基于深度学习的多模态地貌识别算法研究  

Research on Landform Recognition Algorithm in Multi-modal Geomorphologic Data Based on Deep-learning

作　　者：杜琳[1] 王光霞[1] 李科[1] DU Lin;WANG Guangxia;LI Ke(Information Engineering University,Zhengzhou 450052,China)

机构地区：[1]信息工程大学,河南郑州450052

出　　处：《测绘与空间地理信息》2020年第8期21-25,共5页Geomatics & Spatial Information Technology

摘　　要：为了提高地貌识别精度,本文基于深度学习融合多种模态地貌数据的特征对地貌进行识别。该方法利用深度学习网络分别从晕渲图、高程和坡度3种数据中提取地貌的物理和视觉特征,然后采用残差学习模型挖掘不同模态特征之间的深度关联,对多模态地貌数据特征进行学习并融合生成地貌深层特征,从而实现对地貌数据的联合表达。最后,使用3个全连接层和一个SoftMax分类器为每个样本数据生成一个地貌类别标签。实验结果表明,与以往的方法相比,基于深度学习的多模态地貌数据的地貌识别具有更好的性能。This paper presents an algorithm based on deep-learning to improve the accuracy of landform recognition in multi-modal geomorphologic data.To achieve effective representations of landform,the method uses a deep-learning network to generate physical and visual characteristics from multi-modal geomorphologic data(i.e.shaded relief,DEM,slope).It then uses a residual learning unit to acquire deep correlations between both physical and visual modality features,conducts learning and fusion of multi-modal geomorphologic features to generate landform deep feature,and achieve the joint representations of landform data.Finally,it employs three fully-connected layers and a SoftMax classifier to generate landform type label for each sample data.Comprehensive experiments show that the proposed method achieves much better performances as compared to traditional methods.

关 键 词：地貌识别 多模态数据融合 深度学习 卷积神经网络 

分 类 号：P209[天文地球—测绘科学与技术]