A multi-scale supervised contrastive framework for cross-domain soybean disease classification using leaf and UAV...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Precision agriculture faces challenges in soybean crop health monitoring due to environmental variability and differing image scales (ground-level vs. UAV).
• Existing deep learning models often process ground and aerial imagery separately, failing to bridge the domain gap and leverage accessible leaf-level data for aerial-scale improvements.

Purpose of the Study:

• To develop a multi-scale soybean crop health assessment framework integrating ground-level and UAV imagery.
• To address the domain shift between sensing scales and improve the transferability of features from leaf-level to UAV-level analysis.

Main Methods:

• A structured pre-processing pipeline (CLAHE, Gray-World, illumination normalization) was applied to reduce illumination bias.
• Six deep learning backbones were evaluated for leaf-level classification, with MaxViT and ConvNeXt showing top performance.
• A supervised contrastive learning framework was implemented for cross-scale feature alignment, improving feature discriminability and reducing domain discrepancy.

Main Results:

• The static weighted ensemble of MaxViT and ConvNeXt achieved 87.08% accuracy for leaf-level classification.
• Zero-shot transfer from leaf-level to UAV-level yielded 40% accuracy, indicating a significant domain shift.
• Fine-tuning improved UAV classification to 97%, while supervised contrastive learning further boosted accuracy to approximately 98% with enhanced stability.

Conclusions:

• Supervised alignment effectively bridges the domain gap between ground-level and UAV imagery in soybean crop health monitoring.
• The proposed framework generates more class-discriminative representations, enabling scalable multi-scale cross-health monitoring with high accuracy.