自适应多源遥感影像的地面光伏提取方法  

Fine-grained Extraction of Ground Photovoltaics in Adaptive Multi-source Remote Sensing Images

作　　者：陈雨艳 张军[1] 张俊[1] 杜进祥 CHEN Yuyan;ZHANG Jun;ZHANG Jun;DU Jinxiang(School of Earth Sciences,Yunnan University,Kunming 650000,China;International Institute of Rivers and Ecology,Yunnan University,Kunming 650000,China)

机构地区：[1]云南大学地球科学学院,昆明650000 [2]云南大学国际河流与生态研究院,昆明650000

出　　处：《遥感信息》2024年第3期136-143,共8页Remote Sensing Information

基　　金：国家国防科技工业局高分专项(89-Y50G31-9001-22/23);云南大学研究生创新人才培养项目(HXKC202112)。

摘　　要：针对深度学习模型在跨数据应用时的精度和效率问题,采用U-Net框架,结合注意力模块和轻量级DenseNet,提出DA-UNet,并利用多源遥感影像地面光伏数据集进行训练,实现了DA-UNet自适应地从多源高分辨率遥感影像中获取地面光伏特征。与经典网络相比,DA-UNet在视觉效果和精度评价方面具有更好的效果,其中OA、F1评分和IoU指标分别为95.92%、94.68%和90.91%。此外,DA-UNet在其他地区的高分辨率遥感影像上也表现出较强的泛化性和适用性,在实际地面光伏提取方面具有较强的应用价值。To address the accuracy and efficiency challenges of deep learning models across various data sources,this study employs the U-Net framework,combining attention modules and lightweight DenseNet to introduce DA-UNet.Through training on a multi-source remote sensing image ground photovoltaic dataset,DA-UNet autonomously extracts ground photovoltaic features from diverse high-resolution remote sensing images.In comparison to conventional networks,DA-UNet demonstrates superior performance in visual quality and precision evaluation,with an overall accuracy(OA),F1 score,and intersection over union(IoU)of 95.92%,94.68%,and 90.91%,respectively.Furthermore,DA-UNet exhibits robust generalization and applicability to high-resolution remote sensing images in other regions,offering substantial practical value for ground photovoltaic extraction.

关 键 词：深度学习 密集网络 注意力机制 地面光伏 跨数据 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP18[自动化与计算机技术—控制科学与工程]