摘要
【目的】建筑物是地图上的重要组成要素,其形状识别既是空间认知和相似关系领域的研究热点和难点,也可以为地图制图综合的自动化实现提供技术支持。针对目前基于监督学习的栅格化建筑物形状识别方法需要大量标注样本的缺陷,本文融合了对比学习的自监督特征提取策略和迁移学习的监督分类技术,提出了一种基于对比迁移模型(Contrastive Learning Transfer Model,CLTM)的栅格化建筑物形状识别方法。【方法】首先,提取建筑物的形状并对其进行二值化和尺寸标准化处理,消除尺度与像素等因素的干扰;然后,构建对比学习模型对建筑物形状进行编码,获得高维特征向量,利用设计的对比损失函数优化模型;最后,以对比损失进行梯度更新,使用迁移参数预测建筑物形状以验证模型性能。【结果】实验结果表明,该方法的建筑物形状分类准确率达到93.79%,高于AlexNet方法的93.11%,但低于ResNet50方法的96.10%。在形状识别应用中运用t-SNE可视化技术,清晰展示了不同形状类别在特征空间中的聚类趋势,直观显示了形状识别效果,进一步验证了模型的有效性。【结论】尽管该方法的分类准确率低于监督的ResNet50方法,但该模型显著减少了分类准确率对大量标注数据的依赖,同时降低了人工视觉偏差的影响,是一种有效可靠的建筑物识别方法。
[Objectives]Buildings are fundamental elements on maps.In the fields of similarity relations and geospatial cognition,detecting building shapes presents both a significant challenge and a key research focus.Moreover,it has the potential to support cartographic generalization technologically.However,existing supervised learning methods for grid-based building shape recognition require a large number of labeled samples.To address this limitation,this paper proposes a Shape Recognition Method for Rasterized Buildings Based on a Contrastive Transfer Model.This approach combines transfer learning techniques with self-supervised feature extraction strategies.The goal is to extract shape features during a self-supervised learning phase,followed by supervised learning for shape classification,thereby minimizing labeling effort and reducing training costs.[Methods]In this study,the Contrastive Learning-based Transfer Model(CLTM)is applied for shape recognition focused on raster data.The process is as follows:First,the shapes of rasterized buildings are extracted through pre-processing.These shapes are binarized and standardized in size to eliminate the effects of pixel noise and size differences.Second,a high-dimensional feature vector is generated by encoding the building shapes using a contrastive learning model.This step extracts the unique characteristics of each shape and optimizes the model via a contrastive loss function.Finally,the model parameters are updated,and transfer learning is used to evaluate model performance through a downstream shape prediction task.[Results]Experimental results indicate that the classification accuracy of the proposed method reaches 93.79%,which is slightly higher than the AlexNet method but lower than ResNet50.In the shape recognition application,t-SNE visualization is used to clearly display the clustering trends of different shape categories in two-dimensional space.The clustering results indicate that shapes of the same category are closely grouped,while dissimilar shapes are well separated.This confirms the model's effectiveness,as it can accurately distinguish between similar and dissimilar shapes.[Conclusions]The proposed method performs well in building shape classification by leveraging data augmentation and a contrastive loss function to train the model and extract useful features from unlabeled data.This significantly reduces the need for manual data annotation and mitigates the influence of human visual bias.Compared to fully supervised methods,it offers distinct advantages,exhibiting strong shape recognition capabilities and providing an efficient and reliable approach for analyzing geospatial elements.
作者
李文德
师尚杰
闫浩文
LI Wende;SHI Shangjie;YAN Haowen(Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China;National-Local Joint Engineering Research Center of Technologies and Applications for National Geographic State Monitoring,Lanzhou 730070,China;Key Laboratory of Science and Technology in Surveying&Mapping Gansu Province,Lanzhou 730070,China)
出处
《地球信息科学学报》
北大核心
2025年第7期1582-1595,共14页
Journal of Geo-information Science
基金
国家自然科学基金项目(42301513)
甘肃省高等学校青年博士支持项目(2023QB-043)
地理信息工程国家重点实验室、测绘科学与地球空间信息技术自然资源部重点实验室联合资助基金项目(2023-02-08)。
关键词
地图制图综合
栅格数据
建筑物形状识别
对比学习
迁移学习
自监督学习
cartographic generalization
raster data
building shape recognition
contrastive learning
transfer learning
self-supervised learning
