PestCLIP: an incremental pest recognition framework based on a vision-language model

Effective agricultural pest management, crucial for global food security and ecosystem balance, demands robust identification systems capable of adapting to dynamic environments. While deep learning shows promise, current methods often fail in practical class incremental learning scenarios, suffering catastrophic forgetting when encountering learning new pest species. This limitation hinders the development of truly adaptive tools for incremental pest recognition. Addressing this gap, we aimed to create a framework integrating advanced artificial intelligence (AI) with ecological needs for continuous and reliable pest recognition.

We propose PestCLIP, a framework for incremental pest recognition leveraging contrastive language-image pretraining (CLIP) model. To combat forgetting, PestCLIP employs dual-prompt tuning and a unique Concept Pool strategy that captures essential class features without extensive data replay. Crucially, it incorporates Prediction Distribution Calibration through incremental logit adjustment. Tested across diverse agricultural pest datasets (Li's, AgriInsect200, and Farm Insect) and general benchmark (mini-ImageNet), PestCLIP demonstrated superior class incremental learning performance, achieving 97.50% accuracy on Li's dataset with only a 5.55% drop when integrating new pest classes.

Extensive testing on diverse agricultural pest datasets demonstrates the superiority of PestCLIP in incremental pest recognition tasks. The visualization results confirm that PestCLIP effectively preserves class-specific concepts and mitigates prediction bias through distribution calibration. The proposed PestCLIP marks a pivotal step in advancing incremental pest recognition, enhancing the adaptability and reliability of smart pest management systems in dynamic agricultural environments. © 2026 Society of Chemical Industry.