TRACE: Thermal Recognition Attentive-Framework for CO2 Emissions from Livestock

Taminul Islam; Abdellah Lakhssassi; Toqi Tahamid Sarker; Mohamed Embaby; Khaled R Ahmed; Amer AbuGhazaleh

cs.CVMar 27, 2026

TRACE: Thermal Recognition Attentive-Framework for CO2 Emissions from Livestock

Authors: Taminul Islam, Abdellah Lakhssassi, Toqi Tahamid Sarker, Mohamed Embaby, Khaled R Ahmed, Amer AbuGhazaleh

Organizations: Southern Illinois University, Carbondale · University of California, Davis

Abstract

Quantifying exhaled CO2 from free-roaming cattle is both a direct indicator of rumen metabolic state and a prerequisite for farm-scale carbon accounting, yet no existing system can deliver continuous, spatially resolved measurements without physical confinement or contact. We present TRACE (Thermal Recognition Attentive-Framework for CO2 Emissions from Livestock), the first unified framework to jointly address per-frame CO2 plume segmentation and clip-level emission flux classification from mid-wave infrared (MWIR) thermal video. TRACE contributes three domain-specific advances: a Thermal Gas-Aware Attention (TGAA) encoder that incorporates per-pixel gas intensity as a spatial supervisory signal to direct self-attention toward high-emission regions at each encoder stage; an Attention-based Temporal Fusion (ATF) module that captures breath-cycle dynamics through structured cross-frame attention for sequence-level flux classification; and a four-stage progressive training curriculum that couples both objectives while preventing gradient interference. Benchmarked against fifteen state-of-the-art models on the CO2 Farm Thermal Gas Dataset, TRACE achieves an mIoU of 0.998 and the best result on every segmentation and classification metric simultaneously, outperforming domain-specific gas segmenters with several times more parameters and surpassing all baselines in flux classification. Ablation studies confirm that each component is individually essential: gas-conditioned attention alone determines precise plume boundary localization, and temporal reasoning is indispensable for flux-level discrimination. TRACE establishes a practical path toward non-invasive, continuous, per-animal CO2 monitoring from overhead thermal cameras at commercial scale. Codes are available at https://github.com/taminulislam/trace.

Figures & tables

Figure 1: Parameter–mIoU efficiency frontier on the CO 2 Farm Thermal Gas Dataset. Bubble size encodes Boundary F1 (BF1); model families are colour-coded. TRACE (4.1 M) simultaneously achieves the highest mIoU (0.998) and BF1 (0.989), occupying a unique Pareto-optimal position against gas specialists, general transformers, and lightweight backbones.

Figure 2: TRACE per-class Precision, Recall, and F1. Low-Flux has the highest Recall (0.982); High-Flux shows the largest P–R gap (0.784 vs. 0.721), reflecting confusion with Control during transitional breath cycles.

Frames (segmentation)
Split	Total	Train	Val	Test
High-Flux	1,680	1,176	252	252
Control	1,840	1,288	276	276
Low-Flux	1,520	1,064	195	261
Total	5,040	3,528	723	789
Clips (classification)

Table 1: CO 2 Farm Thermal Gas Dataset statistics.

Figure 3: CO 2 Farm Thermal Gas Dataset overview. Each column is a representative frame sampled across varied breathing phases. Top row: raw MWIR thermal frames. Middle row: false-colour CO 2 intensity overlay $\Psi_{t}$ at 4.2–4.4 µm; orange-to-yellow gradient encodes plume concentration. Bottom row: binary ground-truth plume masks. The wide morphological variation — from compact, high-density plumes to diffuse, low-contrast clouds — highlights the segmentation challenge and motivates TGAA’s gas-conditioned attention.

Figure 4: Overview of TRACE. TGAA extracts $\Psi$ -conditioned multi-scale features; the decode head produces per-pixel plume masks $\hat{S}$ . ATF aggregates three streams (mask, encoder, CNN) via cross-frame attention for flux classification. Bottom: four-stage curriculum – S1a/b warm up segmentation, S2 aligns ATF to frozen VideoMAE-Small (discarded after S2), S3 fine-tunes end-to-end. Lock/fire = frozen/trainable.

	Efficiency			Segmentation			Boundary			Classification
Model	Params	GFLOPs	Lat. $\downarrow$	mIoU	Dice	TI	BF1	HD $\downarrow$	CLE $\downarrow$	Acc.	BAcc	F1	$\kappa$	Gini
	(M)		(ms)					(px)	(px)
Mask2Former [ 44 ]	27.9	14.8	34.2	0.5380	0.6849	0.7523	0.0869	50.742	11.117	0.760	0.615	0.562	0.586	0.422
SHViT-S4 $\dagger$ [ 45 ]	16.6	13.1	30.5	0.7663	0.8619	0.9058	0.1381	21.720	3.307	0.375	0.333	0.182	0.079	$-$ 0.290
$\Psi$ -Stats ♮	—	—	1.8	0.8842	0.9378	0.9492	0.3218	8.470	1.342	0.548	0.522	0.491	0.312	0.398
RepViT-M1 $\dagger$ [ 35 ]	5.2	1.4	3.4	0.8923	0.9416	0.9539	0.3554	18.380	1.588	0.375	0.333	0.182	0.082	$-$ 0.104

Table 2: Unified dual-task comparison (789 seg. frames; 104 cls. clips). Sorted by mIoU

\uparrow

; bold = best. ♮

\Psi

-Stats: Otsu seg. of

\Psi_{t}

+ temporal

\Psi

statistics

\to

MLP for classification.

\dagger

: lightweight backbone + SegFormer decode head.

\downarrow

= lower is better; Gini

=2{\times}\text{AUC}{-}1

Figure 5: Qualitative segmentation on three test frames. Columns: raw thermal frame, CO 2 overlay, GT mask, five baseline predictions, and TRACE ( blue ). TRACE closely matches the GT in shape and boundary sharpness, correctly delineating diffuse plume regions that baselines either over-segment (Gasformer, SegFormer-B0), under-segment (MobileNetV4-S), or miss entirely (SHViT-S4).

Figure 6: (a) Per-pixel

\Psi

distribution (plume vs. background); the hatched overlap explains

\Psi

-Stats’ low mIoU (0.884). (b) mIoU by difficulty condition; TRACE’s advantage widens in the hardest cases (rapid motion:

+

3.1 pp; high wind:

+

2.4 pp).

(a) $\Psi$ -fairness (segmentation)
	mIoU		BF1
Model	Base	+ $\Psi$	Base	+ $\Psi$
MobileNetV4-S $\dagger$	0.921	0.943	0.375	0.521
SegFormer-B0	0.954	0.971	0.639	0.782
iFormer	0.961	0.975	0.693	0.810
Gasformer-B0	0.986	0.989	0.894	0.915

Table 3: Supplementary baselines. (a)

\Psi

-fairness: baselines re-trained with

\Psi_{t}

as 4th channel. (b) Temporal classification baselines on 104 clips.

\ddagger

: SegFormer-B0 features;

\star

: end-to-end video model.

	Components					Segmentation				Classification
Variant	TGAA	$\Psi$	ATF	S2	E2E	mIoU	Dice	BF1	HD $\downarrow$	Acc.	F1	$\kappa$	Gini
A4 — No TGAA	✗	✗	✓	✓	✓	0.933	0.964	0.495	5.03	0.503	0.430	0.355	0.830
A6 — No $\Psi$	✗ ‡	✗	✓	✓	✓	0.965	0.982	0.725	2.81	0.769	0.738	0.654	0.824
A3 — No Temporal	✓	✓	✗	✗	✓	0.972	0.981	0.891	1.67	0.573	0.532	0.294	0.752
A8 — No E2E	✓	✓	✓	✓	✗	0.981	0.985	0.923	1.45	0.521	0.461	0.275	0.414
A2 — Concat	✓	✓	✗ †	✓	✓	—				0.721	0.684	0.591	0.770

Table 4: Ablation study. ✓/✗ indicate whether a component is enabled. A4 replaces TGAA with SegFormer-B2 (24.7 M); A6 removes

\Psi

(standard MiT-B0).

\downarrow

= lower is better; bold = best.

Agricultural Image Analysis

Agricultural image analysis uses computer vision and multimodal learning to monitor crops, livestock, produce quality, and field conditions.

See all 94 papers

CowCare: A Unified Visual Framework for Viewpoint-Robust Cattle Tracking and Spatiotemporal Heat-Stress Analysis
Oct 10, 2026Xin Liu, Weiran Li, Minghao Zhu +2Precision Livestock Farming Agricultural Image Analysis
College of Information and Electrical Engineering, China Agricultural University, Beijing, 100083, China · National Innovation Center for Digital Fishery, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China · Key Laboratory of Smart Farming Technologies for Aquatic Animal and Livestock, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China +1
Fully Automatic Trace Gas Plume Detection
May 5, 2026Vít Růžička, David R. Thompson, Jay E. Fahlen +13Remote Sensing
Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena CA, USA. · Independent researcher, Houston TX, 77041, USA. · Carbon Mapper, 680 E Colorado Blvd, Pasadena, CA, 91101, USA. +1
PuTR-CouT: Counting-by-Tracking in Camera-Trap Image Sequences
Sep 4, 2026Fagner Cunha, Juan G. Colonna, Eulanda M. dos SantosObject Counting Wildlife Monitoring
Federal University of Amazonas Manaus, Amazonas, Brazil

Oct 10, 2026J/K move · Enter open · S save

CowCare: A Unified Visual Framework for Viewpoint-Robust Cattle Tracking and Spatiotemporal Heat-Stress Analysis

Xin Liu, Weiran Li, Minghao Zhu, Mina Han+1

College of Information and Electrical Engineering, China Agricultural University, Beijing, 100083, China · National Innovation Center for Digital Fishery, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China · Key Laboratory of Smart Farming Technologies for Aquatic Animal and Livestock, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China+1

Precision Livestock Farming Agricultural Image Analysis Multi-Object Tracking Video Understanding

Open paper AskPDF arXiv

Reliable individual monitoring is essential for dairy cattle welfare and early heat stress detection in Precision Livestock Farming (PLF). However, translating raw multi-camera videos into automated diagnostics remains challenging due to disconnected tracking and recognition pipelines and complex physical environments. Dense mutual occlusions and cross-camera viewpoint variations fragment identities, while environmental noise, including camera vibrations, and limited agricultural data obscure subtle heat stress cues such as rapid breathing and flank undulations. To bridge low-level visual features and high-level behavioral diagnostics, we propose CowCare, a unified framework combining viewpoint-robust tracking with spatiotemporal heat stress recognition. Its tracking module, CowTrack, maintains cross-camera identity consistency through a View-Decoupled Aggregation Network (VDA-Net). Instead of relying on static global representations, VDA-Net aggregates localized textures under heavy occlusion and uses a Viewpoint Decoupling Transformer (VDT) to suppress transient viewpoint variations. CowTrack further incorporates a causal online self-supervision mechanism for continuous adaptation to dynamic farm scenes. The resulting tracklets are subsequently processed by TSDF-Net, a fine-grained video understanding network. Its Temporal Stabilization and Aggregation Module (TSAM) suppresses high-frequency jitter, while a Diffusion Semantic Prior (DSP) uses frozen pre-trained diffusion semantics to enhance subtle behavioral signals and mitigate overfitting. Extensive evaluations on two real-world farm benchmarks demonstrate the framework's effectiveness: CowTrack achieves 83.48% HOTA and 87.72% IDF1 for global multi-object tracking on MVCow, and TSDF-Net achieves 87.72% Top-1 accuracy on HSCow. These results establish CowCare as a robust end-to-end framework for automated livestock health monitoring.

Abstract

Agricultural Image Analysis

Explore similar work

CowCare: A Unified Visual Framework for Viewpoint-Robust Cattle Tracking and Spatiotemporal Heat-Stress Analysis

Fully Automatic Trace Gas Plume Detection

PuTR-CouT: Counting-by-Tracking in Camera-Trap Image Sequences