Comparative Performance of Deep Learning Models for Early Detection of Soybean Nutrient Deficiencies Using Aerial Multispectral Imagery

Abstract

This paper presents a comparative performance analysis of three state-of-the-art deep learning architectures—EfficientNet-B4, ResNet-50 with attention gates, and Vision Transformer—for the identification of nine nutrient deficiencies in soybean crops using aerial multispectral imagery. The study evaluates these models across multiple performance dimensions, including classification accuracy, computational efficiency, robustness under field conditions, and growth-stage-specific performance. Results indicate that Vision Transformer achieves the highest accuracy (94.7%) but with significant computational overhead, while EfficientNet-B4 offers the best trade-off for real-time deployment with moderate accuracy (90.8%). ResNet-50 with attention mechanisms provides a balanced performance profile suitable for practical agricultural applications. The findings provide evidence-based guidelines for model selection in precision agriculture systems.