Estimating wheat disease severity from high-resolution UAV multispectral imagery using deep learning

Bacterial Leaf Streak (BLS) and Fusarium Head Blight (FHB) are among the most damaging diseases of wheat (Triticum aestivum), with severe consequences for grain yield, quality, and ultimately food safety and security. Rapid and precise assessment of disease severity in the fields is crucial for effective field management, potential yield loss evaluation, and high-throughput phenotyping. This research examined the utility of UAV-based multispectral imagery in combination with both traditional machine learning and modern deep learning approaches to estimate wheat disease severity under field conditions. Data collection was carried out at two wheat experimental fields in South Dakota, USA, where Unmanned Aerial Vehicle (UAV) multispectral imagery was acquired in parallel with plot-level measurements of BLS and FHB severity. Spectral and textural metrics extracted from the UAV imagery served as inputs for machine/deep learning-based regression analyses. Regression models evaluated in this work comprised traditional machine learning methods Partial Least Squares Regression (PLSR), Random Forest Regression (RFR), and three deep learning architectures: Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and multi-head self-attention (MHSA)-enhanced CNN (Att-CNN). In addition, a deep transfer learning framework was tested by transferring an Att-CNN model trained on BLS to FHB severity estimation. The results showed that deep learning methods, particularly CNN-based architectures, consistently outperformed conventional machine learning approaches. Incorporation of a MHSA mechanism into the CNN architecture further enhanced performance, especially for BLS severity estimation. Att-CNN achieved the best results for both diseases, with R² = 0.83 and RRMSE = 30.55 % for BLS, and R² = 0.70 and RRMSE = 37.05 % for FHB. While estimation of FHB severity remained more challenging, transfer learning from BLS substantially improved prediction accuracy, raising R² from 0.70 to 0.79 and reducing RRMSE from 37.05 % to 31.31 %. The study highlights the considerable potential of UAV multispectral imagery, though with notable limitations, for monitoring crop diseases. This work also demonstrates the added value of attention-based deep learning and transfer learning techniques in addressing complex applications in agricultural remote sensing.