3D LAND COVER CLASSIFICATION BASED ON MULTISPECTRAL LIDAR POINT CLOUDS

Zou Xiao-liang, Guihua Zhao, Jonathan Li, Yuanxi Yang, Yong Fang2016 · DOI

Abstract

Traditional monochromatic LiDAR systems excel at capturing 3D structure but struggle to distinguish between materials with similar heights. This paper presents an Object-Based Image Analysis (OBIA) approach using the Optech Titan—the first commercial multispectral airborne LiDAR system. By integrating spectral intensity from three wavelengths (532nm, 1064nm, and 1550nm) with vertical structural data, the researchers achieved an overall classification accuracy of 91.63% across nine land cover categories.

Introduction: Beyond Monochromatic LiDAR

Standard airborne LiDAR typically utilizes a single near-infrared wavelength. While effective for topography, it cannot easily differentiate a green lawn from a grey road based on height alone. Historically, researchers fused LiDAR with separate multispectral imagery, but this often led to spatial and temporal registration errors. Multispectral LiDAR solves this by capturing both 3D coordinates and multispectral intensity simultaneously from a single sensor.

The Multispectral Advantage

The Optech Titan system operates across three distinct channels: Channel 1 (1550nm - IR), Channel 2 (1064nm - NIR), and Channel 3 (532nm - Green). These specific wavelengths were chosen because different land covers exhibit unique reflectance patterns across them.

Spectral reflectance of different land covers across the three LiDAR channels.

As shown in the spectral curves, the green channel (532nm) is particularly useful because it can penetrate water, whereas NIR and IR are largely absorbed. This allows for the identification of water bodies and even shallow-water bathymetry within the same data acquisition pass used for terrestrial mapping.

Methodology: Object-Based Image Analysis (OBIA)

The study employed an OBIA framework to group point clouds into meaningful "objects" rather than classifying individual noisy points. The workflow involved:

  1. Multi-resolution Segmentation: Segmenting intensity data into objects using varying scale parameters to capture both broad features (forests) and fine details (roads).
  2. Feature Extraction: Developing customized indices such as pNDVI (pseudo-NDVI using Green and NIR), Ratio_Green, and DIFF_ELEVA (height difference between first and last returns).
  3. Rule-based Classification: A decision tree sorted objects into nine categories, including water, roads, power lines, and varying levels of vegetation.

Key Results

The synergy between spectral and structural data proved highly effective. While a road and a building roof might appear spectrally similar, the DIFF_ELEVA metric easily separates them by height. Conversely, pNDVI distinguishes flat lawns from flat roads based on their "greenness."

Final land cover classification map generated using the OBIA approach on multispectral LiDAR data.

The classification achieved an overall accuracy of 91.63% and a Kappa coefficient of 0.895. Notably, categories like Water, Power Lines, and Buildings reached 100% user accuracy. The primary areas of confusion occurred between "medium" and "high" vegetation, where vertical structures are similar, and at object boundaries where spectral signatures can become mixed.

Conclusion

This research demonstrates that multispectral LiDAR point clouds provide a robust, single-source solution for high-accuracy 3D land cover mapping. By eliminating the need for external imagery fusion, the OBIA approach offers a streamlined and precise method for urban and environmental monitoring. Future work will explore combining these datasets with Digital Surface Models to further enhance boundary precision.

Summary updated

Abstract

Multispectral Lidar System can emit simultaneous laser pulses at the different wavelengths. The reflected multispectral energy is captured through a receiver of the sensor, and the return signal together with the position and orientation information of sensor is recorded. These recorded data are solved with GNSS/IMU data for further post-processing, forming high density multispectral 3D point clouds. As the first commercial multispectral airborne Lidar sensor, Optech Titan system is capable of collecting point clouds data from all three channels at 532nm visible (Green), at 1064 nm near infrared (NIR) and at 1550nm intermediate infrared (IR). It has become a new source of data for 3D land cover classification. The paper presents an Object Based Image Analysis (OBIA) approach to only use multispectral Lidar point clouds datasets for 3D land cover classification. The approach consists of three steps. Firstly, multispectral intensity images are segmented into image objects on the basis of multi-resolution segmentation integrating different scale parameters. Secondly, intensity objects are classified into nine categories by using the customized features of classification indexes and a combination the multispectral reflectance with the vertical distribution of object features. Finally, accuracy assessment is conducted via comparing random reference samples points from google imagery tiles with the classification results. The classification results show higher overall accuracy for most of the land cover types. Over 90% of overall accuracy is achieved via using multispectral Lidar point clouds for 3D land cover classification.
BibTeX
@inproceedings{xiaoliang2016d,
  author = {Zou Xiao-liang and Guihua Zhao and Jonathan Li and Yuanxi Yang and Yong Fang},
  title = {{3D LAND COVER CLASSIFICATION BASED ON MULTISPECTRAL LIDAR POINT CLOUDS}},
  booktitle = {˜The},
  year = {2016},
  doi = {10.5194/isprs-archives-xli-b1-741-2016},
}
Figure