Adaptive Multi-Scale Feature Refinement for Wheat Phenology Recognition Using Cross-Scale Attention Mechanisms

Haifang Sun¹Liang Hou¹Xiaorui Guo²Yanan Wang³Jianan Min¹Xiaoliu Zheng¹Zhehao Tian¹Donghui Zhang⁴Xinshi Zhang¹Senhao Liu⁵Yu Gao⁶Zhaoli An¹Hao Qi^7*Liangjie Lv^2*

• ¹Institute of Agricultural Information and Economy, Hebei Academy of Agriculture and Forestry Sciences, Shijiazhuang, China
• ²Institute of Cereal and Oil Crops, Hebei Academy of Agriculture and Forestry Sciences, Shijiazhuang, China
• ³Hebei Agricultural Technology Extension Station, Shijiazhuang, China
• ⁴Institute of Remote Sensing Satellite, China Academy of Space Technology, Beijing, China
• ⁵Chongqing University, Chongqing, China
• ⁶Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
• ⁷Hebei Academy of Agriculture and Forestry Sciences, Shijiazhuang, China

Accurate delineation of crop growth stages under real-world field conditions remains a long-standing challenge in computational phenotyping, particularly for wheat whose developmental phases are characterized by subtle, continuous morphological transitions and environmental noise. In this study, we propose AMFR-Net, an Adaptive Multi-Scale Feature Refinement Network tailored for fine-grained wheat stage identification using ground-level RGB imagery. Unlike conventional architectures that struggle with ambiguous inter-stage boundaries and rigid receptive structures, AMFR-Net leverages a ResNet-101 backbone augmented by a novel Adaptive Multi-Scale Attention Fusion (AMSAF) module—comprising cross-scale interaction blocks and confidence-weighted feature aggregation—to hierarchically recalibrate spatial–semantic representations. This design enables the network to adaptively amplify phenologically salient cues while suppressing irrelevant context, ensuring robust generalization under constrained annotation and deployment conditions. Evaluated on the expert-labeled CGIAR benchmark, AMFR-Net achieves state-of-the-art performance across all major metrics (Top-1 Accuracy: 89.10%; Macro-F1: 89.10%; AUC: 97.88%) and demonstrates superior discriminability in phenologically adjacent stages compared to lightweight and deep CNN baselines. Ablation studies validate the synergistic effect of multi-level attention and scale-aware refinement. The proposed framework offers a scalable, interpretable, and field-deployable solution for in-situ phenology monitoring, and sets a foundation for future integration of multimodal sensing, weak supervision, and cross-seasonal adaptation.