基于改进全卷积神经网络模型的土地覆盖分类方法研究

Research on Land Cover Classification Method based on Improved Fully Convolutional Neural Network Model

遥感卫星数据是地球表面信息的重要来源,但利用传统的遥感分类方法进行土地覆盖分类局限性大、过程繁琐、解译精度依赖专家经验,而深度学习方法可以自适应地提取地物更多深层次的特征信息,适用于高分辨率遥感影像的土地覆盖分类。文中对高分辨率影像中水体、交通运输、建筑、耕地、草地、林地、裸土等进行高精度分类,结合遥感多地物分类的特点,以DeepLabV3+模型为基础,作出了以下改进:(1)骨干网络的改进,使用ResNeSt代替ResNet作为骨干网络;(2)空洞空间金字塔池化模块的改进,首先在并联的每个分支的前一层增加一个空洞率相对较小的空洞卷积,其次在分支后层加入串联的空洞率逐渐减小的空洞卷积层。使用土地覆盖样本库和自制样本库进行模型训练、测试。结果表明,改进模型在2个数据集的精度和时间效率均明显优于原始DeepLabV3+模型:土地覆盖样本库总体精度达到88.08%,自制样本库总体精度达到85.22%,较原始DeepLabV3+模型分别提升了1.35%和3.4%,时间效率每epoch减少0.39 h。改进模型能够为数据量以每日TB级增加的高分影像提供更加快速精确的土地覆盖分类结果。

Remote sensing satellite data are essential source of earth surface information.However,traditional remote sensing classification methods usually have limitations and include cumbersome processes,and the accuracy of interpretation depends on the expert experience.Deep learning methods can adaptively extract more detailed feature information from field objects and are suitable for land cover classification of high-resolution remote sensing images.Based on the DeepLabV3+model,this paper makes the following improvements:(1)Improvement of the backbone network.We use ResNeSt instead of ResNet as the backbone network;(2)The improvement of the hole space pyramid pooling module.First,a hole convolution with a relatively small hole rate is added to the previous layer of the parallel branch,and then a series of hole convolution layers with a gradually decreasing hole rate are added to the back layer of the branch.We use the land cover sample database and the self-made sample database respectively for model training and classify water bodies,transportation,buildings,cultivated land,grasslands,forests,bare soil,etc.from high-resolution images.Our results show that the accuracy and time efficiency of the improved model using the two databases are significantly higher than those of the original DeepLabV3+model.The overall accuracy using the land cover sample database and selfmade sample database reach 88.08%and 85.22%,respectively,which are 1.35%and 3.4%higher than that using the original DeepLabV3+model,respectively.Also the time cost decreases by 0.39h per epoch.The improved model can provide faster and more accurate land cover classification results for high-resolution imageries that increases in terabytes of data per day.