深度学习语义分割方法在遥感影像分割中的性能分析

Performance Analysis of Semantic Segmentation Method Based on Deep Learning in Remote Sensing Image Segmentation

针对如何应用深度学习语义分割方法实现遥感影像高性能分割的问题,选择了当前流行的SegNet、PSPnet以及Deeplabv3+三种基于深度学习语义分割算法,利用南方某区域无人机高分辨率遥感影像中4类要素分割为实验,以总体精度、平均精度及平均交并比(MIoU)作为精度衡量指标,全面对比分析了3种算法的精度;结果表明,在迁移学习支持下,3种算法总体精度可提升2至5个百分点;通过对PSPNet算法运用不同骨干网络,验证了不同结构网络对精度的贡献,优选出复杂度低的骨干网络;采用集成学习的思路,利用投票法对多算法模型进行结果融合可提升总体精度1%左右;3种算法对植被及水体的分割效果均要优于建筑物及道路,其中Deeplabv3+算法精度最高,总体精度达到89.3%,MIoU达到80.4%,可实现要素的鲁棒分割。

Towards how to use deep learning semantic segmentation method to realize high performance segmentation of remote sensing image, three popular semantics segmentation algorithms based on deep learning are selected, which are SegNet, PSPnet and Deeplabv3+. Based on the classification of four types of elements in high-resolution remote sensing images of a UAV in the south, this paper takes the overall accuracy, average accuracy and MIoU as the accuracy measurement indicators, and comprehensively analyzes the performance of the three algorithms. The experimental results show that the overall accuracy of the three algorithms can be improved by 2 to 5 percentage points with the support of migration learning. By replacing the PSPNet algorithm backbone network, the contribution of different structural networks to accuracy is verified, and a backbone network with low complexity is optimized. Using the idea of integrated learning, the result fusion of multi-algorithm model based on voting method can improve the overall accuracy by about 1%. The three algorithms have better segmentation effects on vegetation and water body than buildings and roads. Among them, Deeplabv3+ algorithm has the highest accuracy, the overall accuracy reaches 89.3%, and MIoU reaches 80.4%, which can achieve the robust segmentation of the elements.