Explainability Output Inventory for a 51-Class Plant Disease Image Dataset

An inventory of explainability outputs from a study benchmarking eight tiny vision transformers and two custom CNNs for plant disease classification. The underlying dataset comprises approximately 45,000 original images across 51 disease and healthy classes, aggregated from multiple public repositories. The study was authored by Md. Sifat Haque Zidan and last updated in April 2026.

Use Cases

• Comparing Grad-CAM and Grad-CAM++ visualizations for model explainability based on the qualitative analyses mentioned.
• Evaluating the trade-off between model accuracy and computational cost for plant disease diagnosis based on the benchmarked architectures.
• Developing efficient convolutional neural networks for multi-crop disease classification based on the described PlantaNet and PlantaNetLite models.

Strengths

• Underlying image dataset contains approximately 45,000 original images.
• Models were evaluated on a curated multi-crop dataset spanning 51 disease and healthy classes.
• Best-performing custom model achieved 99.66% test accuracy with a compact 9.85 MB size.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for the inventory data.
• The 5.5 KB file size indicates the inventory is a very small summary, not the primary image data.

Provenance

Source

Aggregated from multiple publicly available image repositories.

Collection Method

Curated for a study benchmarking vision transformer and CNN architectures.

Time Range

null

Freshness

Last updated 2026-04-24 17:24:17; freshness should be verified.

Geography

null