Detection of Avian Influenza-Infected Chickens using a Multi-Modal Audio-Visual Convolutional Neural Network Optimized for Indian Poultry Farms

Abstract

Avian influenza (AI) poses a recurring threat to poultry production and public health in India, where high-density farming and diverse environmental conditions create challenges for early detection. This study presents a novel Multi-Modal Chicken Health Monitoring System (MM-CHMS) that integrates audio-based spectral analysis of chicken vocalizations with visual-based posture and movement monitoring to detect AI infections. Our architecture combines Mel-spectrogram Convolutional Neural Networks (CNNs) for sound analysis with a ResNet-50 backbone for video feature extraction, fused through a fully connected feature integration layer. The system is designed for field deployment in Indian poultry farms, addressing constraints such as variable lighting, high background noise, and breed-specific vocal variations. A dataset was collected from farms in Goa, Karnataka, and Maharashtra, including both healthy and AI-infected flocks confirmed via RT-PCR tests. The proposed system supports both real-time edge-device inference and offline batch analysis. Field evaluations demonstrate an overall detection accuracy of 94.2 percent, with AI infections detected up to 48 hours prior to visible symptoms, enabling timely intervention.