 at 480 m under spatially held-out macro-regions while using substantially fewer inference-time parameters than competing baselines.Figure 8 reports F1 against deployed inference-time parameters and performance across query scales.
  • Geographic Generalization under Geo-Cluster Splits: Spatially held-out performance is lower than chronological-split performance because region-specific source-context and background information are removed.

F. Ablation Study

The ablation supports lightweight Stage 2 adaptation: MethaneFuse outperforms full fine-tuning across evaluated query scales while reducing false positives and preserving strong AUROC at 480 m.

  • Ablation Study: The proposed adaptation freezes the Stage 1 encoder and updates LoRA expert adapters, routing modules, and downstream heads instead of fully fine-tuning the encoder.
  • Ablation Study: MethaneFuse improves F1, accuracy, and FPR over Full FT at all four query scales.The comparison is reported in the fusion-level Stage 2 adaptation ablation.
  • Ablation Study: At 480 m, MethaneFuse gains 1.37 F1 points over Full FT and raises AUROC from 90.89 to 93.62.
  • Ablation Study: At 960 m, MethaneFuse improves F1 by 1.02 points and reduces FPR by 3.95 points relative to Full FT.
Loading 2609.09762v1…