Towards scalable deep learning for automated microscopy in harmful algal bloom monitoring: Data-centric workflow and multi-region generalisation

Harmful algal blooms (HABs) pose growing risks to drinking-water supplies and ecosystems, yet routine monitoring remains dependent on manual light microscopy. Deep learning offers a potential rapid alternative for automating microscopy, but progress has been constrained by idealised datasets and limited evaluation under real-world conditions. To address these gaps, this study introduces a multi-region microscopy dataset comprising 105 microalgae/cyanobacteria taxa collected from three Australian regions, capturing diversity and challenges of operational monitoring. Given dataset complexity, a state-of-the-art object detection model (YOLOv12) was adopted, and a structured data-centric workflow was developed on one regional dataset, providing a systematic evaluation of dataset design choices in automated microscopy under real-world conditions. The best-performing configuration used an 80/10/10 train/validation/test split, genus-level taxonomic granularity, inclusion of priority taxa, static geometric augmentation, and an input resolution of 640, achieving a mean Average Precision at 0.5 Intersection-over-Union (mAP@0.5) of 0.73 in-domain evaluation. To assess transferability, the best-performing model was evaluated across regions without retraining, where performance declined substantially (mAP@0.5 ≤ 0.10). Limited target-domain fine-tuning provided partial improvement, but performance remained substantially below in-domain levels, indicating cross-region generalisation challenges. To mitigate this, new models were trained on merged datasets from multiple regions, recovering performance to mAP@0.5 of 0.62-0.67 depending on the regions included. Collectively, these results demonstrate that a structured data-centric workflow can substantially enhance automated microscopy, yet domain generalisation remains a critical bottleneck for deployment. This study provides both methodological innovation and empirical insight, advancing progress toward scalable AI systems for rapid HAB monitoring.