Behind the Research: Optimizing Real-Time Thermal Imaging for Agricultural Automation
Published in Computational Sciences
.png)
A deep dive into benchmarking lightweight YOLO architectures for okra maturity detection
Introduction
Agricultural automation faces a critical challenge: how can we develop inspection systems that are both highly accurate and fast enough for real-time processing? In our recent study published in The Journal of Supercomputing, my colleagues and I addressed this question by systematically benchmarking four state-of-the-art YOLO nano-variant object detection models for thermal imaging applications in okra maturity grading.
Why Thermal Imaging Matters
Traditional RGB-based inspection systems struggle under variable lighting conditions—a common reality in agricultural settings ranging from night harvesting to indoor packing facilities. Thermal imaging offers a compelling alternative by capturing temperature-dependent information that remains consistent regardless of ambient lighting. For okra specifically, thermal signatures reveal physiological differences in moisture content and thermal mass between maturity stages, providing a robust basis for automated quality assessment.
However, thermal imagery presents unique computational challenges. Single-channel intensity data, reduced texture information, and temperature-dependent contrast require specialized architectural considerations that haven't been systematically evaluated in prior research.
The Research Gap
While YOLO (You Only Look Once) architectures have been extensively benchmarked on RGB datasets such as COCO and ImageNet, comprehensive evaluation on thermal agricultural imagery remained absent from the literature. This gap is particularly significant because:
- Architectural innovations designed for RGB may not transfer effectively to thermal domains
- Real-time industrial sorting requires sub-50 millisecond inference latency
- Deployment on heterogeneous computing platforms (GPU vs. CPU) demands quantitative performance analysis
Our study addresses these gaps by evaluating YOLOv5n, YOLOv8n, YOLOv11n, and YOLOv12n across multiple performance dimensions.
Methodology Highlights
We developed a dual-source thermal dataset combining passive and active thermal imaging modalities. The active thermal approach—involving controlled preheating to 30°C—proved particularly important. Natural thermal contrast between adequately matured and overripe okra averages only 2.8°C under ambient conditions, which can be insufficient under variable environments. Controlled preheating amplifies this contrast to 4–8°C by exploiting maturity-dependent differences in thermal mass and cooling behavior.
Our experimental design incorporated rigorous statistical validation across five independent training runs, ensuring reproducibility and significance testing of observed performance differences. This methodological rigor is essential for drawing reliable conclusions in machine learning research.
Key Findings
Training Duration Dependency
One of our most significant findings relates to training efficiency. Under the resource-efficient 10-epoch protocol—reflecting rapid development scenarios—YOLOv8n achieved the highest detection accuracy (66.3% mAP@0.5–0.95), while YOLOv5n delivered comparable performance (66.1%) with superior computational efficiency.
However, extended training experiments revealed a critical insight: attention-based architectures (YOLOv11n and YOLOv12n) achieve higher peak accuracy (73.1% and 72.9% respectively) when training budgets permit 45–67 epochs. This performance ranking inversion suggests that architectural complexity translates to gains only with sufficient training iterations—a crucial consideration for practical deployment.
Platform-Specific Performance
Our benchmarking across heterogeneous computing platforms revealed distinct trade-offs:
- GPU deployment (NVIDIA T4 with TensorRT): YOLOv8n achieved 1.6 ms inference latency, supporting throughput exceeding 625 FPS—well above the ≥20 FPS requirement for real-time sorting
- CPU deployment (ONNX Runtime): YOLOv5n exhibited superior performance at 31.1 ms, making it optimal for edge and embedded scenarios
These findings provide practical guidance for selecting architectures based on deployment constraints.
Architectural Insights
Through gradient-weighted class activation mapping (Grad-CAM) analysis, we discovered that decoupled detection heads enable task-specific feature specialization. Classification branches focus selectively on thermal intensity gradients, while localization branches emphasize geometric boundaries—a design principle particularly beneficial for thermal imagery where diagnostic features are spatially distinct.
Ablation studies confirmed that mosaic augmentation improved detection performance by 6.2% without additional latency, while balanced loss weighting outperformed imbalanced configurations in this binary classification task.
Practical Implications
Our research demonstrates that thermal-based YOLO nano-variants are ready for integration into automated sorting pipelines, achieving 75 kg/min throughput with <2% background false positives. The system operates at sustained rates exceeding industrial requirements while maintaining high detection accuracy.
For practitioners, the key takeaway is this: model selection must jointly consider deployment constraints (inference latency, memory footprint) and available training resources (time budget, computational capacity). Simple architectures like YOLOv5n and YOLOv8n excel in rapid-iteration scenarios, while attention-based variants justify their complexity only when extended training is feasible.
Future Directions
Several promising research directions emerge from this work:
- Dataset expansion across cultivars, seasons, and environmental conditions to enhance model robustness
- Multi-class grading spanning the full maturity spectrum
- Multispectral fusion combining thermal and RGB modalities
- Knowledge distillation from transformer-based teachers to lightweight students
- Field deployment trials in commercial facilities for long-term validation
Conclusion
This systematic benchmark establishes quantitative baselines for thermal agricultural imaging and provides practical guidance for architecture selection. The identified strategies—anchor-free detection, decoupled heads, mosaic augmentation, and balanced loss weighting—are transferable to other crops and sensing modalities, offering strong potential for advancing automated agricultural inspection systems.
Article Citation: Ganapathy, M.R., Pugazhendi, P., Periasamy, S., Nagarajan, B. (2026). Benchmarking YOLO nano-architectures for real-time thermal imaging: application to okra maturity grading on heterogeneous computing platforms. The Journal of Supercomputing, 82:97. https://doi.org/10.1007/s11227-026-08226-w
Data Availability: The public thermal dataset is available at Mendeley Data. Code is available at the project GitHub repository.
I'm Dr. Pathmanaban Pugazhendi, an Assistant Professor at Easwari Engineering College, where I combine my passion for Computer Vision, Thermal Imaging, Deep Learning, and Edge AI to solve real-world problems in agriculture, food safety, and sustainable technology.
My Research Journey
My work has been cited over 410 times across various international journals, with a Total Author Impact Factor of 62.20. What drives me is the opportunity to bridge the gap between cutting-edge AI research and practical applications that make a tangible difference.
In Agricultural Technology & Food Science:
- I develop AI-powered systems for precision agriculture, helping farmers detect diseases and defects in crops like guava, sugarcane, and sapota
- I use thermal imaging to assess fruit quality post-harvest, reducing food waste and improving safety
- My recent work includes uncertainty-aware deep learning models that make agricultural AI more reliable and trustworthy
In Sustainable Engineering:
- I apply machine learning to optimize engine performance using alternative fuels like biodiesel and ethanol-water blends
- I'm working on intelligent battery management systems for electric vehicles
- My research explores how AI can accelerate our transition to renewable energy
In Computer Vision & AI:
- I specialize in developing lightweight deep learning models (like YOLO architectures) that can run on edge devices
- I'm particularly interested in making AI accessible for resource-constrained environments—bringing powerful technology to where it's needed most, not just where computational resources are abundant
What Drives My Work
I believe AI should be practical, deployable, and accessible. That's why much of my research focuses on developing efficient models that work in real-world conditions—whether it's detecting bread contamination using thermal imaging, identifying weld defects in manufacturing, or helping farmers spot crop diseases early.
My work has been published in top-tier Q1 journals including Journal of Food Engineering (IF: 5.8), Scientific Reports (IF: 3.9), and Trends in Food Science & Technology (IF: 15.4), but what matters most to me is seeing these technologies actually being used to solve problems.
Looking Forward
I'm currently focused on advancing Edge AI technologies—developing intelligent systems that can make split-second decisions without needing constant cloud connectivity. From smart farming tools to industrial quality control systems, I'm excited about building AI that works anywhere, for anyone.
I'm always open to collaborations, discussions, and connecting with fellow researchers, students, and industry professionals who share a passion for applied AI and sustainable technology.
Follow the Topic
-
The Journal of Supercomputing
The Journal of Supercomputing publishes papers on the technology, architecture and systems, algorithms, languages and programs, performance measures and methods, and applications of all aspects of supercomputing.
Related Collections
With Collections, you can get published faster and increase your visibility.
SI - Simulation-Powered Innovation: Driving the Future of Digital Ecosystems
This Special Issue publishes cutting-edge research on simulation technologies that are revolutionizing industries, fostering innovation, and enabling seamless interactions within complex digital environments. It is based on the 29th International Symposium on Distributed Simulation and Real Time Applications (DS-RT 2025).
The scope of the SI includes, but is not limited to:
• Advanced Paradigms in Large Scale Distributed and Real-Time Simulations: Exploring innovative methodologies and software architectures for extreme scale simulations, including parallel and distributed simulation, multi-agent systems, and hardware-software co-design.
• Large Scale Real-Time Systems and Concurrent Systems Design: Focusing on paradigms, modeling, and architecture for systems with hard and soft real-time constraints, emphasizing the development of robust and efficient environments.
• Innovative Modelling Techniques for Simulation: Delving into advanced modeling approaches, including the reuse of models, development of new modeling languages, agent-based modeling, and spatial modeling and simulation.
• Multi-Level, Multi-Scale, and Multi-Resolution Modeling: Investigating the complexities and challenges in creating simulations that operate across various levels, scales, and resolutions.
• Non-Functional Properties in Distributed Simulation and Real-Time Systems: Addressing key aspects like dependability, reliability, maintainability, safety, security, and quality of service in simulation and real-time systems.
• Theoretical Foundations of Real-Time and Simulation Models: Exploring fundamental concepts such as event systems, causality, space-time models, notions of time, and the coordination of simulators in discrete and continuous systems.
• Large Scale Simulation Studies: Presenting case studies and research in industrial, commercial, ecological, societal, and pervasive computing systems, focusing on large and very large scale applications.
• Performance and Validation in Large Scale Distributed Simulations: Discussing benchmarking, analytical results, and empirical studies to assess the performance and accuracy of large-scale distributed simulations.
• Parallel and Distributed Simulation Algorithms and Methods: Covering synchronization, scheduling, memory management, and load balancing in the context of parallel or distributed simulation.
• Virtual Environments and Mixed Reality Systems in Simulation: Exploring the use of interactive virtual reality, immersive environments, and human communication in simulation-based virtual and mixed reality systems.
• Collaborative Virtual and Augmented Reality: Investigating shared interaction spaces, telepresence systems, and shared workspaces, including 3D video and acoustic reconstruction.
• Serious Gaming and MMOG in Simulation: Addressing applications, architectures, and scalability issues in serious gaming and massive multiplayer online games.
• Human-Computer Interaction in Large Scale DS-RT Systems: Examining design issues, interaction designs, and the challenges raised by large-scale distributed simulation and real-time systems.
• Artificial Intelligence and Machine Learning in Simulation: Focusing on the integration and application of AI and machine learning techniques within simulation environments.
• Advances in Digital Twin Paradigms: Exploring the latest developments, applications, and challenges in the field of digital twins, emphasizing their role in bridging the physical and digital worlds.
• Intersection of Digital Twins and machine learning: enhancing predictive modeling, real-time analytics, and decision-making, focusing on on novel algorithms and use-cases.
• Role of Digital Twins in applications: enhancing automation, connectivity, and real-time data exchange. Contributions are solicited on implementation challenges and future trends on applications such as Industry 4.0, manufacturing, supply chain, construction, aerospace, logistics, energy sector, healthcare.
Submitted papers should present original, unpublished work, relevant to one of the topics of the special issue. All submitted papers will be evaluated on the basis of relevance, significance of contribution, technical quality, scholarship, and quality of presentation, by at least three independent reviewers. It is the policy of the journal that no submission, or substantially overlapping submission, be published or be under review at another journal or conference at any time during the review process. Please note that the authors of selected papers presented at the DS-RT 2025 symposium are invited to submit an extended version of their contributions by taking into consideration both the reviewers’ comments on their conference paper, and the feedback received during presentation at the conference. It is worth clarifying that the extended version is expected to contain a substantial scientific contribution, e.g., in the form of new algorithms, experiments or qualitative/quantitative comparisons, and that neither verbatim transfer of large parts of the conference paper nor reproduction of already published figures will be tolerated. The extended versions of symposium papers will undergo the standard, rigorous journal review process and be accepted only if well-suited to the topic of this special issue and meeting the scientific level of the journal. Final decisions on all papers are made by the Editor in Chief.
Publishing Model: Hybrid
Deadline: Ongoing
Section - Architectures, Systems and Hardware Security
All aspects of high-performance hardware and architectures, including optimizing and evaluating processors, systems issues, and security, especially at the hardware level and sustainability of systems.
Topics include but not limited to the following:
T
• Architectural support for programming languages or software development.
• Architectures to support extremely heterogeneous composable systems (e.g., chiplets)
• Design-space exploration/performance projection for future systems
• Evaluation and measurement on testbed or production hardware systems
• Hardware acceleration of containerization and virtualization mechanisms for HPC
• Interconnect technologies, topology, switch architecture
• I/O architecture/hardware and emerging storage technologies
• Memory systems: caches, memory technology, non-volatile memory, memory system architecture (to include address translation for cores and accelerators)
• Multi-processor architecture and micro-architecture (e.g., reconfigurable, vector, stream, dataflow, GPUs, and custom/novel architecture)
• Sustainable design aspects, including power and energy efficiency and power-management strategies
• Resilience, error correction, high availability architectures
• Scalable and composable coherence (for cores and accelerators)
• Secure architectures, side-channel attacks, and mitigation, covering all attack vectors, including all forms of side-channel attacks, piracy, reverse engineering, tampering, and hardware Trojan attacks, including countermeasures at different stages of system design - i.e., architecture definition, design, validation, and deployment
• The security of hardware and system security at all levels of abstraction
• Interactions between hardware and systems, and between hardware and firmware/software, including in the context of security and trust
• Software/hardware co-design, domain-specific language support
• Interactions among architectures, compilers, programming languages, and operating systems
Architectures, Systems and Hardware Security research relates to multiple United Nations Sustainable Development Goals (SDGs) through advances in health care, education, and energy, among other fields. This Section particularly welcomes submissions related to SDG 9 “Industry, Innovation, And Infrastructure.”
An essential aspect of supercomputing involves solving computer-intensive problems. Paper submissions are expected to address problems that require significant computational resources.
Publishing Model: Hybrid
Deadline: Ongoing
Please sign in or register for FREE
If you are a registered user on Research Communities by Springer Nature, please sign in