MCDCNet: a multimodal cotton disease classification model for visually variable leaf symptoms under natural field conditions

Cotton is typically cultivated on a large scale. Unmanned aerial vehicle (UAV) can efficiently acquire large scale cotton field images for disease classification tasks. However the acquired images are strongly affected by natural environmental conditions. The images usually contain lesion appearance variation uneven illumination and complex background conditions. To address this problem this study proposes a multimodal cotton disease classification model named multimodal cotton disease classification network (MCDCNet) for natural field environments. In this work cotton leaf images were acquired using UAV. Soil temperature and humidity data from the corresponding regions were continuously collected using environmental sensors. In the proposed model, we innovatively integrate discrete wavelet transform and axial attention into the Swin Transformer to improve disease classification performance. Meanwhile we evaluated multimodal feature fusion at different stages and selected a weighted mid-level fusion strategy for integrating cotton leaf image features with environmental textual data. Experimental results demonstrated that the proposed multimodal model achieved significantly superior classification performance under complex field conditions, with a precision of 98.7%, recall of 98.4%, and an F1-score of 98.5%. In conclusion, The proposed multimodal MCDCNet framework provides an effective solution for intelligent agricultural disease monitoring under natural field conditions.