Leakage-aware benchmark of deep learning for underwater propeller blade-count recognition: code, per-fold results, and model checkpoints

Complete companion artifact for a leakage-aware re-evaluation of deep learning on the public propeller blade-count dataset of Yaman et al. (Applied Acoustics 2021). Six classifiers (EfficientNet-B0, ResNet-50, MobileNetV3, DenseNet-121, a from-scratch multi-scale CNN, a raw-waveform 1D-CNN) and a feature-based SVM are evaluated under two 10-fold protocols: the established stratified segment-level protocol and a blocked recording-level protocol with purging. Model selection uses innervalidation only. The deposit contains the full experiment code, the per-fold machine-readable results behind every number in the manuscript (including per-epoch training histories and configurations), the best-fold model checkpoints for all models and ablation variants, and the statistics script for the paired tests. Headline result: a 0.46M-parameter raw-waveform 1D-CNN is the strongest and most temporally robust model, while segment-level evaluation conceals a 20-27 point collapse of ImageNet transfer models on the first temporal block. The audio recordings themselves are not redistributed; users must obtain them from the source repository and cite Yaman et al. (2021) alongside this record. Version 2 adds eleven ablation sweeps in total, each a complete ten-fold run with a single factor changed. Only the input representation proves to matter: removing log compression costs 19.80 accuracy points (p = 0.001) and halving the mel resolution costs 1.53 points (p = 0.014). ImageNet pretraining, the design of the classification head, and every augmentation operation alone or combined produce no significant change. The deposit grows from 17 to 24 per-fold ledgers and from 15 to 22 checkpoints. Keywords: underwater acoustics; propeller recognition; data leakage; cross-validation; deep learning benchmark