IDPDC — Plant Disease Classification
Interpretable, lightweight EfficientViT-based framework for plant disease classification, benchmarked against CNN and transformer baselines. Paper under review.
Overview
Research code for developing efficient, interpretable deep learning models for plant disease identification on the PlantVillage dataset. The accompanying paper is under review.
My role
Researcher
Problem
Plant disease identification needs models that are accurate, efficient enough to deploy, and interpretable.
Approach
A model factory builds ResNet-50/101, EfficientNet-B0/B3, ViT-B/16, Swin-T and EfficientViT models, trained with AdamW, cosine-annealing learning rate, mixed precision, gradient clipping and early stopping. PlantVillage (38 classes) uses a 70/20/10 split with rotation, flip, color-jitter and random-resized-crop augmentation. Evaluation is repeated over multiple runs to report confidence intervals, confusion matrices and per-class performance.