Why LiDAR makes hyperspectral imagery more valuable for forest species mapping OLI 2018 Andrew Brenner, Scott Nowicki & Zack Raymer
Presentation Outline Why hasn’t hyper-spectral met the hype? Why has LiDAR become foundational to forestry? Potential of data fusion Species classification Health classification Summary Questions
QSI Introduction Quantum Spatial (QSI) Largest Purely Geospatial Company in the US Full Service Geospatial Firm
QSI Overview Acquire Analyze Answer Airborne LiDAR (NIR & Bathy) Ground Mobile LiDAR Large Format Cameras Oblique Cameras Thermal Infrared Imaging Hyperspectral Imaging HD Videography Analyze Ortho Imagery Classified LiDAR Point Cloud Full Feature Classification Topographic /Planimetric Mapping 3D and 4D Models Structure from Motion Geospatial Cloud Solutions Enterprise GIS Answer Emergency Response Disaster Preparedness Hazard Detection Law Enforcement/Interdiction Regulatory Compliance Vegetation Analysis Environmental Monitoring Infrastructure Management
Hyperspectral data has been around for decades in forestry Why Hasn’t Hyperspectral Data Met the Hype Hyperspectral data has been around for decades in forestry Was anticipated to be the next great technology Rarely used operationally Why Expensive to collect Vegetation spectra change seasonally Need field training Need perfect acquisition conditions Otherwise …..
Lots of Noise Within class variation > between class variation Low accuracy and high cost means not operational
Sub-optimal conditions and hyperspectral only (no Lidar) SVM PC Classification SAM results over RGB image Sub-optimal conditions and hyperspectral only (no Lidar)
LiDAR is now considered vital to forest operations LiDAR is Foundational LiDAR is now considered vital to forest operations Being collected extensively by both private and public sector With increase of acquisition More players More competition Better sensors Lower cost Better delivery Better datasets
LiDAR is used in Forestry for Operations and infrastructure Hydrology Tree canopy Wildlife management Fisheries management Silviculture Archeology Hazard Assessment Fire risk …………
Automated Point Classification (APC) Point cloud classification –Any source Airborne/mobile/terrestrial scanning LiDAR Single photon/Geiger mode LiDAR Photo derived (PhoDAR) Advanced machine learning algorithms Feature identification and delineation: structures, infrastructure, streetscape and landscape
Lidar Point-Based Vegetation Delineation: Tree Top Points
Lidar Point-Based Vegetation Delineation: Tree Polygons
Lidar Point-Based Vegetation Delineation: Tree Polygons
Data Fusion High density Lidar provides physical characteristics at landscape, stand and tree-level, but only coarse tree type differentiation Many difference between genus/species types are physically observable only at the <cm scale Reflectance spectroscopy can be quantitatively used to measure many of the characteristics used to determine species in the field Airborne hyperspectral data can be collected at resolutions that capture those characteristics and map them at the individual crown level Applications: Forest inventory, utility veg management, invasive species and disease detection
---VSWIR--------------- 465 Channels (400-2500) VNIR Reflectance ---VSWIR--------------- Wavelength (nm) Visible wavelengths: color, chlorophyll absorption and other pigments Short-Wave Infrared (SWIR) wavelengths: canopy water content, carbon content, drought stress Near-Infrared (NIR) wavelengths: leaf IR-scatter, leaf structure, general health
VSWIR push-broom hyperspectral imaging system Acquisition System VSWIR push-broom hyperspectral imaging system 400-2500nm Max 465 bands co-acquired with high-density LiDAR ± 1.5 hours of solar noon Leaf on Clear – low cloud . VNIR SWIR Max cross-track pixels 1600 320 Max Spectral Resolution 335 130 Spectral Range 400-1000 nm 950-2500 nm
Hyperspectral Classification Acquisition: Hyperspectral imagery (VNIR or VSWIR) and Lidar Lidar analytics: Vegetation segmentation, tree polygons, DEM for image orthoprojection Field collected ground truth training database, validation database Hyperspectral image cube data reduction and optimization Classification: Support Vector Machine (SVM) Raster classification results to tree polygon Verification and Validation
True-color image derived from hyperspectral cube
Vegetation height model
Principal Component Analysis
Tree Species Assigned to Polygons
2016 Urban tree assessment in Louisville, KY. A hyperspectral imagery based species and health assessment aided with municipal tree inventory data. Collection date: August, 2016 Area: 2500 acres
Individual tree species can be differentiated based upon spectral variation due to leaf chlorophyll content, leaf structure, flowers and canopy structure that changes throughout the growing season.
Verification Results Total Accuracy 84%
Tree polygons allow the reduction of noise in the HSI classification Summary Combining hyperspectral imagery and LiDAR provide the data needed for accurate tree level classification Tree polygons allow the reduction of noise in the HSI classification HSI can also support health analyses at a tree level The combination provides the level of data that will support the cost of acquisition and processing
Questions