基于深度学习的无人机土地覆盖图像分割方法

Deep Learning Based Unmanned Aerial Vehicle Landcover Image Segmentation Method

编制土地覆盖图需要包含精准类别划分的土地覆盖数据,传统获取方法成本高、工程量大,且效果不佳。提出一种面向无人机航拍图像的语义分割方法,用于分割不同类型的土地区域并分类,从而获取土地覆盖数据。首先,按照最新国家标准,对包含多种土地利用类型的航拍图像进行像素级标注,建立无人机高分辨率复杂土地覆盖图像数据集。然后,在语义分割模型DeepLabV3+的基础上进行改进,主要包括:将原始主干网络Xception+替换为深度残差网络ResNet+;引入联合上采样模块,增强编码器的信息传递能力;调整扩张卷积空间金字塔池化模块的扩张率,并移除该模块的全局池化连接;改进解码器,使其融合更多浅层特征。最后在本文数据集上训练和测试模型。实验结果表明,本文提出的方法在测试集上像素准确率和平均交并比分别为95.06%和81.22%,相比原始模型分别提升了14.55个百分点和25.49个百分点,并且优于常用的语义分割模型FCN-8S和PSPNet模型。该方法能够得到精度更高的土地覆盖数据,满足编制精细土地覆盖图的需要。

Compilation of land-cover maps needs high qualified land-cover data with precise classification.Traditional techniques to obtain these have the problem of high cost,heavy workload and unsatisfied results.To this end,a semantic segmentation method was proposed for unmanned aerial vehicle(UAV)images,which was used to segment and classify different types of land areas to obtain land-cover data.Firstly,the UAV images were annotated which contained various land use types at pixel level according to the latest national standards,and the high-resolution complex land-cover image data set of UAV was established.Then,several significant improvements based on original design of semantic segmentation model DeepLabV3+were made,including replacing the original backbone network Xception+with the deep residual network ResNet+;adding joint upsampling unit after backbone network to enhance the encoder’s capability of information transfer and conduct preliminary upsampling;adjusting dilated rates of atrous spatial pyramid pooling(ASPP)unit to smaller ones and removing global pooling connection of the module;and improving the decoder by fusing more low-level features.Finally,the models were trained and tested on the UAV high-resolution land-cover dataset.The presented model achieved good experimental results with pixel accuracy of 95.06% and mean intersection-over-union of 81.22% on the test set,which was 14.55 percentage points and 25.49 percentage points higher than that of the original DeepLabV3+model respectively.The proposed method was also superior to the commonly used semantic segmentation methods FCN-8S(pixel accuracy was 32.39%,mean intersection-over-union was 8.39%)and PSPNet(pixel accuracy was 87.50%,mean intersection-over-union was 50.75%).The results showed that the proposed method can obtain more accurate land-cover data and meet the needs of compiling fine land-cover maps.