A Knowledge-Driven Incremental Learning Framework for Automating and Enhancing Plant Identification in Airports

This paper presents a knowledge-driven incremental learning framework for automated plant species recognition in airport environments, aimed at enhancing biodiversity monitoring through mobile applications and robotic platforms. The proposed framework integrates a fine-tuned Contrastive Language–Image Pretraining (CLIP) model based on the Vision Transformer (ViT-L/14) architecture, along with an evolving feature-based knowledge base that enables the incremental addition of new species without retraining. This architecture mitigates catastrophic forgetting by using a similarity-based inference mechanism that updates a compact feature knowledge base, thereby avoiding the need for full model retraining. Unlike memory-based or feature-replay methods, which require storing and replaying past data or embeddings, the proposed framework incrementally updates class prototypes using only expert-validated features, improving scalability and reducing annotation effort. Experiments on the Oxford 102 Flower and PlantCLEF 2015 datasets show improved accuracy across incremental stages.