融合物候趋势特征与时序感知网络的冬小麦种植区提取方法

Method for Extracting Winter Wheat Planting Area Integrating Phenological Trend Features and Temporal-Aware Networks

摘要:

为提高冬小麦种植区空间分布信息提取精度,针对现有方法在多时相物候信息利用不足、显式物候趋势特征融合效率不高以及复杂农业景观下局部结构恢复能力有限等问题,本文提出一种融合物候趋势特征的时序感知网络ST-EdgeNet(Spatial-temporal edge-aware network)。以关中平原冬小麦关键生育阶段为对象,基于2025年3—5月Sentinel-2月尺度合成影像,选取7个原始波段和7个植被指数构建三时相基础光谱特征,并依据冬小麦在返青、抽穗和灌浆阶段的时序变化规律,构建包含旺盛期高值、峰值突出度、返青速率、峰后变化及软先验约束在内的9 维物候趋势特征。在模型设计上,构建物候感知时序融合模块(PhAT)以增强三时相特征的自适应交互能力,设置趋势特征独立编码支路(TrendEncoder)以提高显式物候特征的有效表达,并结合局部结构增强与解码精化策略改善复杂地块区域恢复效果。基于关中平原9个县区构建县级空间分离数据集,其中训练集、验证集和测试集分别来自6个、1个和2个县区,并与PSPNet、FCN 和SegFormer等主流语义分割模型进行对比。结果表明:融合物候趋势特征的STEdgeNet在空间独立测试区取得最优识别效果,总体精度(OA)、F1分数和平均交并比(mIoU)分别达到98.53%、97.60%和96.61%,较最优对比模型SegFormer的平均交并比提高10.25个百分点;引入物候趋势特征后,ST-EdgeNet的平均交并比由95.32%提升至96.61%,分割错误率相对降低27.6%。研究表明,物候趋势特征与时序感知网络的协同设计能够显著提升冬小麦种植区提取精度,所提方法可为冬小麦种植区高精度遥感制图与智能化调查提供方法参考。

Abstract:

A spatial-temporal edge-aware network, ST-EdgeNet, integrated with phenological trend features, was proposed to improve the extraction accuracy of winter wheat planting areas. The method aimed to address three limitations of existing methods: insufficient utilization of multi-temporal phenological information, low efficiency in fusing explicit phenological trend features, and limited capability for local structure recovery in complex agricultural landscapes. Focusing on the key growth stages of winter wheat in the Guanzhong Plain, monthly composite Sentinel-2 imagery from March to May 2025 was used. Seven original spectral bands and seven vegetation indices were selected to construct three-phase basic spectral features. According to the temporal variation characteristics of winter wheat during the green-up, heading, and grain-filling stages, a 9-dimensional phenological trend feature set was further developed, including vigorous-growth high-value response, peak prominence, green-up rate, post-peak variation, and soft-prior constraints. In terms of model design, a phenology-aware temporal fusion module (PhAT) was constructed to enhance the adaptive interaction among multi-temporal features; an independent TrendEncoder branch was introduced to improve the representation of explicit phenological trend features; and local structure enhancement and decoder refinement strategies were incorporated to improve recovery performance in complex field regions. A county-level spatially separated dataset covering nine counties in the Guanzhong Plain was established, in which the training, validation, and test sets were derived from six, one, and two counties, respectively. The proposed method was compared with mainstream semantic segmentation models, including PSPNet, FCN, and SegFormer. The results showed that ST-EdgeNet with phenological trend features achieved the best recognition performance in spatially independent test areas, with overall accuracy ( OA), F1-score, and mean intersection over union (mIoU) reaching 98.53%, 97.60%, and 96.61%, respectively. Its mIoU was 10.25 percentage points higher than that of the best comparative model, SegFormer. After phenological trend features were introduced, the mIoU of ST-EdgeNet was increased from 95.32% to 96.61%, while the segmentation error rate was reduced by 27.6%. These results indicated that the synergistic design of phenological trend features and a temporal-aware network was effective in improving the accuracy of winter wheat planting area extraction, providing methodological support for high-precision remote sensing mapping and intelligent surveying of winter wheat planting areas.