FUME

Overview

Why FUME Matters

Early, non-invasive rumen health checks through precise multi-gas plume segmentation

Rumen acidosis quietly erodes animal welfare and production. Farmers need a rapid, hands-off indicator rather than invasive probes or lab work. FUME reads thermal CO₂ and CH₄ plumes directly to surface early warning signals.

The network aligns paired modalities through a weight-shared FastSCNN encoder, modality-specific self-attention, and channel-attention fusion, producing unified segmentation masks tailored to gas behavior.

With only 0.47M parameters and 12ms latency, FUME runs on edge devices while delivering 93.2% Dice and 87.8% mIoU, outperforming Gasformer, GasTwinFormer, and CarboFormer in both accuracy and efficiency.

Why FUME?

Key Contributions

Novel components enabling efficient multi-gas emission analysis

Lightweight & Fast

Only 0.47M parameters with 12ms inference latency, enabling real-time deployment on edge devices for practical farm applications.

State-of-the-Art Results

Achieves 93.2% Dice and 87.8% mIoU, outperforming all existing OGI-specific and general segmentation methods.

Dual-Modal Fusion

Novel channel attention fusion mechanism effectively combines CO₂ and CH₄ thermal streams for comprehensive analysis.

Architecture

Method

A unified architecture for multi-gas emission segmentation and health classification

1 Learning-to-Downsample

Efficient spatial reduction module that reduces input resolution to H/8 × W/8 using depthwise separable convolutions for minimal computational overhead.

2 Global Feature Extractor

Pyramid Pooling Module captures multi-scale contextual information at H/32 × W/32 resolution for comprehensive scene understanding.

3 Modality Self-Attention

Dedicated self-attention modules for each gas stream (CO₂ and CH₄) enable modality-specific feature refinement before fusion.

4 Channel Attention Fusion

Attention-based combination of dual streams produces unified representations for downstream segmentation and classification tasks.

Experiments

Results

Comprehensive evaluation against state-of-the-art methods

Quantitative Comparison with State-of-the-Art

Comprehensive performance comparison across segmentation metrics (Dice, IoU, HD95, ASD) and efficiency measures. † indicates best performance. For HD95, ASD, and Latency, lower is better.

SOTA Comparison — Performance Analysis

(A) Mean IoU comparison showing FUME substantially outperforms both general segmentation baselines and OGI-specific methods (Gasformer, GasTwinFormer, CarboFormer). (B) Per-class F1 radar chart demonstrating balanced classification performance across Healthy, Transitional, and Acidotic classes.

Result Comparison Spotlight

The qualitative panel below is the key evidence: FUME cleanly segments CO₂ and CH₄ plumes across Acidotic, Transitional, and Healthy cases. Note the tight boundaries, modality-consistent masks, and correct health tags even under faint, elongated, or fragmented gas shapes—critical for trustable early acidosis alerts.

Qualitative Results — Qualitative Segmentation Results

Visual comparison across Acidotic, Transitional, and Healthy classes for both CO₂ and CH₄ modalities. FUME achieves precise boundary delineation of gas plumes with predictions closely matching ground truth annotations, demonstrating robustness to varying plume morphologies.

Dataset Distribution — Dataset Statistics

Distribution across pH levels, health classes, and train/validation/test splits.

Per-class F1 Scores — Classification Performance

Per-class F1 scores (%) and balanced accuracy metrics across all health classes.

Analysis

Ablation Study

Understanding the contribution of each architectural component

Ablation Table — Component Analysis

Quantitative ablation showing the impact of removing each module. Δ denotes change relative to the full FUME configuration.

Ablation Visualization — Visual Ablation Analysis

(A) Dice coefficient comparison across architectural variants. (B) Multi-metric radar chart highlighting the critical importance of CO₂ modality and dual-task learning.

Reference

Citation

If you find our work useful in your research, please consider citing

@misc{islam2026fumefusedunifiedmultigas,
    title={FUME: Fused Unified Multi-Gas Emission Network
           for Livestock Rumen Acidosis Detection},
    author={Taminul Islam and Toqi Tahamid Sarker and
            Mohamed Embaby and Khaled R Ahmed and
            Amer AbuGhazaleh},
    year={2026},
    eprint={2601.08205},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2601.08205},
}