SEF-RegNet: An efficient method for identifying straw field operation scenarios

Determining field operation scenarios is essential before detecting straw coverage. This study introduces a lightweight and efficient SEF-RegNet model for accurate recognition of field operation environments. The model innovatively incorporates a Shuffle_DSC module that enhances inter-channel information exchange through the integration of channel mixing operations and depthwise separable convolutions, effectively reducing parameter count and computational complexity. Additionally, an EMA attention module is introduced, effectively aggregating spatial information through feature grouping, parallel multi-scale subnetworks, and cross-spatial learning, which enhances the focus on key straw features. Focal Loss is utilized for loss calculation to dynamically adjust modulation and weighting factors, optimizing the model’s robustness against difficult samples and class imbalance issues. The model was trained and tested on datasets from typical black soil regions such as Jilin and Liaoning, achieving an overall recognition accuracy of 90.6 % on the independent test sets, a 6.7 % improvement over RegNetY. With FLOPs of 0.26 G, a model size of 14.8 MB, and a single-frame inference time of 23.6 ms, SEF-RegNet strikes a balance between high accuracy and lightweight design. Ablation studies confirmed the effectiveness of the Shuffle_DSC, EMA, and Focal Loss modules, while comparisons with other models further validated SEF-RegNet’s advancement. In tests involving 1,390 actual operational images across regions, the model achieved an overall accuracy of 85.9% with a single-frame inference time of 21.7 ms, demonstrating its practical deployment capability. This study provides a reliable technical foundation for accurate and equitable straw coverage detection and subsidy assessment under complex field operation scenarios.