杭州师范大学学报（自然科学版）

为准确提取高分辨率遥感影像中的光伏空间分布信息，解决因背景信息复杂导致的分类精度低和边界分割模糊等问题，提出了EMAHRNet模型.该模型使用HRNet(high-resolution net)作为主干网络，采用高、低分辨率并行融合的方式学习图像特征信息并引入可训练的融合权重，优化融合过程.同时，引入OCR(object-contextual representations)模块，根据图像的像素特征和目标对象的区域特征形成图像上下文信息，建立像素与区域的相关性，增强模型对光伏边缘信息的提取能力.此外，使用多尺度注意力模块EMA(efficient multi-scale attention)捕捉不同尺度区域特征，有效提升模型在复杂空间背景下的分割能力.基于浙江省2022年2 m分辨率的遥感影像自建光伏数据集，实验结果表明，EMAHRNet模型的F₁分数、交并比和准确率分别达到95.1%、84.8%和97.7%,相较于HRNet模型，分别提升1.8、2.4和0.8百分点，且在建筑、水面、耕地、林地和工业用地等复杂背景下均取得较好的分割效果.

Accurate extraction of photovoltaic(PV) spatial distribution from high-resolution remote sensing imagery is often challenged by complex backgrounds, leading to low classification accuracy and blurred boundaries. To address these issues, this paper proposed the EMAHRNet model. The model employed HRNet(high-resolution net) as its backbone, utilizing a parallel architecture that integrated high-and low-resolution features for learning image characteristics, and introduced trainable weights to optimize the feature fusion process. An OCR(object-contextual representations) module was incorporated to establish contextual relationships between pixels and object regions, enhancing the extraction of PV edge information. Furthermore, the efficient multi-scale attention(EMA) module was applied to capture regional features at different scales, significantly improving the model's segmentation capability in complex spatial contexts. Experimental results on a self-built PV dataset comprising 2-meter resolution remote sensing images from Zhejiang Province in 2022 demonstrated that the EMAHRNet model achieved a F₁-score of 95.1%, an intersection of union of 84.8%, and an accuracy of 97.7%. Compared to HRNet model, these metrics represented improvements of 1.8,2.4 and 0.8 percentage points, respectively. The proposed model also exhibited superior segmentation performance across diverse and complex backgrounds, including buildings, water bodies, farmland, forested areas, and industrial land.