Smart Disease Detector in Crops: A Deep Learning-Based Approach for Automated Plant Disease Identification

Abstract

This paper presents a practical pipeline for image-based plant disease detection that operates without pre-trained weights or internet access. A compact Efficient Net-like convolutional network is trained from random initialisation on the Plant Village dataset (∼ 41,000 snapshots, 15 classes) at 224 × 224 resolution. Adam, Intensive Augmentation, Reduce LROn Plateau, Early Stopping, and the sampling of the optimal checkpoint to stabilise convergence and avoid overfitting are used to optimise the geometry. Despite early validation oscillations typical of de novo training, the model converges reliably and attains 98.4%–98.9% accuracy on held-out data, with a highly diagonal confusion matrix and uniformly strong per-class precision/recall. These results show that carefully tuned schedules and lightweight regularization can substitute for transfer learning when bandwidth or policy constraints prevent downloading external backbones, enabling classroom, extension, and field deployments. The paper details dataset preparation, architecture and training choices, learning curve behavior, and common failure modes, and concludes with limitations and a roadmap for higher-resolution, field-domain generalization and efficient on-device inference.

Plant phenotyping relevance

植物病害状態を画像から推定する深層学習パイプラインが研究の中心であり、モデル、データ準備、学習、検証性能が具体的に記述されているため。

Code and data availability

The paper's phenotyping input is the public PlantVillage Kaggle dataset (15-class pepper/potato/tomato subset), explicitly cited with the allowed Kaggle URL. No author code or model checkpoints are released.