EATMamba: Evolutionary Token-Refined Vision Mamba for Tomato Leaf Disease and Pest Classification

Accurate and efficient recognition of tomato leaf diseases and pests is essential for precision agriculture, yet practical deployment remains challenging due to complex backgrounds, domain shifts, and subtle inter-class visual differences. From a broader mathematical perspective, this task can be viewed as token-level representation refinement under noise, ambiguity, and distribution shift. Recent state-space-model-based vision backbones provide favorable efficiency for high-resolution imagery but lack explicit mechanisms for adaptive feature refinement under noisy conditions. To address these issues, we propose EATMamba, an evolution-inspired Vision Mamba framework for tomato leaf disease and pest classification. Rather than being limited to a task-specific classifier, EATMamba is formulated as an evolution-inspired differentiable token-refinement mechanism designed to be compatible with state-space visual recognition backbones. EATMamba introduces two lightweight and fully differentiable modules—Evolutionary Crossover–Interaction and Knowledge-Guided Mutation–Selection—which perform token-level recombination and selective refinement to emphasize discriminative disease cues while suppressing irrelevant background information. These modules are inspired by crossover, mutation, and selection concepts, but are implemented as trainable differentiable operations over visual tokens, forming a generate–recombine–select style mechanism for representation refinement. The scope of this study is low-cost RGB-based visible-symptom disease and pest classification, rather than pre-symptomatic early disease detection. Extensive experiments on two complementary tomato datasets, including a controlled high-resolution dataset and an in-the-wild farm dataset, demonstrate that EATMamba consistently outperforms representative CNN-, Transformer-, and state-space-model-based baselines. Ablation studies and visualization analyses further confirm the complementary contributions of the proposed modules. Overall, EATMamba provides an effective and efficient framework for fine-grained plant disease recognition and illustrates how evolution-inspired principles can be incorporated into modern vision backbones for robust agricultural image analysis.