Spectral-Based Paddy Crop Health Monitoring Using Dual Vegetation Indices and Swin Transformer

Paddy crop health monitoring is fundamental to achieving agricultural productivity and food security. Conventional field-based inspection methods are labor-intensive, subjective, and impractical for large-scale cultivation. This paper presents an automated system that leverages Sentinel-2 satellite imagery, dual vegetation indices (NDVI and EVI), and an ensemble of deep learning architectures—CNN, ResNet-50, EfficientNet-B3, and Swin Transformer—to classify paddy crop conditions as Healthy, Moderate Stress, or Severe Stress. Real-time weather parameters (rainfall, temperature, humidity) are integrated with a rule-based recommendation engine to deliver context-aware alerts and actionable advisories for farmers. Experimental evaluation on paddy fields in Chidambaram, Tamil Nadu demonstrates that the Swin Transformer achieves the highest validation accuracy of 97.95%, with an F1-score of 97.91%, outperforming all baselines. The Swin Transformer’s hierarchical shifted-window self-attention mechanism enables fine-grained local feature extraction at multiple spatial scales, making it particularly effective at capturing subtle spectral transitions in NDVI and EVI patches that correspond to early stress signatures in paddy canopies—a capability that standard CNNs, which rely on fixed local receptive fields, cannot achieve. The proposed system enables timely early-stress detection, reduces manual monitoring dependency, and supports precision agriculture at scale.