AI-Generated Synthetic Video as a Training Resource for Invasive Species Classifiers: A Reproducible Bio-Prompting Framework for Synthetic-to-real Generalisation
The limited availability of labelled visual data remains a major constraint for training reliable species classifiers in ecological applications. Although synthetic data generation using generative artificial intelligence represents a promising alternative, robust methodological frameworks are still lacking to evaluate its effectiveness and its ability to generalise to real-world data.In this work, we present a reproducible framework based on biologically informed prompt engineering to generate ecologically plausible synthetic data and systematically assess synthetic-to-real generalisation. Two widely used generative paradigms, Text-to-Video (T2V) and Image-to-Video (I2V), were employed to construct training datasets. These datasets were used to train five convolutional neural network architectures, ranging from lightweight models to pre-trained deep networks, using invasive psittacines as a case study.Evaluation on an independent set of unseen real-world images showed that both paradigms achieve comparable overall performance across architectures, with no statistically significant differences in F1-score. However, clear differences emerged in error profiles: models trained with I2V exhibited higher precision and fewer false positives, whereas those trained with T2V achieved higher recall at the cost of increased false positives. These patterns were consistent across runs, indicating that they reflect systematic differences between generative paradigms rather than random variability.The results demonstrate that synthetic video data can serve as a robust and cost-effective complement to real-world data in ecological classification tasks, provided that its generation and evaluation are embedded within a rigorous and reproducible methodological framework.The code and data used in this study are available for peer review in an anonymous private repository: