Attention-based pretrained deep learning framework for nutrient deficiency diagnosis in oilseed rape using UAV multispectral imagery

Timely and accurate diagnosis of crop nutrient deficiencies using imagery is essential for advancing precision agriculture. Conventional approaches relying on RGB imagery are limited by their narrow spectral range, whereas multispectral (MS) imagery offers richer spectral information for crop monitoring. However, transferring deep learning models trained on RGB data to the MS domain is problematic due to spectral-domain shifts. To address this issue, we developed a deep transfer learning framework for classifying nitrogen, phosphorus, and potassium deficiencies in oilseed rape using unmanned aerial vehicle (UAV) imagery. UAV imagery was collected from four experimental fields in Wuhan and Wuxue, China, between 2021 and 2025. Representative deep learning architectures (VGG16, ResNet50, GoogLeNet, and MobileNetV2) were evaluated on both RGB and MS data, followed by multi-strategy transfer learning and the integration of a channel attention mechanism to enhance cross-modal feature representation. The results indicate that GoogLeNet consistently achieved the highest classification accuracy across both modalities. MS-based classification (F1-score = 0.849) outperformed RGB-based classification (0.831). The fine-tuning (FT) strategy further improved F1-scores to 0.844 (RGB) and 0.865 (MS), while reducing trainable parameter count by 76.3% and accelerating model convergence. Incorporating a Squeeze-and-Excitation (SE) module increased the MS F1-score to 0.914 and shortened training time by 35.7%. On an independent cross-year validation, the optimal MS-based model achieved a competitive F1-score of 0.787, demonstrating reliable generalization. Overall, this study highlights the promising potential of deep transfer learning for efficient, robust nutrient-stress monitoring in precision agriculture.