A Compact and Efficient 7.9 Million Parameter Machine Learning Model PD36-B for Real-Time Plant Disease Detection: A Case Study

Deep learning has markedly advanced image-based plant disease diagnosis as improved hardware and dataset quality have enabled increasingly accurate neural models. This paper presents PD36-B, a compact convolutional neural network (7.9 M parameters, 30.2 MB, 20 layers) for multi-class plant disease classification across 38 disease categories spanning 14 crop species. Trained with TensorFlow/Keras on the New Plant Diseases Dataset (87k images, 38 classes), PD36-B employs a five-block progressive convolutional hierarchy (filter depth 32->512), mixed same/valid padding, and two-stage dropout regularization (r1=0.25 and r2=0.4) to achieve robustness and edge-deployability. Training accuracy reached 98.18% by epoch 10, and average test accuracy was 96.57% over 38 classes. An ablation study confirms that each architectural component: dropout regularization, convolutional depth, and dense head, contributes to the final accuracy. Per-class analysis reveals uniformly strong performance: precision/ recall 0.9014/0.9366 for the most challenging class (Corn Cercospora/Gray leaf spot), and 0.9978/0.9874  for the best case (Cherry (including sour) - Powdery mildew), indicating low false positives and strong coverage. Compared against lightweight baselines (MobileNetV2, MobileNetV3-Small, EfficientNet-Lite0), PD36-B achieves competitive accuracy without relying on ImageNet pre-training, making it suitable for offline, plant disease detection in smart agriculture. The model is integrated into a Qt-for-Python desktop application providing real-time inference with disease description and treatment recommendations. These results demonstrate that a carefully designed compact CNN can achieve accuracy competitive with recent baselines while remaining practical for resource-constrained edge scenarios in smart agriculture.