Deep Learning-Based Phenotyping of Lettuce Diseases Using Efficient-FBM-FRMNet for Precision Agriculture

Parul Nasra¹Sheifali Gupta¹Mudassir Khan^2*Jaibir Singh^3,4,5Bayan Alabdullah⁶Abrar Almjally⁷Ruby Pant⁸Nitin Kumar⁹Salil Bharany^1*

• ¹Chitkara University, Chandigarh, India
• ²King Khalid University, Abha, Saudi Arabia
• ³Lovely Professional University, Phagwara, India
• ⁴Galgotias university, noida, India
• ⁵Galgotias university, greater noida, India
• ⁶Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia
• ⁷Imam Muhammad Ibn Saud Islamic University, Riyadh, Saudi Arabia
• ⁸Uttaranchal University, Dehradun, India
• ⁹Graphic Era Deemed to be University, Dehradun, India

Lettuce (Lactuca sativa), a widely cultivated leafy vegetable, is highly susceptible to bacterial and fungal infections that significantly reduce crop yield. Timely and accurate disease identification is therefore crucial for precision agriculture. This study introduces Efficient-FBM-FRMNet, a deep learning framework for automated lettuce disease detection. The proposed architecture combines EfficientNetB4 with dilated convolutions, a Feature Bottleneck Module (FBM), a Reasoning Engine, and a Feature Refinement Module (FRM) in a sequential manner that strengthens both accuracy and interpretability. EfficientNetB4 with dilated convolutions captures fine-grained and multi-scale lesion patterns by expanding the receptive field without increasing parameters. The FBM then condenses redundant features into compact, noise-reduced representations that emphasize lesion-specific cues. These condensed features are processed by the Reasoning Engine, a lightweight non-linear projection that models higher-order feature interactions and enhances semantic separability. Finally, the FRM This is a provisional file, not the final typeset article calibrates and regularizes the feature space, suppressing overfitting and improving prediction stability. This integration allows the framework not only to achieve superior accuracy with fewer parameters but also to provide interpretable, lesion-focused predictions that conventional CNNs lack. The proposed model was evaluated on a publicly available dataset of 2,813 lettuce leaf images covering bacterial, fungal, and healthy classes, it achieved high accuracy, precision, and recall which consistently outperformed conventional CNN architectures such as EfficientNetB4, ResNet50, and DenseNet121. These improvements highlight the framework's superiority in discriminative capability, interpretability, and stability. The results underscore the potential of Efficient-FBM-FRMNet for deployment in greenhouse monitoring, UAV-assisted field surveillance, and mobile diagnostic platforms, contributing to sustainable, AI-driven agricultural practices.