基于高分二号影像和改进DeepLabV3+的冬小麦种植区提取模型研究

Extraction Model Study of Winter Wheat Planting Areas Based on GF-2 Imagery and Improved DeepLabV3+ Model

摘要:

准确获取冬小麦的空间分布信息是长势监测和产量预估的基础。针对复杂种植环境下冬小麦遥感提取面临的光谱混淆严重、频域信息利用不足以及田块边缘细节丢失等问题,提出了一种基于国产高分二号影像和改进DeepLabV3+ 的冬小麦种植区提取模型RSMANet(Residual SCConv MFMSA network with adaptive feature fusion)。首先,通过引入空间和通道重构卷积(Spatial and channel reconstruction convolution,SCConv)设计SC-ResNet18 替换原有主干网络,提高模型对空间细节和语义特征的提取能力;其次,在编码阶段设计多尺度多频注意力(SCConv withmulti-scale and multi-frequency attention module,SCC-MFMSA)模块,该模块主要包括尺度分解、多频通道注意力(Multi-frequency channel attention,MFCA)和多尺度空间注意力(Multi-scale spatial attention,MSSA)3部分,通过双注意力协同提取影像频域和空域特征,有效区分冬小麦冠层纹理与背景高频噪声,并提高模型的多尺度感知能力;最后,在特征融合阶段引入多尺度注意力融合(Multiscale attention fusion,MAF)模块,实现局部细节特征和全局语义信息的动态融合,通过增强通道、空间和多尺度层次的特征表达和语义关联,改善田块边界模糊和光谱混淆问题。试验结果表明,RSMANet模型能有效提升冬小麦种植区提取精度,IoU和F1值分别达到89.09%和94.52%,较原始DeepLabV3+模型分别提升3.19、1.39个百分点,与其他主流模型(如CMTFNet、SACANet、TransUNet等)相比分别提升1.08~3.81个百分点和0.23~1.74个百分点。

Abstract:

Accurate acquisition of winter wheat spatial distribution information is fundamental to growth monitoring and yield prediction. To address the problems of severe spectral confusion, insufficient utilization of frequency-domain information, and loss of field edge details in remote sensing extraction of winter wheat under complex planting environments, an improved DeepLabV3+ model named RSMANet was proposed for extracting winter wheat planting areas by using Chinese GF-2 imagery. Firstly, SC-ResNet18 was designed to replace the original backbone network by introducing the spatial and channel reconstruction convolution (SCConv), which improved the model's ability to extract spatial details and semantic features. Secondly, a multi-scale and multi-frequency attention (SCC-MFMSA) module was constructed in the encoding stage, consisting of three components: scale decomposition, multi-frequency channel attention (MFCA), and multi-scale spatial attention (MSSA). By synergistically extracting frequency-domain and spatial-domain features of the imagery through a dual-attention mechanism, this module effectively distinguished winter wheat canopy texture from background high-frequency noise, and improved the model's multi-scale perception ability. Finally, a multi-scale attention fusion MAF) module was introduced in the feature fusion stage to achieve dynamic fusion of local detailed features and global semantic information. By enhancing feature representation and semantic correlation at channel, spatial, and multi-scale levels, the problems of blurred plot boundaries and severe spectral confusion were improved. Experimental results demonstrated that the RSMANet model effectively improved the extraction accuracy of winter wheat planting areas; the IoU and F1 score reached 89. 09% and 94. 52%, respectively, which were 3.19 and 1.39 percentage points higher than those of the original DeepLabV3+ model, and 1.08~3.81 and 0.23~1.74 percentage points higher than those of other mainstream models (such as CMTFNet, SACANet, and TransUNet). The research results can provide a reference for winter wheat planting area extraction based on high-resolution imagery.