Hierarchical Classification of Urban ALS Data by Using Geometry and Intensity Information

Xiaoqiang Liu, Yanming Chen, Shuyi Li, Liang Cheng, Manchun Li2019 · DOI

TL;DR

This paper introduces a novel hierarchical classification method for urban Airborne Laser Scanning (ALS) data that leverages both stable geometric features and more volatile intensity information. By employing supervised learning for geometry and unsupervised learning for intensity, the method achieves improved classification accuracy, particularly for ground-level objects, and demonstrates greater robustness and transferability to data acquired under different conditions.

What Problem Does This Paper Address?

Urban environments are complex, and accurately classifying objects within them using ALS data is crucial for applications like urban morphology and ecology research. While ALS provides rich 3D geometric information, its intensity data is notoriously unstable, varying significantly due to factors like flight height, atmospheric conditions, and sensor calibration. This variability makes it challenging to directly use intensity in traditional supervised learning models, often leading to poor generalization to new datasets. Consequently, many studies either ignore intensity or struggle to integrate it effectively, leading to incomplete or less accurate classifications, especially for ground-level features like roads and low vegetation.

Methods

The proposed method employs a hierarchical approach, separating the classification of geometry and intensity information.

  1. Supervised Learning for Geometry: The initial step uses stable geometric features (including "absolute" features like height and normal, and "relative" features like point distribution) to train a supervised classifier. This provides a baseline classification.
  2. Unsupervised Learning for Intensity: After the initial classification, ground-level points are isolated. These points are then reclassified using an unsupervised learning method (specifically, Gaussian Mixture Models - GMM) applied to their intensity information. This step is designed to handle the fluctuating nature of intensity without requiring prior calibration.
  3. Heuristic Rule for Integration: A heuristic rule is developed to combine the results from the supervised and unsupervised classifications. This rule uses confidence scores (ratios of probabilities) from both classifiers to make a final decision, prioritizing the supervised result for elevated objects and carefully integrating the unsupervised result for ground-level objects. The outputs of both classifiers are probabilities, allowing for a nuanced combination.

Datasets

The study utilizes the benchmark dataset provided by the International Society for Photogrammetry and Remote Sensing (ISPRS) for the 3D Semantic Labeling Contest. This dataset includes training and testing areas from Vaihingen city, with pre-defined classes such as power line, low vegetation, impervious surface, car, fence, roof, façade, shrub, and tree.

Key Insights and Ideas

  • Decoupling Intensity: The core innovation is decoupling the use of geometry and intensity. By treating geometry as stable and intensity as volatile, the method can leverage intensity's discriminative power without succumbing to its instability.
  • Unsupervised Intensity for Robustness: Using unsupervised learning for intensity allows the model to adapt to variations in intensity across different datasets without needing retraining or calibration, making it more transferable.
  • Hierarchical Advantage: The hierarchical approach effectively utilizes the strengths of both supervised (for stable geometry) and unsupervised (for variable intensity) learning, leading to improved performance, especially for ground-level objects.
  • Intensity's Importance for Ground-Level Objects: The results strongly suggest that intensity information is particularly valuable for distinguishing ground-level features like low vegetation and impervious surfaces, where geometric features alone might be insufficient.

Key Results

  • The proposed method significantly improves classification accuracy compared to supervised learning methods that do not use intensity.
  • For low vegetation, precision increased by up to 10.20% and recall by up to 13.38% when using the proposed method compared to a baseline supervised method.
  • For impervious surfaces, recall increased by up to 13.38%.
  • When tested on data acquired under the same conditions as the training data, the proposed method performed comparably to supervised methods that directly incorporated intensity.
  • Crucially, when tested on data acquired under different conditions (simulated intensity variation), the proposed method outperformed supervised methods that directly used intensity, demonstrating its robustness.
  • The F1 scores for ground-level objects (low vegetation, impervious surface) showed substantial improvements, while elevated objects saw less dramatic gains, reinforcing the idea that intensity is most beneficial for ground-level features.

Limitations and Future Work

  • Imbalanced Data: The classification of façades, fences, and shrubs remains relatively poor, partly due to imbalanced class distribution in real-world data and lower point density for these objects.
  • Point Density: The low point density of certain objects (fences, façades, shrubs) poses a challenge.
  • Generalization: While the method shows improved transferability, clearly defining land cover types and testing on more diverse environments is suggested for a more generalized ALS classification model.
  • Future Directions: The authors suggest exploring rebalance machine learning methods, optimal or multi-scale radius selection for feature extraction, and graph-structured regularization frameworks for handling large wrongly labeled regions. They also aim to develop more generalized classifiers applicable to a wider range of urban environments.

Summary updated

BibTeX
@misc{liu2019hierarchical,
  author = {Xiaoqiang Liu and Yanming Chen and Shuyi Li and Liang Cheng and Manchun Li},
  title = {{Hierarchical Classification of Urban ALS Data by Using Geometry and Intensity Information}},
  year = {2019},
  doi = {10.3390/s19204583},
}
Figure