Reproducibility package for "Resolution–Compression Trade-Offs in In-the-Wild Plant Disease Classification: Accuracy, Reliability, and Inference Efficiency"

This record provides the reproducibility package for the study “Resolution–Compression Trade-Offs in In-the-Wild Plant Disease Classification: Accuracy, Reliability, and Inference Efficiency.” It contains Python code for data-integrity auditing, duplicate-aware cleaning and deterministic group-aware splitting, CNN training, controlled JPEG evaluation, confidence calibration, selective prediction, GPU efficiency benchmarking, Pareto analysis, paired bootstrap uncertainty analysis, and class-specific degradation/calibration analysis. Processed split indices, duplicate-review decisions, benchmark metadata, and compact numerical result tables are included. PlantWild_v2 source images, trained checkpoints, and full prediction/logit archives are not redistributed. The source dataset must be obtained independently from the original PlantWild repository.