CarboNeXT and CarboFormer: Dual Semantic Segmentation Architectures for Detecting and Quantifying Carbon Dioxide Emissions Using Optical Gas Imaging

2025-05-23

arXiv.org

AI Summary

To address the need for real-time, lightweight semantic segmentation of CO₂ plumes in optical gas imaging (OGI), this paper proposes two novel architectures: CarboNeXT and CarboFormer. Methodologically, we introduce a joint modeling paradigm integrating multi-scale contextual aggregation, the UPerHead decoder, and an auxiliary FCN branch, alongside a lightweight Transformer design optimized for edge deployment. Our contributions include: (1) establishing the first OGI benchmark dataset covering both controlled-release (CCR) and real-time cattle rumen emission (RTA) scenarios; (2) achieving 88.46% and 92.95% mIoU on CCR and RTA with CarboNeXT at 60.95 FPS; and (3) attaining 84.88% and 92.98% mIoU with CarboFormer—only 5.07M parameters—at 84.68 FPS, significantly improving detection accuracy for low-flow leaks. These models establish a new paradigm for carbon emission monitoring under resource-constrained conditions.

Technology Category

• Computer Vision: Segmentation
• Search and Optimization: Learning to Search
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Sustainability and carbon-aware systems for Web, mobile, and WoT
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Carbon dioxide (CO$_2$) emissions are critical indicators of both environmental impact and various industrial processes, including livestock management. We introduce CarboNeXT, a semantic segmentation framework for Optical Gas Imaging (OGI), designed to detect and quantify CO$_2$ emissions across diverse applications. Our approach integrates a multi-scale context aggregation network with UPerHead and auxiliary FCN components to effectively model both local details and global relationships in gas plume imagery. We contribute two novel datasets: (1) the Controlled Carbon Dioxide Release (CCR) dataset, which simulates gas leaks with systematically varied flow rates (10-100 SCCM), and (2) the Real Time Ankom (RTA) dataset, focusing on emissions from dairy cow rumen fluid in vitro experiments. Extensive evaluations demonstrate that CarboNeXT outperforms state-of-the-art methods, achieving 88.46% mIoU on CCR and 92.95% mIoU on RTA, with particular effectiveness in challenging low-flow scenarios. The model operates at 60.95 FPS, enabling real-time monitoring applications. Additionally, we propose CarboFormer, a lightweight variant with only 5.07M parameters that achieves 84.68 FPS, with competitive performance of 84.88% mIoU on CCR and 92.98% on RTA, making it suitable for resource-constrained platforms such as programmable drones. Our work advances both environmental sensing and precision livestock management by providing robust tools for CO$_2$ emission analysis, with a specific focus on livestock applications.

Problem

• Detecting and quantifying CO2 emissions using Optical Gas Imaging
• Modeling local details and global relationships in gas plume imagery
• Achieving computational efficiency for resource-constrained environments

Innovation

• Lightweight encoder-decoder architecture for segmentation
• Multi-scale feature fusion with auxiliary supervision
• Optimized for real-time performance on constrained devices