Coffee is a major agricultural export crop in Vietnam, yet leaf diseases such as Leaf Rust, Cercospora leaf spot, Phoma leaf spot, and Leaf Miner cause significant yield losses each year. Early and accurate identification of these diseases remains a practical challenge for smallholder farmers. This study proposes an improved lightweight convolutional neural network (CNN) architecture for automated detection of coffee leaf diseases from smartphone photographs. Built upon a truncated EfficientNet-B0 backbone, the model integrates a dual attention mechanism combining Efficient Channel Attention (ECA) and Coordinate Attention (CA), a Ghost Transition Block for efficient feature expansion, and a Fusion Pooling head that merges Global Average Pooling and Global Max Pooling outputs. A multi-stage preprocessing pipeline, encompassing Otsu-based leaf segmentation, random background substitution with agricultural imagery, and Contrast Limited Adaptive Histogram Equalization (CLAHE), is applied to reduce background noise. Training employs Class-Weighted Focal Loss, MixUp augmentation, and a three-stage progressive fine-tuning strategy to address the severe class imbalance present in the BRACOL dataset. Evaluated via 5- Fold Stratified Cross-Validation on 1,685 annotated images from BRACOL, the proposed model achieves a mean accuracy of 94.97% and macro F1-Score of 89.26% with only approximately one million parameters and 0.52 GFLOPs, a fourfold reduction in parameters compared with the original EfficientNet-B0, while remaining competitive with the larger ResNet50 baseline (95.32%). After quantization and conversion to TensorFlow Lite, the model is deployed in a cross-platform mobile application built with React Native, achieving an inference time of approximately 65.59 ms per image and a compressed model size of 1.2 MB. The proposed system provides a practical, offline-capable tool for field-level coffee disease monitoring.

Coffee leaf disease, Lightweight CNN, Attention mechanism, Ghost module, BRACOL dataset