Enhancing Plant Disease Detection in Smart Agriculture Using Deep Learning Models and Advanced Data Augmentation Techniques

Abstract

Deep learning for automatic detection of plant diseases has emerged as an innova- tive method to address agricultural productivity and food security issues. While deep CNNs have achieved near-ceiling performance on benchmark datasets under controlled conditions, the results do not necessarily translate to real-world agricultural environments. The reason for this is domain shift: models are typically trained on clean laboratory images. Field images include background clutter, illumination variations, occlusions, and inconsistent capture conditions [1, 2, 10, 11]. This thesis investigates cross-domain plant disease classification under a controlled- to-field setting using transfer-learned CNN architectures, namely EfficientNet-B0, MobileNetV3-Large, and ResNet-50 [21, 22, 20]. PlantVillage is used as the source domain for training, while PlantDoc is used as the target domain for evaluation and adaptation. The study evaluates a 12-class controlled-to-field transfer task using a harmonized subset of overlapping classes from both datasets. The implemented framework compares four training pipelines: P1, which uses preprocessing only and no augmentation; P2, which applies basic geometric and photometric augmentation; P3, which applies stronger environmental augmentation together with mixed-image regularization using MixUp and CutMix; and P4, which serves as an ablation study by applying P3’s environmental transforms without image mixing. [24, 26, 27]. The results show that source-domain validation performance does not translate into strong field performance. In the zero-shot cross-domain setting, where models are trained on PlantVillage and evaluated directly on PlantDoc without target-domain training, the best configuration reaches only 24.84% accuracy on the filtered PlantDoc FULL evaluation set of 1405 images, confirming a substantial lab-to-field generalization gap. When target-domain supervision is introduced through fine-tuning on the filtered PlantDoc training split and evaluation on the untouched filtered PlantDoc test split, performance improves substantially, reaching up to 59.32% accuracy and 0.5734 macro-F1. The results further show that basic augmentation is more reliable than the most aggressive augmentation setting for zero-shot transfer, while supervised fine-tuning on target-domain data provides the largest reduction in the deployment gap. Furthermore, a target-domain data efficiency ablation reveals that supervised adaptation scales logarithmically. Models recover the majority of their fine-tuned performance using only a fraction of the target dataset, demonstrating that bridging the lab-to-field gap does not strictly require massive field-collected datasets. Overall, this research studies plant disease recognition using field data. It shows that while domain adaptation improves robustness and makes models more deployable, augmenting the source training data alone is not sufficient to solve the controlled- to-field generalization problem.