A morphologically diverse freshwater microalgae dataset for deep learning-based classification with transfer learning analysis

Accurate identification of freshwater microalgae is critical for early detection of harmful algal blooms, water quality monitoring, and sustainable biofuel production—applications that are increasingly important as climate change intensifies pressures on aquatic ecosystems. However, progress in automated classification has been constrained by the scarcity of domain-specific microscopy datasets, forcing most studies to rely on ImageNet-based transfer learning despite fundamental domain mismatches. To address this gap, we introduce the Light Microscopy Freshwater Microalgae (LMFM-12) dataset, comprising 7555 curated images from 12 morphologically diverse species—including unicellular, colonial, and filamentous forms—captured under multiple magnifications. This represents the largest publicly available freshwater microalgae light microscopy dataset to date. Comprehensive evaluation of seven convolutional neural network architectures reveals that randomly initialized models achieve high accuracies (>97%), approaching the performance of fully fine-tuned ImageNet-pretrained networks. Representational similarity analysis demonstrates that random initialization develops distinct representational strategies that contrast sharply with ImageNet-adapted features. Despite comparable accuracy, these divergent approaches suggest that effective microalgae classification arises from learning specialized microscopic features rather than adapting generic visual patterns. Cross-domain evaluation indicates that while ImageNet pretraining achieves superior generalization, Grad-CAM++ analysis highlights distinct attention patterns between ImageNet and LMFM-12 initialization strategies. These results position LMFM-12 as a valuable resource for advancing automated microalgae classification research. • Largest publicly available freshwater microalgae dataset (LMFM-12) • Lightweight CNNs achieve >98% accuracy without ImageNet pretraining • SVCCA reveals distinct domain-specific representational strategies • Cross-domain evaluation shows different model attention strategies