Work

Resource-constrained hybrid attention-driven approach for enhanced interpretability and scalability in multi-event livestock condition classification and monitoring

Year: 2025

Type: article

Abstract: • Proposes CATT-Net framework by integrating PCA, LSTM and Transformer encoders. • Achieves high accuracy (99.74 %) across four diverse datasets excelling SOTA studies. • Optimizes for real-time use with 1.72 MB model, 144.15k parameters and 3.77 MFLOPs. • Incorporates SHAP-based explainability, enhancing model interpretability. • Validated on real agrotech data in Bangladesh with extensive ablation experiments. Precision Livestock Farming (PLF) enhances dairy farming by integrating artificial intelligence (AI) and sensor technologies for health, productivity, and welfare monitoring. However, existing methods remain resource-intensive, difficult to scale, and limited in handling diverse livestock events. To address these challenges, we propose CATT-Net, a lightweight and interpretable hybrid deep learning framework for multi-event livestock condition classification, covering health-related, reproductive, and stress-induced behavioral events. It integrates Principal Component Analysis (PCA) for dimensionality reduction, Squeeze-and-Excitation (SE) blocks for feature recalibration, LSTMs for temporal dynamics, Transformer encoders for global representation, and an Additive Attention mechanism for focused decision-making. Evaluated on four diverse datasets from experimental and commercial farms, CATT-Net achieves high accuracy (99.74 %, 96.10 %, 98.81 %, and 95.85 %) while maintaining efficiency (1.72 MB, 144k parameters, 3.77 MFLOPs), enabling deployment in resource-limited settings. SHAP-based analysis ensures transparency, and ablation studies confirm the contribution of each component. Compared to state-of-the-art transformer models and simpler baselines, CATT-Net achieves higher accuracy while requiring substantially less computational cost. External validation using real-world data from Adorsho PraniSheba, Bangladesh, further demonstrates its practical applicability for livestock monitoring. Code availability: https://github.com/fariya17/CATT-Net

Source: Results in Engineering

DOI: 10.1016/j.rineng.2025.107391

Authors: Fariya Bintay Shafi, Md. Faysal Ahamed, Amith Khandakar, Wesam Rohouma, Mohamed Arselene Ayari, Mamun Bin Ibne Reaz

Institutions: Rajshahi University of Engineering and Technology, Qatar University, University of Doha for Science and Technology, Universiti Malaysia Perlis

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $1,940