Abstract
Terrestrial landscape has complex three-dimensional (3D) features that are difficult to extract using traditional methods based on 2D representations. These methods often relegate such features to raster or metric-based (two-dimensional) representations based on Digital Surface Models (DSM) or Digital Elevation Models (DEM), and thus are not suitable for resolving morphological and intensity features for fine-scale land cover mapping. Small-footprint LiDAR provides an ideal way for capturing these 3D features. This research develops a novel method of integrating airborne LiDAR derived features and multi-temporal Landsat images to classify land cover types. We tested our approach in Williamson County, Illinois, which has diverse and mixed landscape features. Specifically, our method applied a 3D convolutional neural network (CNN) approach to extract features from LiDAR point clouds by (1) creating an occupancy grid, an intensity grid at 1-meter resolution, and then (2) normalizing and incorporating data into the 3D CNN. The extracted features (e.g., morphological and intensity features) from the 3D CNN were finally combined with multi-temporal spectral data to enhance the performance of land cover classification based on a Support Vector Machine classifier. Visual interpretation from both hyper-resolution photos and point clouds was used for training and preparation of testing data. The classification results show that our method outperforms a traditional method by 2.65% (from 81.52% to 84.17%) when solely using LiDAR and 2.19% (from 90.20% to 92.57%) when combining all available imageries. We demonstrate that our method can effectively extract LiDAR features and improve fine-scale land cover mapping through fusion of complementary types of remote sensing data.
| Original language | English (US) |
|---|---|
| Pages (from-to) | 423-434 |
| Number of pages | 12 |
| Journal | ISPRS Journal of Photogrammetry and Remote Sensing |
| Volume | 144 |
| DOIs | |
| State | Published - Oct 1 2018 |
Fingerprint
Keywords
- Big data analysis
- Convolutional neural network
- Land cover classification
- LiDAR
- Multi-temporal Landsat imagery
ASJC Scopus subject areas
- Atomic and Molecular Physics, and Optics
- Engineering (miscellaneous)
- Computer Science Applications
- Computers in Earth Sciences
Cite this
A 3D convolutional neural network method for land cover classification using LiDAR and multi-temporal Landsat imagery. / Xu, Zewei; Guan, Kaiyu; Casler, Nathan; Peng, Bin; Wang, Shaowen.
In: ISPRS Journal of Photogrammetry and Remote Sensing, Vol. 144, 01.10.2018, p. 423-434.Research output: Contribution to journal › Article
}
TY - JOUR
T1 - A 3D convolutional neural network method for land cover classification using LiDAR and multi-temporal Landsat imagery
AU - Xu, Zewei
AU - Guan, Kaiyu
AU - Casler, Nathan
AU - Peng, Bin
AU - Wang, Shaowen
PY - 2018/10/1
Y1 - 2018/10/1
N2 - Terrestrial landscape has complex three-dimensional (3D) features that are difficult to extract using traditional methods based on 2D representations. These methods often relegate such features to raster or metric-based (two-dimensional) representations based on Digital Surface Models (DSM) or Digital Elevation Models (DEM), and thus are not suitable for resolving morphological and intensity features for fine-scale land cover mapping. Small-footprint LiDAR provides an ideal way for capturing these 3D features. This research develops a novel method of integrating airborne LiDAR derived features and multi-temporal Landsat images to classify land cover types. We tested our approach in Williamson County, Illinois, which has diverse and mixed landscape features. Specifically, our method applied a 3D convolutional neural network (CNN) approach to extract features from LiDAR point clouds by (1) creating an occupancy grid, an intensity grid at 1-meter resolution, and then (2) normalizing and incorporating data into the 3D CNN. The extracted features (e.g., morphological and intensity features) from the 3D CNN were finally combined with multi-temporal spectral data to enhance the performance of land cover classification based on a Support Vector Machine classifier. Visual interpretation from both hyper-resolution photos and point clouds was used for training and preparation of testing data. The classification results show that our method outperforms a traditional method by 2.65% (from 81.52% to 84.17%) when solely using LiDAR and 2.19% (from 90.20% to 92.57%) when combining all available imageries. We demonstrate that our method can effectively extract LiDAR features and improve fine-scale land cover mapping through fusion of complementary types of remote sensing data.
AB - Terrestrial landscape has complex three-dimensional (3D) features that are difficult to extract using traditional methods based on 2D representations. These methods often relegate such features to raster or metric-based (two-dimensional) representations based on Digital Surface Models (DSM) or Digital Elevation Models (DEM), and thus are not suitable for resolving morphological and intensity features for fine-scale land cover mapping. Small-footprint LiDAR provides an ideal way for capturing these 3D features. This research develops a novel method of integrating airborne LiDAR derived features and multi-temporal Landsat images to classify land cover types. We tested our approach in Williamson County, Illinois, which has diverse and mixed landscape features. Specifically, our method applied a 3D convolutional neural network (CNN) approach to extract features from LiDAR point clouds by (1) creating an occupancy grid, an intensity grid at 1-meter resolution, and then (2) normalizing and incorporating data into the 3D CNN. The extracted features (e.g., morphological and intensity features) from the 3D CNN were finally combined with multi-temporal spectral data to enhance the performance of land cover classification based on a Support Vector Machine classifier. Visual interpretation from both hyper-resolution photos and point clouds was used for training and preparation of testing data. The classification results show that our method outperforms a traditional method by 2.65% (from 81.52% to 84.17%) when solely using LiDAR and 2.19% (from 90.20% to 92.57%) when combining all available imageries. We demonstrate that our method can effectively extract LiDAR features and improve fine-scale land cover mapping through fusion of complementary types of remote sensing data.
KW - Big data analysis
KW - Convolutional neural network
KW - Land cover classification
KW - LiDAR
KW - Multi-temporal Landsat imagery
UR - http://www.scopus.com/inward/record.url?scp=85051758921&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85051758921&partnerID=8YFLogxK
U2 - 10.1016/j.isprsjprs.2018.08.005
DO - 10.1016/j.isprsjprs.2018.08.005
M3 - Article
VL - 144
SP - 423
EP - 434
JO - ISPRS Journal of Photogrammetry and Remote Sensing
JF - ISPRS Journal of Photogrammetry and Remote Sensing
SN - 0924-2716
ER -