An Approach for UAV-based DSM Filtering Using Image Classification
and Geometric Constraints
Dr Mehdi Ravanbakhsh Remote Sensing Consultant at Astron & Senior Research Fellow at
RMIT
email: [email protected]
• Easy and fast deployment
• Low cost of operation for small sites
• Less prone to atmospheric effects
• Ultra-high resolution of up to 2 cm
• Good quality orthophotos and Digital Surface Model
• Ideal for mapping the natural environment
Photogrammetric Processing Workflow
Multi-view Images
Geometric Alignment
Image Matching
Bundle Block
Adjustment
3D Point Cloud &
Orthophoto Generation
Automatic Aerial
Triangulation
Spatial Product:Multi-spectral Orthoimagery
RBG overlaid on NIR mosaic at 4cm resolution
Image mosaic generated from 194 images
Spatial Product: 3D Image
UAV versus LiDAR
• High level of detail at pixel size density (1cm-4cm)
• No need for co-registration of LiDAR and Imagery
• Unlike LiDAR spectral information is available
• A fraction of LiDAR data aquisition cost
LiDAR UAV Point Cloud
Accuracy Evaluation - Points outline Ground Control Points (GCPs): Red
Check Points (CKPs): Yellow
GCP No: 9
CKP No: 17
Accuracy Evaluation - Points outline Ground Control Points (GCPs): Red
Check Points (CKPs): Yellow
GCP No: 9
CKP No: 17
Accuracy Evaluation - Quantitative analysis
GCPs No 9
CKPs No 17
RMSE(∆E)(pl) 1.2
RMSE(∆N)(pl) 0.9
RMSE(∆H)(cm) 19
Image Resolution: 6cm pl: pixel E: Easting N: Northing
Random distribution of planimetric residuals
DEM Generation - Point Cloud Processing
Above ground objects (gray) & Bare ground (brown)
• Major high differences indicate obove ground objects
• Window size should be greater than the size of the largest feature
• Average slope known
• Still misclassification exists
DEM Generation - Image Classification RGB image mosaiec Clustring result Mask (red=1 & white=0)
• Unsupervised image clssification in 10 classes
• Merge classes to generate an image mask
• Point clouds on vegetation area are eleminated
DEM Generation - Final Result RGB image mosaic DEM with geometry Image Mask DEM with combined
geometry & spectral
• Semi-automatic approach vs manual approach
• True DEM can be generated
• Other elevation products like Digital Canopy Model
DCM Generation RGB Image DCM=DSM-DEM
Suitable dataset for tree stocking assessment and tree crown delineation
Normalised DSM (nDSM) Generation
NIR Image
nDSM
Vegetation
Mask
Suitable dataset for building extraction
Thank you & questions?