Cross-Domain Plant Disease Classification Reliability Benchmark

Kun Xiang's dataset quantifies the reliability gap for machine learning models transferring from controlled lab to field conditions. It contains results from a benchmark study applying techniques like temperature scaling and selective prediction to a ResNet-50 model fine-tuned on PlantVillage and evaluated on PlantDoc. The dataset was last updated on 2026-05-22 and is shared under a CC-BY-4.0 license.

Use Cases

• Benchmarking model calibration techniques like temperature scaling based on reported ECE scores.
• Evaluating selective prediction methods for plant disease classifiers based on reported risk-at-coverage metrics.
• Comparing domain adaptation baselines like adaptive batch normalization based on reported accuracy improvements.
• Analyzing the impact of image-level domain shift based on reported effect sizes for saturation and edge density.

Strengths

• Includes specific, quantitative results such as a 67.7-percentage-point accuracy drop upon domain transfer.
• Reports detailed metrics like a calibrated ECE of 0.3645 after temperature scaling and selective risk of 64.30% at 80% coverage.
• Benchmarks multiple techniques and backbones, including DINOv2-S/14 achieving an unknown-detection AUROC of 0.764.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is very small at 1.4 KB, suggesting it contains summary statistics rather than raw image data.

Provenance

Source

figshare, author Kun Xiang.

Collection Method

Results from a benchmark study evaluating models trained on PlantVillage and tested on PlantDoc under a parent-image-aware split protocol.

Freshness

Last updated 2026-05-22 06:11:02; freshness should be verified.