Enabling Country-Scale Land Cover Mapping with Meter-Resolution Satellite Imagery

Xin-Yi Tong, Gui-Song Xia, Xiao Xiang ZhuarXiv 2022 · DOI · OpenReview

TL;DR

This paper introduces "Five-Billion-Pixels," a massive dataset of high-resolution satellite imagery for detailed land cover mapping. They also propose a novel unsupervised domain adaptation method that effectively transfers knowledge from this labeled dataset to unlabeled satellite images, enabling country-scale land cover mapping with unprecedented detail.

What Problem Does This Paper Address?

Mapping land cover at a country-scale with high resolution is incredibly challenging. While high-resolution satellite imagery offers rich detail, it suffers from:

  • Complex Patterns: Densely populated urban areas and diverse natural landscapes present intricate patterns that are hard to classify.
  • Costly Annotation: Creating large, detailed land cover datasets requires extensive manual labeling, which is time-consuming and expensive.
  • Distribution Shifts: Satellite images from different geographical regions or acquired under varying conditions (e.g., lighting, atmospheric effects) exhibit significant "distribution shifts," meaning models trained on one set of images perform poorly on another.

Existing large-scale land cover datasets often use lower-resolution imagery or have a limited number of categories, failing to capture the nuances of built environments. This paper aims to bridge this gap by enabling detailed, large-scale land cover mapping using high-resolution imagery, even when target data is unlabeled.

Methods

The authors propose a two-pronged approach:

  1. The Five-Billion-Pixels Dataset: This is a new benchmark dataset containing over 5 billion labeled pixels from 150 high-resolution Gaofen-2 (4m resolution) satellite images. It features a comprehensive 24-category system that covers artificial, agricultural, and natural land cover types, offering a much richer classification scheme than previous datasets.
  2. Unsupervised Domain Adaptation (UDA) Approach: To overcome the distribution shift problem, they introduce a novel UDA method based on a Siamese network.
  3. Siamese Network: This network has two identical branches that process images from the labeled source domain (Five-Billion-Pixels) and the unlabeled target domain.
  4. Dynamic Pseudo-Labeling: Instead of relying on fixed thresholds, their method dynamically assigns pseudo-labels to pixels in the target domain based on their classification confidence (quantified by Shannon Entropy). The number of pseudo-labeled pixels increases as the model adapts.
  5. Class Balancing: To mitigate bias from imbalanced categories, the joint classification loss is weighted according to the class distribution in the source domain.
  6. Adaptive Domain Joint Learning: The model learns from both the labeled source data and the pseudo-labeled target data simultaneously, allowing it to adapt gradually without rigid domain alignment.

Datasets

  • Training/Source Domain: The newly introduced Five-Billion-Pixels dataset. This dataset consists of 150 Gaofen-2 (GF-2) satellite images (4m resolution) with over 5 billion labeled pixels across 24 categories.
  • Testing/Target Domains:
    • Chinese Megacities: Five megacities in China (Beijing, Chengdu, Guangzhou, Shanghai, Wuhan) mapped using PlanetScope (3m), Gaofen-1 (8m), and Sentinel-2 (10m) satellite imagery.
  • Additional Asian Cities: Six cities in five Asian countries (Bangkok, Delhi, Naypyidaw, Seoul, Tokyo, Yangon) mapped using Sentinel-2 (10m) satellite imagery.
  • The total study area across all these cities is approximately 60,000 km².

Key Insights and Ideas

  • Granularity Matters: A detailed 24-category system is crucial for understanding complex urban environments, going beyond the basic categories found in many existing datasets.
  • UDA is Essential for Scale: Large-scale mapping with diverse imagery necessitates robust UDA techniques to handle varying sensor characteristics and geographical conditions.
  • Dynamic Pseudo-Labeling is Key: Simply assigning pseudo-labels isn't enough. Dynamically adjusting the number of pseudo-labeled samples based on model confidence is vital for stable adaptation and avoiding noisy labels.
  • Class Balancing Prevents Bias: Imbalanced datasets are common. Explicitly balancing classes during adaptation helps prevent the model from being overly biased towards majority classes.
  • Multi-Scale Source Data Improves Adaptation: Using multi-scale images from the source domain (Five-Billion-Pixels) significantly enhances the model's ability to adapt to target images with different resolutions.

Key Results

  • Dataset Creation: The Five-Billion-Pixels dataset provides a comprehensive, high-resolution benchmark for land cover classification research.
  • UDA Performance: The proposed UDA approach achieved promising results in land cover mapping across 11 cities using unlabeled satellite imagery from multiple sensors (PlanetScope, GF-1, Sentinel-2).
  • Country-Scale Mapping: The method successfully enabled detailed land cover mapping across large geographical areas, demonstrating its scalability.
  • Comparison to Baselines: The proposed method outperformed other leading UDA approaches, particularly in scenarios with significant domain shifts. For instance, in Chinese megacities, their approach achieved higher overall accuracy (OA) and mean F1-score (mF1) compared to baselines and other UDA methods. For example, on dense labels for Chinese megacities, their method achieved an OA of 81.28% and mF1 of 46.79%, outperforming other methods.
  • Qualitative Results: The generated land cover maps showed clear identification of transportation networks, river systems, and detailed urban and agricultural areas, even in complex environments.

Limitations and Future Work

  • Limitations:
    • Natural Class Confusion: The approach still struggles with distinguishing certain natural categories (e.g., forest, irrigated field) when there are significant spectral shifts and inherent similarities between them, especially in winter conditions. This is attributed to the inherent constraints of UDA and the difficulty in distinguishing vegetation based solely on spectral information.
  • Road Discontinuity: Road extraction can be discontinuous, particularly in areas with diverse architectural styles and landscapes, suggesting that feature distribution differences can still pose challenges.
  • Computational Cost: While not explicitly stated as a limitation, the use of a Siamese network and iterative training implies significant computational resources are required.
  • Future Work:
    • Multi-Source Data Fusion: Integrating data from multiple sources (e.g., multi-season imagery, different resolutions) could improve the mapping of natural and urban areas.
  • Semi-Supervised Domain Adaptation: Exploring semi-supervised UDA with sparse annotations (patches or polygons) for regions with very different land distributions and category systems than China.
  • Improved Feature Representation: Investigating combinations of "low-level" spectral/textural information with "high-level" spatial contextual information for better classification in complex systems.

Summary updated

Abstract

High-resolution satellite images can provide abundant, detailed spatial information for land cover classification, which is particularly important for studying the complicated built environment. However, due to the complex land cover patterns, the costly training sample collections, and the severe distribution shifts of satellite imageries, few studies have applied high-resolution images to land cover mapping in detailed categories at large scale. To fill this gap, we present a large-scale land cover dataset, Five-Billion-Pixels. It contains more than 5 billion labeled pixels of 150 high-resolution Gaofen-2 (4 m) satellite images, annotated in a 24-category system covering artificial-constructed, agricultural, and natural classes. In addition, we propose a deep-learning-based unsupervised domain adaptation approach that can transfer classification models trained on labeled dataset (referred to as the source domain) to unlabeled data (referred to as the target domain) for large-scale land cover mapping. Specifically, we introduce an end-to-end Siamese network employing dynamic pseudo-label assignment and class balancing strategy to perform adaptive domain joint learning. To validate the generalizability of our dataset and the proposed approach across different sensors and different geographical regions, we carry out land cover mapping on five megacities in China and six cities in other five Asian countries severally using: PlanetScope (3 m), Gaofen-1 (8 m), and Sentinel-2 (10 m) satellite images. Over a total study area of 60,000 square kilometers, the experiments show promising results even though the input images are entirely unlabeled. The proposed approach, trained with the Five-Billion-Pixels dataset, enables high-quality and detailed land cover mapping across the whole country of China and some other Asian countries at meter-resolution.
BibTeX
@misc{tong2022enabling,
  author = {Xin-Yi Tong and Gui-Song Xia and Xiao Xiang Zhu},
  title = {{Enabling Country-Scale Land Cover Mapping with Meter-Resolution Satellite Imagery}},
  year = {2022},
  eprint = {2209.00727},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2209.00727},
}
Figure