DOI: 10.3390/f14091788 ISSN:

Mapping Vegetation Types by Different Fully Convolutional Neural Network Structures with Inadequate Training Labels in Complex Landscape Urban Areas

Shudan Chen, Meng Zhang, Fan Lei
  • Forestry

Highly accurate urban vegetation extraction is important to supporting ecological and management planning in urban areas. However, achieving high-precision classification of urban vegetation is challenging due to dramatic land changes in cities, the complexity of land cover, and hill shading. Although convolutional neural networks (CNNs) have unique advantages in remote sensing image classification, they require a large amount of training sample data, making it difficult to adequately train the network to improve classification accuracy. Therefore, this paper proposed an urban vegetation classification method by combining the advantages of transfer learning, deep learning, and ensemble learning. First, three UNet++ networks (UNet++, VGG16-UNet++, and ResNet50-UNet++) were pre-trained using the open sample set of urban land use/land cover (LULC), and the deep features of Sentinel-2 images were extracted using the pre-trained three UNet++ networks. Subsequently, the optimal deep feature set was then selected by Relief-F and input into the Stacking algorithm for urban vegetation classification. The results showed that deeper features extracted by UNet++ networks were able to easily distinguish between different vegetation types compared to Sentinel-2 spectral features. The overall classification accuracy (OA) of UNet++ networks and the Stacking algorithm (UNS) was 92.74%, with a Kappa coefficient of 0.8905. The classification results of UNet++ networks and the Stacking algorithm improved by 2.34%, 1.8%, 2.29%, and 10.74% in OA compared to a single neural network (UNet++, VGG16-UNet++, and ResNet50-UNet++) and the Stacking algorithm, respectively. Furthermore, a comparative analysis of the method with common vegetation classification algorithms (RF, U-Net, and DeepLab V3+) indicated that the results of UNS were 11.31%, 9.38%, and 3.05% better in terms of OA, respectively. Generally, the method developed in this paper could accurately obtain urban vegetation information and provide a reference for research on urban vegetation classification.

More from our Archive