FUME: Fused Unified Multi-Gas Emission Network for Livestock Rumen Acidosis Detection

TL;DR

FUME tackles non-invasive rumen health monitoring by leveraging dual-gas optical imaging of CO$_2$ and CH$_4$ to classify health states $y \in \{Healthy, Transitional, Acidotic\}$ and segment gas plumes. The approach uses a lightweight, weight-shared Fast-SCNN encoder with modality-specific self-attention and channel attention fusion to jointly perform segmentation and classification under in vitro conditions, achieving 80.99% mean IoU and 98.82% accuracy with only 1.28M parameters and 1.97G MACs. A first dual-gas OGI dataset with 8,967 annotated frames across six pH levels supports robust evaluation, and results show CO$_2$ as the primary discriminative signal while CH$_4$ enhances boundary delineation; dual-task learning is essential for optimal performance. This work demonstrates the feasibility of gas-emission–based livestock health monitoring and enables practical, real-time acidosis detection, with future work extending to in vivo validation and temporal modeling.

Abstract

Ruminal acidosis is a prevalent metabolic disorder in dairy cattle causing significant economic losses and animal welfare concerns. Current diagnostic methods rely on invasive pH measurement, limiting scalability for continuous monitoring. We present FUME (Fused Unified Multi-gas Emission Network), the first deep learning approach for rumen acidosis detection from dual-gas optical imaging under in vitro conditions. Our method leverages complementary carbon dioxide (CO2) and methane (CH4) emission patterns captured by infrared cameras to classify rumen health into Healthy, Transitional, and Acidotic states. FUME employs a lightweight dual-stream architecture with weight-shared encoders, modality-specific self-attention, and channel attention fusion, jointly optimizing gas plume segmentation and classification of dairy cattle health. We introduce the first dual-gas OGI dataset comprising 8,967 annotated frames across six pH levels with pixel-level segmentation masks. Experiments demonstrate that FUME achieves 80.99% mIoU and 98.82% classification accuracy while using only 1.28M parameters and 1.97G MACs--outperforming state-of-the-art methods in segmentation quality with 10x lower computational cost. Ablation studies reveal that CO2 provides the primary discriminative signal and dual-task learning is essential for optimal performance. Our work establishes the feasibility of gas emission-based livestock health monitoring, paving the way for practical, in vitro acidosis detection systems. Codes are available at https://github.com/taminulislam/fume.

Figures (5)

• Figure 1: Efficiency vs. segmentation quality trade-off across all compared methods including OGI-specific baselines. The upper-left region represents optimal performance (high mIoU, low computational cost). FUME achieves the best trade-off, delivering state-of-the-art segmentation quality while requiring significantly fewer MACs than competing methods.
• Figure 2: FUME Architecture. Our dual-stream network processes paired CO$_2$ and CH$_4$ thermal images through a weight-shared Fast-SCNN encoder with depthwise separable convolutions. The Learning-to-Downsample module reduces spatial resolution to $\frac{H}{8} \times \frac{W}{8}$, while the Global Feature Extractor with Pyramid Pooling Module captures multi-scale context at $\frac{H}{32} \times \frac{W}{32}$. Modality-specific self-attention modules refine each stream before channel attention fusion combines the representations. The fused features feed dual segmentation decoders and a classification head.
• Figure 3: Comparison with state-of-the-art methods. (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, where FUME maintains consistent high performance while several baselines exhibit severe degradation on challenging classes.
• Figure 4: Qualitative segmentation results across Acidotic, Transitional, and Healthy classes for both CO$_2$ and CH$_4$ modalities. FUME achieves precise boundary delineation of gas plumes with predictions closely matching ground truth annotations, demonstrating robustness to varying plume morphologies and emission intensities.
• Figure 5: Ablation study results. (A) Dice coefficient comparison across architectural variants, demonstrating that the proposed FUME configuration achieves optimal segmentation performance. (B) Multi-metric radar chart visualizing the critical importance of the CO$_2$ modality and dual-task learning framework for achieving balanced performance across accuracy, segmentation quality, and classification metrics.