To address the challenges of structural variations among different crops and the performance degradation of deep learning models in unseen crop scenarios,this paper proposes a generalized crop region recognition method based on adaptive excess color enhancement and multi-scale feature fusion,termed ExCF-Net (Excess Fusion Network).First,global feature embedding is applied to the input images to generate adaptive excess color features.Then,by integrating deep learning techniques with traditional excess color methods,the model effectively shares feature parameters across domains.This enables it to learn cross-domain feature correlations,thereby enhancing generalization capability for crops with structures that differ significantly from the training data.To evaluate the effectiveness of the proposed method,we conducted ablation studies and comparative experiments against mainstream algorithms such as SwiftFormer and DeepLabV3+.Experimental results on the AgroScapes dataset show that the proposed method achieves an mIoU (mean Intersection over Union) improvement of 18.45 percentage points over baseline algorithms for test crops with substantial structural differences.Furthermore,compared to SwiftFormer,ExCF-Net yields mIoU gains of 5.66 and 9.30 percentage points on the AgroScapes and CRDLD datasets,respectively.Against DeepLabV3+,it achieves mIoU gains of 5.48 and 7.67 percentage points,respectively.Moreover,visualization results demonstrate that the proposed method effectively reduces boundary misclassification in unseen crop recognition tasks,significantly enhancing model robustness in complex agricultural scenarios.To support research in farmland perception,the source code and a visual demonstration are available at:https://github.com/GoldenUpwinds/Agro_domain_generalization.