Enhancing satellite-based detection of marine fauna using synthetic data

Satellite imagery could offer a powerful, non-invasive tool for large-scale whale population monitoring. However, its effectiveness is constrained by limited annotated training data, the high cost of very high-resolution imagery, and restrictions on data sharing due to proprietary commercial sources. To address these challenges, we develop an open-source dataset generated using a score-based diffusion model to create realistic, high-resolution synthetic satellite imagery of two ecologically and geographically distinct whale species: southern right whales (Eubalaena australis), including mother–calf pairs, in South African coastal waters, and beluga whales (Delphinapterus leucas) in the Canadian Arctic. We evaluate whether object detection models trained exclusively on synthetic data can generalize to real, unseen satellite imagery. For southern right whales, a model trained solely on synthetic images achieves a mean average precision of 0.86 (mAP@50) when tested on real satellite data. When synthetic and real imagery are combined during training, performance improves further, exceeding 0.95 mAP@50. Similarly, for beluga whales, models trained exclusively on synthetic imagery achieve 0.90 mAP@50, increasing to 0.93 when trained with mixed datasets. Despite marked differences in body size, coloration, and habitat conditions—southern right whales are large, dark-bodied mysticetes in relatively clear coastal waters, whereas belugas are smaller, white odontocetes often found in turbid Arctic environments—our approach performs consistently across species. These results demonstrate cross-species generalizability and identify synthetic imagery as a scalable solution for training deep learning models for whale detection by addressing the large data requirements that are required to adequately train such models, and which have historically limited progress in this domain. • Synthetic satellite imagery generated via diffusion models aids whale detection. • Open-source dataset covers two species: southern right whales and belugas. • Synthetic-only training achieves up to 0.90 mAP@50 across both species. • Combining synthetic and real data pushes performance above 0.93–0.95 mAP@50. • Approach generalises across species despite differences in size and habitat.