Abstract
In recent years, Deep Learning (DL) has become the core technology in machine learning (ML) and has gained popularity in a variety of applications. Due to the great success in other fields of application (computer vision, robotics etc.), DL approaches have also got much attention of the remote sensing community. Adopting and combining DL and LiDAR (Light Detection And Ranging) technologies opens up new possibilities, especially for solving more complex problems like the individual tree species classification based on ALS (Airborne Laser Scanning) data.
Our approach uses various strategies (for pre-processing) and DL approaches to perform the tree species classification based solely on the geometric properties of the individual trees that exist in the ALS data. The process starts with the automatic detection of single trees from the ALS data. This makes the basis for the subsequent pre-processing and the tree species classification. The pre-processing step involves the extraction of individual tree point clouds (of tree crowns) from the ALS data. The data (point clouds - x, y, z) extracted in this way can be used directly as input for the deep learning process. Depending on which DL approach has been selected, an additional transformation into raster could be necessary.
Based on the type of input from pre-processing (2D raster or 3D point cloud) two different DL based approaches were developed for further processing. In case of 2D rasters, several existing network architectures (Inception V3, VGG16) were re-trained to perform the classification by transfer learning. Furthermore a proprietary network architecture was developed. For 3D point cloud data, two different network architectures (PointNet ++, PointCNN) were used and trained from scratch for the new task.
The ALS data used covers a mixed forest area (2 x 9 km) in Burgau, Austria, with different tree species, the most common being spruce and pine as well as some deciduous tree species (beech, oak, birch, ash and alder). The classification is successfully performed for these three different tree species. The classification accuracy ranges from 70 % to 90 % depending on data type, network architecture and tree species. Our results also show that using the point clouds directly (without the transformation to raster) gives better accuracies, especially in the case of PointCNN network architecture.
Our approach uses various strategies (for pre-processing) and DL approaches to perform the tree species classification based solely on the geometric properties of the individual trees that exist in the ALS data. The process starts with the automatic detection of single trees from the ALS data. This makes the basis for the subsequent pre-processing and the tree species classification. The pre-processing step involves the extraction of individual tree point clouds (of tree crowns) from the ALS data. The data (point clouds - x, y, z) extracted in this way can be used directly as input for the deep learning process. Depending on which DL approach has been selected, an additional transformation into raster could be necessary.
Based on the type of input from pre-processing (2D raster or 3D point cloud) two different DL based approaches were developed for further processing. In case of 2D rasters, several existing network architectures (Inception V3, VGG16) were re-trained to perform the classification by transfer learning. Furthermore a proprietary network architecture was developed. For 3D point cloud data, two different network architectures (PointNet ++, PointCNN) were used and trained from scratch for the new task.
The ALS data used covers a mixed forest area (2 x 9 km) in Burgau, Austria, with different tree species, the most common being spruce and pine as well as some deciduous tree species (beech, oak, birch, ash and alder). The classification is successfully performed for these three different tree species. The classification accuracy ranges from 70 % to 90 % depending on data type, network architecture and tree species. Our results also show that using the point clouds directly (without the transformation to raster) gives better accuracies, especially in the case of PointCNN network architecture.
Original language | English |
---|---|
Publication status | Published - 8 Oct 2019 |
Event | SilviLaser 2019 - Iguazu Falls, Brazil Duration: 8 Oct 2019 → 10 Oct 2019 |
Conference
Conference | SilviLaser 2019 |
---|---|
Country/Territory | Brazil |
City | Iguazu Falls |
Period | 8/10/19 → 10/10/19 |