Active replay for continual learning in plant species classification

Abstract

Continual Learning (CL) is essential in dynamic environments, such as biodiversity monitoring, due to the continuous emergence of new species or changes in their distribution over time. A major challenge in CL is catastrophic forgetting, where a model loses performance on previously learned classes as it learns new ones. One common approach to mitigate this issue is replay, in which samples from previous tasks are reintroduced during training. We investigate how Active Learning (AL) can improve replay strategies for the classification of plant species in a class-incremental learning setting. Specifically, we compare two AL-based sampling methods: selecting samples with the lowest confidence and those with the highest confidence, with a baseline random sampling strategy. The results show that while all methods achieve similar final accuracy, AL-based strategies significantly reduce catastrophic forgetting. High-confidence sampling reduces forgetting by 11.9 percentage points, while low-confidence sampling achieves a reduction of 9.2 percentage points compared to random sampling. We apply CL in a real-world scenario rather than a benchmark dataset, demonstrating its practical relevance for dynamic and evolving environments. Hence, our approach contributes to biodiversity monitoring, with the potential to support the development of a biodiversity index for tracking species diversity over time.