融合多尺度特征的高分辨率森林遥感图像分割

Fusing Multi-scale Features for Segmentation of High Resolution Forest Remote Sensing Images

为实现对青海三江源国家级自然保护区高原森林的有效监测,基于深度学习技术提出一种融合多尺度特征的遥感图像分割算法。首先,构建了该地区首个2 m空间分辨率的高原森林数据集;其次,为解决遥感图像真值标签不足影响网络模型训练的问题,针对森林遥感图像分割的特点提出一种将图像打乱重组的数据增强方法,将训练数据扩充至1 600张;然后,为解决主流分割网络处理大范围遥感图像存在无法聚焦细节的缺陷,基于编解码结构,提出一种融合多尺度特征的高分辨率森林遥感图像分割网络模型,该模型融合了所设计的卷积模块、多尺度特征融合模块和特征放大提取模块。实验结果表明,所提数据增强方法提升了模型的分割精度,同时该模型经数据增强训练,交并比(intersection over union, IoU)高达89.64%,结果优于当前主流图像分割模型。

To achieve the effective monitoring of plateau forests in the Sanjiangyuan National Nature Reserve in Qinghai,a fusing multi-scale features remote sensing image segmentation algorithm based on deep learning technology was proposed.First,the first 2 m spatial resolution plateau forest dataset in the region was constructed.Second,to solve the problem of insufficient ground-truth label of remote sensing images which affects the training of network models,a data augmentation method of shuffling and reorganizing images was proposed according to the characteristics of forest remote sensing image segmentation,and the training data was expanded to 1600 images.To address the problem of mainstream segmentation networks that cannot focus on details in processing large-scale remote sensing images,a fusing multi-scale features high-resolution forest remote sensing image segmentation network model based on encoding and decoding structures was proposed.The model incorporated the designed convolution block,multi-scale feature fusion block and feature amplification extraction block.Results show that the data augmentation algorithm proposed improves the segmentation accuracy of the model,while the proposed model trained by the proposed data augmentation achieves an intersection over union(IoU)of 89.64%,and the results are better than that of the current mainstream image segmentation models.