Abstract
This paper presents an automated workflow for pixel-wise land cover (LC) classification from multispectral airborne laser scanning (ALS) data using deep learning methods. The workflow consists of three main procedures: data pre-processing, land cover classification, and accuracy assessment. By testing 1D, 2D, and 3D Convolutional Neural Networks (CNNs) across six input data combinations, the study achieved a peak overall classification accuracy of 97.2% with a 3D CNN. However, the 2D CNN emerged as the most efficient model, offering 97.0% accuracy with a runtime five times faster than the 3D alternative. These results demonstrate that multispectral ALS data, combined with spatial-context-aware deep learning, can successfully automate high-precision land cover mapping.
Introduction: The Multispectral Advantage
Accurate land cover mapping—distinguishing between water, vegetation, buildings, and roads—is essential for urban planning and environmental monitoring. Traditional LiDAR systems typically use a single laser wavelength, which provides excellent 3D structure but struggles to differentiate between materials with similar heights, such as a concrete roof versus a parking lot.
Multispectral ALS addresses this by firing lasers at three different wavelengths (532 nm, 1064 nm, and 1550 nm) simultaneously. This allows the system to capture unique spectral signatures for different land covers, much like how different materials reflect sunlight differently in satellite imagery.
Methodology and Workflow
The researchers developed an automated pipeline to process over 94 million points collected from a 1.99 km² area in Tobermory, Ontario, Canada. The workflow is divided into three critical stages:
- Data Pre-processing: Raw 3D point clouds were converted into 2D raster images. Nine layers were generated, including intensity maps for the three laser channels and various height metrics (e.g., maximum and average height).
- Classification via CNNs: The team compared three deep learning architectures. While 1D CNNs treat data as simple spectral signatures, 2D and 3D CNNs incorporate spatial relationships between neighboring pixels, which is crucial for identifying continuous features like roads or forests.
- Accuracy Assessment: Eighteen different models were evaluated using high-resolution (20 cm) orthophotos as ground truth to determine the optimal balance between accuracy and computational cost.
Key Insights and Results
- Spatial Context is Critical: Moving from 1D to 2D/3D CNNs improved accuracy by an average of 8.4%. This confirms that land cover classification benefits significantly from analyzing how a pixel relates to its neighbors.
- The 2D CNN "Sweet Spot": While the 3D CNN achieved the highest accuracy (97.2%), it was only 0.2% more accurate than the 2D CNN (97.0%) while requiring five times the processing time. For large-scale applications, the 2D CNN provides the best efficiency.
- Spectral vs. Structural Data: The study found that laser intensity (spectral info) is generally more useful for CNN classification than height information alone. The best-performing data stack utilized all three spectral bands plus the maximum height of the first laser return.
Conclusion
The integration of multispectral ALS with deep learning represents a significant leap forward for remote sensing. By capturing both the 3D structure and the spectral "color" of the landscape, this method overcomes the limitations of traditional LiDAR and optical imagery. The proposed 2D CNN workflow offers a highly accurate and computationally feasible solution for automated, large-scale land cover mapping.
Summary updated