基于特征分离机制的深度学习植被自动提取方法

随着遥感影像分辨率的提高,植被信息的高精度提取对于了解地表植被变化规律、评价生态区域具有重要意义。针对传统方法跨季节植被提取不完整问题,本文基于高分2号(GF-2)卫星数据,提出一种基于特征分离机制的深度学习语义分割网络植被提取方法。该网络在Densenet的基础增加可分离卷积和空间金字塔结合的特征分离机制来增大感受野,更有效利用植被的特征信息,提升了模型的精度。本文通过构建高精细跨季节植被样本库,使用本文所提方法,完成了遥感影像植被信息提取,并选取总体准确度、F1值和交并比作为评价指标,对不同的传统方法和深度学习方法进行精度对比与分析。实验结果表明,本文方法提取植被的效果较好,其中F1分数达到91.91%,总体准确度达到92.79%,交并比达到85.10%。对高分1号、高分6号和高景1号遥感影像进行植被提取通用性验证,结果表明本文方法具有一定的通用能力,可以从高分辨率遥感影像中准确地、自动地提取植被。本文研究成果可为城市生态环境评价和植被的应用研究提供数据参考。

Automatic Vegetation Extraction Method based on Feature Separation Mechanism with Deep Learning

With the improvement of the spatial resolution of remote sensing images, the high-precision extraction of vegetation information is of great significance for understanding the changing laws of surface vegetation and evaluating ecological regions. Aiming at the problem that the existing vegetation extraction methods are difficult to extract the yellow vegetation information and it is difficult to realize the vegetation cross-season extraction, this paper proposes a deep learning semantic segmentation network of vegetation extraction method based on the feature separation mechanism using the GaoFen-2 satellite data. The network adds a feature separation mechanism that combines separable convolution and atrous spatial pyramid on the basis of Densenet. The atrous spatial pyramid effectively reduces the loss of information while acquiring spatial features of different scales. This network takes the high-level semantic information of vegetation into account in complex background. The feature information is enhanced while the accuracy of the model is improved. In order to reduce the calculation amount and the parameter amount of the atrous spatial pyramid, a separable convolution layer is used to replace its original convolution layer. In this paper, we constructed a high-precision cross-season vegetation sample database. Using the method proposed in this article, vegetation information is extracted from remote sensing images, which solves the problem that it is difficult to effectively extract the yellow vegetation information. This paper selects overall accuracy, F1 score, and intersection over union as evaluation indicators to compare and analyze the accuracy of different traditional methods and deep learning methods. The experimental results show that the method proposed in this paper is better than traditional vegetation extraction methods and other deep learning methods according to the three evaluation indicators. The F1 score reaches 91.91%, the overall accuracy reaches 92.79%, and the intersection ratio reaches 85.10%. The general verification experiment of the different vegetation types in the remote sensing image of GaoFen-2 has been carried out. The experimental results show that the method in this paper can completely extract the vegetation types of woodland, arable land, and grassland in the image. The generalization of vegetation extraction is verified on the remote sensing images of GaoFen-1, GaoFen-6, and SuperView-1.The results show that the method proposed in this paper has a certain general ability. It can realize the automatic and high-precision extraction of vegetation from high resolution remote sensing images. The results of this paper can provide data reference for urban ecological environment evaluation and vegetation application research.