Wheat, a vital food source across many regions, faces significant threats from various diseases during cultivation, such as rust and leaf spots, leading to substantial crop damage within a short period. Traditional disease detection methods rely on continuous monitoring by experts familiar with wheat leaf characteristics, a labor-intensive and costly process. This study intends to focus on these challenges via developing a robust automated system for detecting different sorts of wheat leaf diseases and distinguishing healthy crops from them, thereby minimizing yield losses. To achieve this goal, we utilize a dataset having 3 classes of wheat comprising a healthy class and other two classes with diseases: brown rust and yellow rust. Our proposed approach involves leveraging an EfficientNet-based feature extraction model, which relies on extracted features from various deep transfer learning models after preprocessing the dataset. A modification to the EfficientNetB0 architecture with 237 layers is proposed to enhance the classification of wheat leaf diseases. Additional layers including ‘AveragePooling2D with a pool size of (5, 5),’ ‘Flatten,’ ‘Dense with 512 units and Leaky ReLU activation,’ ‘Batch Normalization.’ ‘Dropout,’ and a ‘Custom Dense’ layer are introduced. These layers are strategically inserted into the EfficientNetB0 architecture to improve feature extraction and distinguishability. Extensive testing is conducted using a Kaggle dataset, and the proposed model demonstrates superior performance compared to VGG16 and ResNet101 in experimental evaluations. Notably, the suggested approach achieves an outstanding identification accuracy of 99%.