基于深度学习的海岸带土地利用信息提取方法

Extraction Method of Coastal Land Use Information Based on Deep Learning

针对海岸带地物类型复杂多样,监测难度较大,本文选取山东省烟台市海岸带为研究区,运用时间序列Sentinel-2遥感影像数据,基于面向像元分类方法,利用不同地物在时间序列遥感影像不同波段上表现出的光谱差异特征,通过构建轻量级卷积神经网络提取出研究区土地利用信息,并对分类结果进行精度评估.结果表明:(1)像元时序特征值作为网络输入形式,提取出烟台市8种土地利用类型信息,很好区分出草/林地、耕地、裸地等地物,并能提取细长河流和道路,有效降低了“同物异谱”和“异物同谱”现象.(2)该方法总体分类精度、Kappa系数分别达到了91.32%,0.8965,比采用支持向量机、随机森林分类器总体精度提高4.17%,5.66%,分类制图中有效地避免了“椒盐”现象.基于像元时序特征值分类方法分类精度较高,为利用中分辨率遥感数据对海岸带土地利用信息快速、准确分类提供支持.

Aiming at the complex and diverse types of coastal features and the difficulty of monitoring,this paper takes the coastal zone of Yantai City,Shandong Province as the research area to study the extraction method of the coastal land use information.Based on the time series Sentinel-2 remote sensing image data and pixel-oriented classification method,we construct a lightweight convolutional neural network to extract the land use information in the study area and evaluate the accuracy of the classification results according to the characteristics of spectral differences displayed in different bands of time series remote sensing images.The results show as follows:(1)The time series feature value of the pixels is used as the network input form to extract the information of 8 land use types in Yantai coastal zone,which can distinguish grass/woodland,cultivated land,bare land and other ground features,and can extract slender rivers and roads,effectively reducing the phenomenon of“same-spectrum heterospectrum”and“foreign-substance homospectrum”.(2)The overall classification accuracy and Kappa coefficient of the classification method based on pixel time series feature reached 91.32%and 0.8965,respectively,which are 4.17%and 5.66%higher than those of the support vector machine and random forest classifier.The phenomenon of“pepper salt”is effectively avoided in the classification drawing.The classification method based on deep learning can provide support for fast and accurate classification of coastal land use information using middle-resolution remote sensing data.