Deep Learning-Based Plant Disease Prediction Using Real-Field Image Data

Plant diseases remain a persistent threat to global agricultural output, annually destroying an estimated 20–40% of total crop yields worldwide. Prompt and reliable identification of infection at early stages is indispensable for guiding timely intervention and protecting food security. Conventional diagnostic workflows, which depend on field visits by trained agronomists, are inherently subjective, time-consuming, and impractical at the scale of modern large-area farming. To address these operational gaps, this study proposes a fully automated deep learning pipeline tailored for classifying plant diseases from photographs collected under uncontrolled, realistic field conditions. The core of the architecture is a hybrid model that couples EfficientNet-B4 [3] with a Convolutional Block Attention Module (CBAM) [4], equipping the network with the capacity to localize and emphasize abnormal leaf tissue while filtering out irrelevant scene elements. The system is developed and benchmarked on a purpose-built dataset of 54,306 labeled images representing 26 disease classes across 14 crop species. A structured preprocessing workflow—encompassing Contrast Limited Adaptive Histogram Equalization (CLAHE), mosaic-based class balancing, and mixup regularization [10]—is incorporated to enhance the model’s tolerance to lighting inconsistencies and skewed class distributions. On the held-out test partition, the proposed model attains a top-1 accuracy of 96.7%, outperforming six well-established CNN baselines. Gradient-weighted Class Activation Mapping (Grad-CAM) [11] visualizations further confirm that the model directs its attention toward pathologically relevant regions, lending credibility to its predictions. These properties collectively position the framework as a strong foundation for lightweight, smartphone-deployable disease advisory tools for smallholder farmers.