基于多尺度特征融合Mamba网络的冬小麦制图方法

Network for Remote Sensing Winter Wheat Mapping Based on Mamba with Multi-scale Feature Fusion

摘要:

冬小麦作为我国最主要的粮食作物之一，对其空间分布进行准确、及时制图，具有重要经济与社会价值。传统深度学习方法在作物制图中取得了一定成果，但卷积神经网络偏重捕捉小尺度局部细节信息，难以建模遥感影像中的长程依赖关系;而Transformer结构擅长提取大尺度全局信息，却易弱化细节表达且计算复杂度高。为此，本文提出了一种基于多尺度特征融合Mamba的遥感制图方法STM-Mamba，该网络组合Swin Transformer与多尺度卷积注意力（Multi-scale convolutional attention，MSCA）机制，利用Swin Transformer模块提取原始影像多尺度的语义特征图，利用MSCA模块提取特征图中多尺度信息，构建高效空间感知模块，并结合Mamba状态空间单元捕获遥感影像全局与局部上下文信息，从而提升了作物制图精度并有效降低了计算复杂度。模型训练阶段设计了边界感知损失，以强化对作物地块边缘判别。在陕西省渭南市Sentinel-2冬小麦数据集上试验结果表明，STM-Mamba平均交并比（mIoU）达87.14%，优于其他对比模型，模型参数量为2.717×10^7，相比CNN-Transformer混合架构的MAE-UPerNet减少74.29%。可视化结果显示，本文所提方法在复杂区域的地块边界识别更清晰、错分误差更少，田块分割更加完整。

Abstract:

As one of the most important staple crops in China, accurate and timely mapping of winter wheat's spatial distribution carries significant economic and social value. Traditional deep learning methods achieved notable progress in crop mapping;however, convolutional neural networks (CNNs) emphasize capturing small-scale local details and fail to effectively model long-range dependencies in remote sensing imagery. Conversely, Transformer architectures excel at extracting large-scale global information but tend to weaken fine-grained feature representation and suffer from high computational complexity. To address these challenges, STM-Mamba, a novel remote sensing mapping framework was proposed based on multi-scale feature fusion with Mamba. The network synergistically combined Swin Transformer and multi-scale convolutional attention (MSCA) mechanisms: Swin Transformer modules hierarchically extracted multi-scale semantic feature maps from raw imagery, while MSCA modules captured multi-scale spatial information within these feature maps to construct an efficient spatial perception module. By integrating Mamba state-space units that effectively captured both global and local contextual information, the proposed method achieved enhanced crop mapping accuracy while significantly reduced computational complexity. Additionally, a boundary-aware loss was designed during model training to enhance the discriminative capability along crop field boundaries. Experimental results on the Sentinel-2 winter wheat dataset from Weinan City, Shaanxi Province demonstrated that STM-Mamba achieved a mean intersection over union (mIoU) of 87.14%, outperforming all comparative models. The model had 2.717×10^7 parameters, representing a 74.29% reduction compared with the CNN-Transformer hybrid architecture MAE-UPerNet. Visualization results further confirmed that the proposed method achieved clearer field boundary identification, fewer misclassification errors, and more complete field segmentation in complex agricultural landscapes.