Abstract
Accurate pest classification is critical for precision agriculture, yet existing deep learning methods face challenges including computational inefficiency from uniform sample processing and inadequate modeling of complex feature relationships. This paper proposes AdaptPest-Net, a task-adaptive architecture with three key innovations: (1) Sample-Difficulty-Aware Dynamic Routing (SDADR) employs Gaussian-gated path selection to adaptively route samples through shallow, medium, or deep networks based on predicted classification difficulty, improving accuracy by matching network capacity to sample complexity; (2) Graph Convolution-Mamba Fusion (GCMF) synergistically combines a 3-layer GCN with adaptive adjacency for explicit spatial structural modeling and a selective Mamba state-space model (T = 8, input-dependent parameters) for temporal feature dynamics, capturing complementary topological relationships and long-range dependencies through parallel dual-branch extraction; (3) Bidirectional Cross-Modal Attention enables deep integration between spatial and temporal modalities through mutual enhancement with batch-level knowledge transfer and adaptive gate fusion, where graph topology directly guides Mamba's feature evolution. Comprehensive experiments on IP102 and D0 demonstrate that AdaptPest-Net achieves 78.4% accuracy on IP102, representing a significant improvement over existing methods, while D0 experiments validate strong cross-dataset generalization capability with 99.85% accuracy.