The detrimental effects of harmful algal blooms (HABs) on the marine ecosystem, human health, and shellfish and aquaculture industry are well known. Anthropogenic activities have led to an increase in frequency, extent and magnitude of HAB activity. As a result, the detection, monitoring and forecasting of HABs are key to agencies and marine managers, allowing them to implement prevention and remediation strategies. However, HAB events are relatively rare events, that can be challenging to detect. HAB detection with satellite image data improves the coverage and efficiency of tracking HABs. Existing remote sensing-based methods frequently rely on statistical classification algorithms. While comparison with cell concentration in situ data has identified two issues: reduced accuracy for the detection of certain species, and accuracy dependency with satellite training data availability. This talk will present a deep-learning technique to improve the performance of the existing models for HAB detection from ocean colour. To this end, we developed a Machine Learning (ML) system using a few-shot learning approach for the detection of Phaeocystis and Pseudo-nitzschia HABs across the French-English channel. We assessed the performance of the ML model in comparison to in situ cell abundance data. The ML system showed better performance than the S-3 EUROHAB model, with results for the detection of Phaeocystis blooms being particularly promising.