Identifying Urban Wetlands through Remote Sensing Scene Classification Using Deep Learning: A Case Study of Shenzhen, China

Renfei Yang, Fang Luo, Fu Ren, Wenli Huang, Qianyi Li, Kaixuan Du, Dingdi YuanISPRS Int. J. Geo Inf 2022 · DOI

TL;DR

This paper introduces a novel scene-level deep learning framework for identifying urban wetlands, overcoming the limitations of traditional pixel-based methods. By treating image patches as scenes, the approach achieves higher accuracy and provides a more intuitive understanding of wetland distribution, crucial for effective urban planning and conservation.

What Problem Does This Paper Address?

Urban wetlands are vital for ecosystem services but face significant degradation due to rapid urbanization. Accurately identifying and mapping these wetlands is a critical first step for their conservation and management. Traditional remote sensing classification methods, which analyze individual pixels, struggle with the scattered, dynamic, and complex nature of urban wetlands. These wetlands often appear as a mix of vegetation, water, and other land cover types within a single pixel, leading to misclassification. Furthermore, urban wetlands are not static entities; their characteristics change dynamically, making single-time-point pixel analysis insufficient. This paper addresses the challenge of accurately classifying these complex urban wetland "scenes" rather than just individual pixels.

Methods

The researchers developed an urban wetland identification framework based on scene-level classification using deep learning. The core of their approach involves:

  • Scene Segmentation: Sentinel-2 high-resolution multispectral images of Shenzhen were segmented into 320m x 320m square patches, treating each patch as a "scene" for classification.
  • Deep Learning Models: Twelve different Convolutional Neural Network (CNN) models (including VGG16, ResNet, DenseNet, MobileNet, and EfficientNet) were adapted and compared for their performance in classifying these wetland scenes. These models were pre-trained on the ImageNet dataset and then fine-tuned for the specific task.
  • Classification Criteria: A patch was classified as a wetland scene if a specific wetland type (or water) covered more than 50% of its area. Conversely, if a non-wetland class dominated, it was classified as a non-wetland scene.
  • Spatial Pattern Analysis: The standard deviation ellipse method was employed to analyze the spatial distribution and patterns of the identified urban wetland scenes.

Datasets

  • Sentinel-2 High-Resolution Multispectral Images: Used for generating the sample datasets and performing the scene classification. These images were resampled to a 10m resolution.
  • Very High-Resolution Optical Images (0.2m resolution): Used to create the local classification system and sample datasets, serving as a reference for ground truth.
  • GlobeLand30 and GLC_FCS30: Existing pixel-based land cover products were used for comparative analysis with the scene-level classification results.

Key Insights and Ideas

  • Scene-level classification is superior for urban wetlands: The study demonstrates that treating urban wetland areas as "scenes" (patches) rather than individual pixels significantly improves classification accuracy and provides a more ecologically relevant representation.
  • DenseNet121 excels: Among the tested CNN models, DenseNet121 emerged as the top performer, highlighting its effectiveness in capturing the complex features of urban wetland scenes.
  • Urban wetlands are challenging: The F1-scores for wetland subcategories were generally lower than for non-wetland categories, indicating the inherent difficulty in distinguishing between different types of urban wetlands. Marsh and pond scenes, in particular, were harder to classify accurately.
  • Open water dominates Shenzhen's urban wetlands: A striking finding is that open water constitutes a substantial majority (78%) of the identified urban wetland scenes in Shenzhen, with a distinct southward distribution.
  • Other wetland types are scarce and scattered: Swamps, marshes, tidal flats, and ponds, while ecologically important, represent a very small fraction (4.64%) of Shenzhen's total area, underscoring the severity of wetland degradation.

Key Results

  • DenseNet121 achieved the highest performance: The best model, DenseNet121, reached an overall accuracy (OA) of 0.89 and a kappa coefficient of 0.86.
  • Good consistency with existing products: The scene-classification results showed high agreement with pixel-based products like GlobeLand30 (OA: 0.959, Kappa: 0.870) and GLC_FCS30 (OA: 0.956, Kappa: 0.859).
  • Spatial distribution: Wetland scenes in Shenzhen exhibit an east-west balance, with open water showing a southward pattern, while other wetland types are more scattered and show a slight westward distribution.
  • Composition: Open water accounted for 78% of wetland scenes, while swamps, marshes, tidal flats, and ponds collectively made up only 4.64% of Shenzhen's total area.

Limitations and Future Work

  • Resolution limitations: The Sentinel-2 images used were resampled to 10m resolution, which might limit the fine-grained feature extraction for very small wetland patches.
  • Data availability: While very high-resolution optical images provided detailed ground truth, their limited spectral information and application scale restrict their use for city-wide classification.
  • Future directions: The authors suggest exploring higher-resolution multispectral remote sensing images and investigating more advanced CNN models to further improve classification accuracy and address the challenges of identifying small, scattered urban wetlands. They also emphasize the potential of scene-level classification for more intuitive and effective urban wetland management and policy formulation.

Summary updated

BibTeX
@article{yang2022identifying,
  author = {Renfei Yang and Fang Luo and Fu Ren and Wenli Huang and Qianyi Li and Kaixuan Du and Dingdi Yuan},
  title = {{Identifying Urban Wetlands through Remote Sensing Scene Classification Using Deep Learning: A Case Study of Shenzhen, China}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2022},
  doi = {10.3390/ijgi11020131},
}
Figure