基于星机协同与边界约束的夏玉米提取方法

Summer Maize Extraction Based on UAV-Satellite Collaboration and Boundary Constraints

摘要:

准确获取夏玉米种植区空间分布信息对农业资源监测、种植结构调查和农田精细化管理具有重要意义。针对Sentinel-2卫星影像在小地块农区田块边界处易受混合像元和邻近地物干扰、传统总体精度指标难以充分反映边界识别差异等问题，本文提出一种基于星机协同与边界约束的夏玉米提取方法。以2025年7、8月Sentinel-2双时相光谱-植被指数特征作为主体输入，每个时相包含7个原始波段和7个植被指数，共构成28通道卫星物候特征;同时利用无人机红边影像生成夏玉米软标签和田块边界先验，并设计边界感知门控融合模块（BAGF），将无人机提供的局部高分辨率空间结构信息定向注入田块边界和混合像元区域。在陕西省关中平原9个县区夏玉米样本上开展试验，并采用县区级空间分离方式划分训练集、验证集和测试集。结果表明，CS-FPNet+BAGF在测试集上总体精度（OA）为97.22%、交并比（IoU）为95.23%、F1值为97.56%，优于DeepLabV3+、FPN、PSPNet、SegFormer和U-Net等模型。Trimap边界分区评价结果表明，不同模型在田块内部区域均具有较高识别精度，性能差异主要集中在田块边界区域;本文方法Boundary IoU（B-IoU）达91.18%，边界-内部精度差（B-I Gap）降至8.66%，较SegFormer和U-Net分别降低2.98、3.40个百分点。模型消融试验结果表明，FPN多尺度解码器、无人机软标签和BAGF边界门控均能提高模型识别能力，其中显式边界门控较学习型门控在边界约束方面表现更稳定;输入消融试验结果表明，7—8月双时相影像是夏玉米识别的主要信息来源，植被指数对总体分类精度具有小幅增益。研究结果表明，有限无人机影像可作为局部边界结构先验，与Sentinel-2双时相特征形成互补，从而提升复杂农区夏玉米种植区提取和田块边界识别精度。

Abstract:

Accurate acquisition of the spatial distribution of summer maize planting areas is important for agricultural resource monitoring, planting structure investigation, and refined farmland management. To address the problems that Sentinel-2 imagery is susceptible to mixed pixels and spectral interference from adjacent objects at field boundaries in smallholder agricultural areas, and that conventional overall accuracy metrics are insufficient to reflect differences in boundary recognition, a summer maize extraction method based on UAV-satellite collaboration and boundary constraints was proposed. In this method, Sentinel-2 dual-temporal spectral and vegetation-index features from July and August 2025 were used as the main input. Each temporal phase contained seven original spectral bands and seven vegetation indices, forming 28-channel satellite phenological features. Meanwhile, UAV red-edge imagery was used to generate summer maize soft labels and field boundary priors, and a boundary-aware gated fusion module, namely BAGF, was designed to directionally inject local high-resolution spatial structure information provided by UAV imagery into field boundary and mixed-pixel regions. Experiments were conducted by using summer maize samples from nine counties and districts in the Guanzhong Plain of Shaanxi Province, and a county-level spatially independent strategy was adopted to divide the training, validation, and test sets. The results showed that CS-FPNet+BAGF achieved an overall accuracy (OA) of 97.22%, an intersection over union (IoU) of 95.23%, and an F1 score of 97.56% on the test set, outperforming DeepLabV3+, FPN, PSPNet, SegFormer, and U-Net. The Trimap boundary partition evaluation showed that all models achieved relatively high recognition accuracy in field interior regions, while their performance differences were mainly concentrated in field boundary regions. The proposed method achieved a B-IoU of 91.18% and reduced the B-I Gap to 8.66%, which was 2.98 and 3.40 percentage points lower than those of SegFormer and U-Net, respectively. Model ablation experiments showed that the FPN multi-scale decoder, UAV soft labels, and BAGF boundary gating all improved the recognition ability of the model, and the explicit boundary gating mechanism was more stable than the learnable gating mechanism in boundary constraint. Input ablation experiments showed that the July-August dual-temporal imagery was the main information source for summer maize identification, while vegetation indices provided a slight gain in overall classification accuracy. The results indicated that limited UAV imagery could serve as a local boundary structure prior and complement Sentinel-2 dual-temporal features, thereby improving summer maize planting area extraction and field boundary recognition accuracy in complex agricultural areas.