Maize is a major staple food crop worldwide. However, its yield is threatened by foliar diseases, such as northern corn leaf blight, common rust, and gray leaf spot, due to which timely disease detection is essential to ensure food safety and minimize pesticide usage. We propose a hybrid deep-learning model by combining a CNN trunk (7 & 7 convolution, batch normalization, ReLU) with EfficientNet inverted-bottleneck blocks, along with Efficient Channel Attention (ECA) to improve cross-channel feature calibration in a lightweight model with a competitive approach for 4.5M parameters. PlantVillage maize recall images were resized to 224&224 (Maize-leaf-PV) and augmentation (rotations, flips, zooms, brightness, and channel shifts) to mimic field diversity. The model was trained using the Adam optimizer with a cosine-annealing schedule for 350 epochs on TPU pods, batch size 64, dropout rate 0.5, early stopping, and model checkpointing to avoid overfitting. The model achieved an accuracy of 98.85 with F1-scores of 93.37 (Blight), 99.45 (Common Rust), 89.57 (Gray Leaf Spot), and 99.90 (Healthy). These are significantly better than state-of-the-art baselines, including ResNet-50 (95.58%), Inception-v3 (95.99%), and EfficientNetB2 (94.00%), while decreasing the computational cost and model memory footprint. The lightweight structure is conducive to real-time application in a precision-agriculture system and allows targeted response to reduce crop yield losses and environmental harm. Towards implementation: robustness validation under diverse field conditions with UAV and smartphone imaging, transfer learning for additional crops, ensemble and severity-estimation modules, optimization for mobile-device implementation to democratize access to automated disease monitoring, collaborative farmer feedback loops, and cloud-based deployment pipelines for continuous model improvements.