Reproducing a year-round expert survey with deep learning

Custom acoustic classifiers are increasingly used in biodiversity monitoring, yet their performance under limited training data remains poorly understood. We evaluated how sample size influences detection accuracy for the Bangwa Warbler (Bradypterus bangwaensis), an Afrotropical songbird underrepresented in global recognisers. We compared four ResNet architectures (18–152) trained on 320 samples with a custom BirdNET classifier, identifying ResNet18 as the best‑performing model. To assess data efficiency, this model was retrained using four progressively reduced sample size thresholds. Validation on over 141 hours of passive acoustic monitoring recordings showed that the full 320‑sample ResNet18 model achieved F1 = 0.88 ± 0.02 and AUPRC = 0.95 ± 0.01, outperforming deeper architectures (ResNet50, ResNet152) and all reduced‑sample models, which exhibited consistent performance declines. ResNet34 performed comparably to ResNet18. When applied to an independent year-long dataset (293 hours), ResNet18 closely reproduced expert-derived seasonal detection patterns (12,361 vs. 11,856 acoustic events), while BirdNET classifiers achieved high recall but substantially underestimated species presence (F1 = 0.61–0.62). Our findings show that custom classifiers perform well with sufficient labelled data and can reproduce seasonal vocal activity patterns derived from manual surveys, although their performance declines under sparse‑data conditions.