Airborne Laser Technology, Data Processing and Applications
Transcription
Airborne Laser Technology, Data Processing and Applications
Airborne Laser Technology, Data Processing and Applications E. Baltsavias Institute of Geodesy and Photogrammetry ETH Zurich, CH-8093 Zurich, Switzerland [email protected] www.photogrammetry.ethz.ch E. Baltsavias 1 Acknowledgements For this presentation material has been used, without or with modifications, from various colleagues, organisations and companies, who I want to thank: - C. Brenner (Leibniz University of Hannover) - N. Pfeifer and team (Technical University of Vienna) - D. Fritsch, J. Kilian, A. Wehr (Univ. of Stuttgart) - M. Doneus (Univ. of Vienna) - G. Vosselman, G. Sithole (ITC, TU Delft) - N. Demir (ETHZ) - A. Streilein (Swisstopo = Swiss Federal Office of Topography) - U. Lohr (at that time with firm Toposys) - P. Friess (Optech) - Y. Kuwano, A. Rohrbach (Leica) - A. Ullrich (Riegl) - Firms Leica, Riegl, Optech, IGI, Toposys, Fugro E. Baltsavias 2 1 Contents - History - Basic components of Airborne Laser Scanning (ALS) and functioning - Range measurement principles - Interaction of laser beam with targets - Full waveform digitising - Basic error sources and data accuracy - Processing overview - Strip adjustment - Point classification (filtering) - Overview of commercial systems (hardware and software) - Overview of applications and examples - Quality control of data - Bathymetric lidar - Short comparison of airborne laser scanning to other remote sensing technologies E. Baltsavias 3 History First optical laser developed in 1960 (Maiman, USA) First airborne laser ranging tried out in 1960s Started to be developed from early and middle 1970s, espec. in N. America, particularly for hydrographic and bathymetric applications First in late 1980s, the use of GPS made accurate range measurements from airborne laser profilers possible (Univ. of Stuttgart, Prof. Ackerman) Beginning of 90s profilers replaced by scanners (ALS), and GPS combined with IMU 1996 ISPRS Congress in Viennna: one ALS manufacturer, some reports and tests with ALS 1996-2000: work of several ISPRS Working Groups (WG) and one OEEPE WG on ALS. Publication of a special issue of ISPRS Journal of Photo & RS gives a good overview. Since 2000: ALS increasingly used in practice and in various applications; increasing scientific investigations and tests and better methods; continuously improving ALS systems, incl. waveform digitizing (espec. from 2004) and simultaneous double ALS (from 2006); more and better software; more service providers and users Often the term LiDAR is used (espec. In N. America): Light Detection And Ranging E. Baltsavias 4 2 Basic components and functioning - Active sensor: a very narrow, high energy ray is sent from a source (laser) to the scene, reflected back and recorded. Active means it works day and night (even better at night due to no sun interference). Active also means it can measure in textureless areas including shadows - Here we treat only airborne Lidar. What is recorded is the Time of Flight (TOF) or rarely the phase (called also continuous wave (CW) lasers), and almost always the intensity, although intensity is rarely used. All commercial ALS systems use TOF. There is one experimental CW ALS, called SCALARS (Univ. of Stuttgart) - Basically: measurement of distance via polar technique, e.g. the direction of the ray and the distance from the ray source to the scene are measured - Most ALS work in near infrared (NIR), so are influenced by clouds, snow, rain etc. (no weather independence, as radar) E. Baltsavias 5 Basic components - Laser transmitter and detector/receiver = range measuring unit Receiver consists of: detector, electronics, optical filters, optics - Deflection mechanism of the laser ray, e.g. mirror, polygon - GNSS (not only GPS) / IMU (Intertial Measurement Unit ; sometimes term INS used) Offsets (called also lever arm) between GNSS/IMU and laser, and misalignment between IMU and laser (called boresight calibration) must be known. GNSS is in differential modus with GNSS reference stations closeby. Dual frequency GNSS is mostly used. - Computer, onboard software (e.g. for navigation and flight management) and storage devices (data size is huge), including precise timing device that synchronises all sensors. Sensors synchronised (e.g. using GNSS PPS (Pulse Per Second) time port) and their data time stamped (e.g. with resolution ≤ 1 µsec) - Aircraft platform, mostly stabilised (important roll compensation, espec. for low altitude flights). For helicopters possible use of pod. E. Baltsavias 6 3 Basic components Optionally cameras: - digital cameras (frame or less line CCDs) -> especially for generation of orthoimages - Video or standard CCDs for attributation or annotation (see powerline example in applications) - less thermal, hyperspectral Clear tendency: as of 2008 ca. 50% of ALS had additional imaging sensors. Platforms - Airplanes - Helicopters (espec. for mapping of corridors or small areas) - Unpiloted Airborne Vehicles (UAVs), including small ones (developments from firm Riegl for lightweight ALS, e.g. http://www.riegl.com/products/uasuav-scanning/) - Much less, balloons and other nonconventional platforms. Terrestrial (many) and satellite (very few) laser scanners not treated here. E. Baltsavias 7 Basic components of an ALS system DGNSS Laser transmitter Control & data recording IMU Deflection unit Detector/ Receiver Ground E. Baltsavias 8 4 Basic components of an ALS system No fixed rules exist for distance of ground reference GNSS stations from airplane. Often 10-50 km, depending on topography (GNSS satellite visibility) and possible GNSS signal disturbances. E. Baltsavias 9 Pulse laser measurement principle (pulse in reality not trapezoidal as shown here, but Gaussian) • Travelling time: h=R ttravel = 2R / c with c speed of light Ground • Example: h = R = 1000 m ttravel = 6.7 µs • Range resolution: AT AR next pulse t ttravel tp E. Baltsavias t ΔR = ½ c Δttravel • Maximum pulse repetition frequency (theoretical, assuming no transmit / receive overlap): fmax = 1 / ttravel = c / 2R • Example: h = R = 1000 m fmax = 150 kHz 10 5 Pulse laser measurement principle tp Typical characteristics of sent pulse: Signal amplitude • Pulse width tp = 10 ns ( 3 m @ speed of light) t trise • Pulse rise time trise = 1 ns ( 30 cm @ speed of light) • In TOF, a position of the incoming pulse rising edge is used to determine TOF and range. Intensity is measured by the maximum or better the area of the incoming pulse. • Peak power Ppeak = 2,000 W • Energy per pulse E = Ppeak · tp = 20 µJ 3m • Average power (@ pulse repetition rate F = 10 kHz) Pav = E · F = 0.2 W Average power constant for a system. Thus, energy and Ppeak decrease with higher PRF F. Ground E. Baltsavias 11 CW laser measurement principle Characteristics (example ScaLARS): Modulation signal t • Two modulation frequencies fhigh = 10 MHz, flow = 1 MHz λshort = 30 m, λlong = 300 m Signal amplitude Modulation signal • High frequency used for accurate phase measurement t • Low frequency used for wavelength count (ambiguity for number of wavelengths) Left: measurement with f high AT, AR, transmitted and received amplitude Signal amplitude E. Baltsavias T: period tL : travelling time (corresponds to phase difference) 12 6 CW laser operation • Maximum unambiguous range determined by λlong : Rmax = λlong / 2 • Example: λlong = 300 m Rmax = 150 m Ground • Range resolution: ΔR = λshort / 4π Δφ AT t AR t λ • Range gating: Range differences known to be < λlong / 2 • Range tracking: If no sudden surface steps > λlong / 2 are present CW lasers will not treated further here E. Baltsavias 13 Some definitions - Pulse repetition frequency (PRF) or pulse rate: number of pulses sent per second - Echoes (also called pulses or returns): received pulse reflections from multiple objects recorded for one sent pulse - Minimum vertical object separation: minimum distance between 2 separable echoes - Scan rate: number of scan patterns (e.g. scan lines) per second - Field of View (FOV) or scan angle: across-flight angle that laser beam can cover - Beam divergence: the angle showing the deviation of the laser beam from parallelity E. Baltsavias 14 7 Other important parameters - Minimum and maximum flying height: maximum depends mainly on transmitted power, minimum on national/local regulations and eyesafe distance - Maximum swath width: depends on flying height and FOV - Laser footprint (ground area illuminated by laser beam): depends on beam divergence and flying height. In ideal case a circle, in reality an ellipse or even more irregular pattern - Wavelength: important for measuring certain objects (object should reflect well at laser wavelength) - Across and along track point density (these 2 define also the average point density): they depend on many parameters, like scan pattern, PRF, scan rate, flying height, aircraft velocity, FOV etc. -> necessity for good flight planning and selection of acquisition parameters - Number of echoes for which intensity is recorded - GNSS/IMU measurement frequency and accuracy (accuracy espec. for IMU) - Use of additional imaging sensors (digital cameras, video, etc.) - Weight, dimensions, power consumption, environmental operational conditions (Temp., Humid. etc.) - Range resolution and accuracy (note difference!) - Software! (flight planning, post-processing etc.) E. Baltsavias 15 Scanning mechanisms & ground patterns Oscillating mirror Rotating polygon Nutating mirror (Palmer scan) Fiber switch (Toposys Falcon) Laser (same for receiving optics) Z-shaped, sinusoidal Parallel lines “Elliptical” Parallel lines Flight direction E. Baltsavias 16 8 Scanning mechanisms & ground patterns - Most systems use oscillating mirrors with equal angle increments - Some systems offer selectable scan patterns (though differences not large) - Accuracy at the edge of the swath with oscillating mirrors often worse due to deflection inaccuracies, espec. with high inertia mirrors - Point density is inhomogeneous with all scan patterns, for some patterns more for other less (try to compensate this with appropriate flight planning parameters) - There are gaps in ground coverage and depending on laser footprint also overlaps. This is one of the reasons why laser images (intensity) is much inferior to digital camera images - With Palmer scan the point density at swath edge is higher. This is positive for connecting neighbouring overlapping swaths (see strip adjustment below) E. Baltsavias 17 Scanning mechanisms & ground patterns Some systems offer selectable scan patterns (though differences not large), here example Leica ALS E. Baltsavias 18 9 Beam divergence • Laser beam widens with distance • Beam divergence γ • Theoretical limit by diffraction: D, aperture diameter γ ≥ 2.44 λ / D • Large receiving optics aperture generally advantageous (collection of more reflected energy) • Example: h λ = 1064 nm, D = 10 cm 0.026 mrad (γ / 2) • Typical values for ALS: γ = 0.15 – 1 mrad • Ground laser beam diameter (assuming a circle) DI = D + 2h tan(γ / 2) ≈ 2h tan(γ / 2) ≈ h γ DI, diameter of illuminated area • Example: γ = 1 mrad 1 m diameter @ h = 1 km flying height • Small divergence general advantageous: - more homogeneous objects and terrain, less surface smoothing - better XY and Z accuracy Ground E. Baltsavias 19 Polygon mirror example Parameters in green … … 0.08° = constant angular step of mirror … Pre-decided 25 cm @ 0.5 mrad beam divergence θ/2=20° h=500 m 0.08° 70 cm Swath width 2h tan (θ / 2) = 0.7 h = 364m … … … ∝ 1 / cos2 θi 364 m 0.08° 79 cm 83 cm v = 150 km/h ; 50 lines / s E. Baltsavias Pulse repetition frequency 25 kHz 500 pulses / line 20 10 Basic components – The laser ray Laser beam properties (for pulse lasers) - High power, so that enough energy can return back to the detector (high flying height). - Very narrow beam: laser can illuminate and measure small targets, more energy per area. - A very narrow high energy pulse is emitted, with a width of under 10 ns (note 1 ns means 0.3 m distance). The narrower the pulse, the better the range accuracy. - Pulse modelled as a Gaussian function, pulse width measured at half maximum of the amplitude. E. Baltsavias 21 Basic components – The laser ray Spectral properties - Mostly used laser: Nd:YAG = neodymium-doped yttrium aluminium garnet Emits at 1064 nm wavelength - Other systems: e.g. 810 nm (ScaLARS), 900 nm (FLI-MAP), 1540 nm (TopoSys, Riegl) - Laser systems emit in one wavelength only. Exception bathymetric lasers emit at 1064 and 532 nm, to measure both water surface and water bottom - Emitted light has very narrow spectral width, e.g. for Nd:YAG 0.1-0.5 nm - Experimental multispectral laser scanners developed for terrestrial applications (multiple laser diodes, no range measurement). But no commercial ALS product exists. E. Baltsavias 22 11 Spectral properties Substantial differences E. Baltsavias 23 Spectral properties (from JPL Aster spectral library) Note the huge reflectance difference between 1.047-1.064 µm (60%) (most ALS) and 1.540 µm (0.6%) (Toposys Falcon) E. Baltsavias 24 12 Reflectivity Reflectivity vs. Material (for 900nm laser wavelength) MATERIAL REFLECTIVITY Dimension lumber (pine, clean, dry) 94% Snow 80-90% White masonry 85% Limestone, clay up to 75% Deciduous trees typ. 60% Coniferous trees typ. 30% Carbonate sand (dry) 57% Carbonate sand (wet) 41% Beach sands, bare areas in desert typically 50% Rough wood pallet (clean) 25% Concrete, smooth 24% Asphalt with pebbles 17% Lava 8% Black rubber tire wall 2% Range vs. reflectivity Correction factor for maximum laser range, depending on target reflectivity (example for lasers of firm Riegl, 900 nm wavelength, diffuse targets, maximum range in the specifications given for 80% reflectivity). Range proportional to square root of reflectivity. E. Baltsavias 25 Types of reflection Isotropic (Lambertian) reflection (same in all directions) Specular reflection (water, glas etc.) no received signal, if angle α not very small Local normal to surface α α Local normal to surface Mixed (hybrid) reflection (partly isotropic, partly specular) E. Baltsavias 26 13 Power balance Example • PT = 2000 W Ar R 1 - Power transmitted: PT • Atmospheric transmission M = 0.8 • Receiver area Ar = 80 cm2 (for Dr = 10 cm) 2 - Power received on object: M PT • Range = 1 km 3 - Power reflected, assuming Lambertian reflection: Pr = 4 · 10-10 PT = 800 nW PT M ρ / π • Reflectivity ρ = 0.5 • H u g e d i f f e r e n c e b e t w e e n transmitted and received power. Main influence from R. 4 - Power received: Ground Pr = ρ PT (M2Dr2 Dtar2) / (4R4γ2) = ρ PT (M2Ar2) / (πR2) Note influence of R2 (extended target); R3 , R4 (for linear , point target) E. Baltsavias 27 For most ALS systems, pulse shape deteriorates with higher PRF Pulse magnitude decreases with increasing PRF. Clockwise from top left: 33, 50, 70, 100 KHz. E. Baltsavias 28 14 Laser intensity Computation of intensity - usually for each received echo - computed from echo magnitude, or integration of area of each echo (better) - quantisation with 8- to 16-bit Laser intensity has several problems: - Laser footprints with area gaps or overlaps. Footprint has varying irregular shape. - Laser is monochromatic source. Some objects may reflect very low to nothing. - Noise is high (e.g. 10% estimated using homogeneous surfaces) - Intensity for same object may be inhomogeneous (depends on flying height, scan angle etc.). Thus, needs normalisation within one image and across images for multi-temporal analysis (see Hoefle and Pfeifer, 2007). - Saturation with highly specular surfaces E. Baltsavias 29 Laser intensity Intensity rarely used. Possible uses: - Visualisation (poor quality ; see next slide) - Matching of intensity to camera images (e.g. for co-registration) - Detection of tie points for laser strip co-registration, or matching of tie with control points - Boresight calibration (in-flight) using tie points in intensity images (e.g. used by Leica) - Classification of laser points in various object classes (difficult; used more in forestry, glaciology) E. Baltsavias 30 15 Examples of laser intensity image Comparison of digital camera (top) and laser intensity (bottom). Quality much worse than camera images. Objects with low reflectivity at laser wavelength or specular objects reflecting away from the sensor or very far objects appear dark. E. Baltsavias 31 Interaction with targets First echoe from tree canopy 2nd, 3rd etc. echoes from tree branches Last echoe from ground Multiple echoes generally with vegetation (semi-transparent objects), also at abrupt surface discontinuties (e.g. building edges) and overhanging objects (e.g. power lines) E. Baltsavias 32 16 Interaction with targets AT t AR first pulse last pulse t 5m tp Usually returned pulse magnitude lower and width wider (not as shown in figure above) Assuming returned pulse width of 10 ns 3 m Min distance of separable objects Δh = 1.5 m (half the pulse duration) In theory, in reality minimal vertical separation is larger. A more realistic measure is: t/2 < Δh < t , i.e. 1.5m < Δh < 3m E. Baltsavias 33 Interaction with targets AT t AR,1 t1 Detection accuracy ≈ 10-15% t of rise time ≈ 3 - 4.5 cm, for t rise = 1 ns For range estimation, usually the leading edge of the sent pulse and each received echo is used. For edge detection some use a fixed threshold (poor method), others select the threshold at half the rise time (time corresponding to half amplitude). For flat surfaces with good homogeneous reflectivity in the laser footprint, received pulse very similar to sent one -> small rise time (good range accuracy), no range averaging of various targets. E. Baltsavias 34 17 Interaction with targets AT t AR,2 t2 t rise 7 ns Multiple irregular surfaces close to each other reflect an incoming pulse. The reflected pulses are combined to a wider pulse with lower magnitude and longer rise time -> lower range accuracy, range averaging E. Baltsavias t 35 Interaction with targets (same reflectivity as left but for sloped terrain) • Measured range depends on surface slope and roughness, e.g. for the 2 left figures return pulse on the right is wider than on the left and measured range is an average of the range of the laser footprint • Minimum detectable object size depends on reflectivity (e.g. thin power cables are detectable). The returned pulse on the 2 right figures may be detectable if the yellow area has high reflectivity, even if it covers a small area within the laser footprint E. Baltsavias 36 18 Interaction with targets E. Baltsavias 37 Tree penetration - Occlusions Canopy penetration decreases with increasing scan angles and increasing leaves (for deciduous trees). Typical values with laser profiler (with leaves): 30-40% coniferous, 20-25% deciduous (> 60% in winter). Similarly, occlusions increase with increasing scan angle (important for 3D building and city modeling). E. Baltsavias 38 19 ground bush tree crown AR power line Echoes - Conventional ALS t t1 First echoe t2 2nd echoe t3 3rd echoe Data recorder t4 Last pulse Above: returned full continuous waveform, but only 4 discrete echoes are registered. First conventional ALS registered first, last or both echoes. Current conventional ALS register 4-5 echoes. E. Baltsavias 39 ground bush tree crown AR power line Echoes - Full waveform ALS t First echoe Last echoe AR 1 Giga-samples/s t Above: returned full continuous waveform Data recorder Below: discrete sampling of continuous waveform. Unnecessary data (e.g. when only one echo) are discarded. E. Baltsavias 40 20 Conventional versus full waveform digitising E. Baltsavias 41 Example of full waveform digitising – Tree profile vertical profile of raw full-waveform ALS data E. Baltsavias 42 21 Multiple Pulses in Air (MPiA) technology • Introduced by Leica in 2006 – Send next pulse before receiving the previous one – Allows Laser system to operate at double the pulse rate of current Leica systems at any given altitude ( from 1000m?) up to 5,000 m? • Used also by other major ALS manufacturers (Optech, Riegl) with other terms than MPiA, • Other possibilities for multiple pulses - use two ALS systems, sharing IMU and some electronic parts - splitting laser pulse in two, one at nadir, one slightly off (mentioned in literature for Leica ALS but implemented)? • Benefits – Double the data density at current swath – Double the swath at current density – Data acquisition cost savings Problem with sending 2 pulses before receiving one, in steep terrain. E. Baltsavias 43 MPiA: Single-pulse technology limits pulse rate 2 E. Baltsavias 5 3 1 4 44 22 MPiA allows doubling of pulse rate 1 2 3 2 3 4 4 5 5 E. Baltsavias 45 Other multiple pulse technology (used by Optech) First consider „What influences maximum flying height ?“ 1. The so-called timing limit - Normally one pulse must be received before the next is transmitted - This poses restrictions to the PRF in relation to flying height, i.e. PRF must be decreased with higher flying height. - Example: - Assume 100kHz PRF - This leaves 10 microseconds between 2 consecutive laser shots - In 10 microseconds light travels 3 km. Dividing by 2 (travel back and forth) gives 1.5 km flying height. Accounting for atmospheric interference actual limit is closer to 1.1 km. E. Baltsavias 46 23 What influences maximum flying height ? 2. The received signal strength is influenced by A) the PRF and B) the sensitivity of the laser detector (including noise level), assuming a given object reflectance. A) Normally the strength (amplitude) decreases with increasing PRF almost linearly Leica has developed a technology where this does not occur. B) The received power decreases with the SQUARE of the distance to the object In most ALS systems, the timing limit occurs before the signal strength limit. E. Baltsavias 47 Other multiple pulse technology (used by Optech) For most ALS systems, the timing limit occurs before the signal strength limit. A technology where both limits are similar is the Optech multipulse technology. A pulse can be recorded before the next is sent. According to informal sources, the Optech multi-pulse technology uses range gating and tracking. Some tests at Swisstopo with such a system have led to many data errors, occuring at abrupt and large surface discontinuities. E. Baltsavias 48 24 Error sources in 3D point estimation • Range errors: due to clock errors, low reflection/large range, double bounce (called also multi-path, leading to longer range), rough or discontinuous surfaces, secondary received reflections (sun, clouds), atmospheric attenuation, detector and electronic errors. • Errors in laser beam deflection (e.g. mirror angle measurement) or deflection mechanism calibration • DGNSS (receiver type, satellite constellation and topography, ground reference constellation and distance to aircraft, frequency, satellite signal disturbances) • IMU (accuracy, frequency, drift) • Calibration values between GNSS, IMU, laser scanner (offsets, alignment) • Dynamic bend of IMU / scanner mounting plate, effects from temperature, pressure, humidity • Missing or wrong in-flight calibration • Method and software to combine GNSS, IMU and calibration data to estimate position and orientation of each laser pulse • Time synchronization and interpolation (Frequency: .g. GNSS: 1-10/s, IMU 200-500/s, Laser finder very hight, e.g. 100,000/s ; interpolation errors increase with decreasing GNSS/IMU frequency and more turbulent flight) • Transformation to local map coordinate system Etc. E. Baltsavias 49 Error budget (geometry) Simulation using only some error sources Z κ Δκ ϕ ΔX0, ΔY0, ΔZ0 ω h Δβ Δϕ Δω Y ΔR ΔZ ΔY X ΔX E. Baltsavias 50 25 due to error in Error budget (geometry) β Δω Δϕ 0 15 ΔX Δκ Total Total @ h=1000m 22.4 53.0 23.6 56.2 27.6 66.6 26.4 63.5 26.4 63.5 2.5 26.5 63.6 0 5 9.4 9.4 4 5 11.7 19.1 8 4 17.0 37.3 Δβ 0 20.9 7.5 0 ΔZ0 0 8 0 0 0 20.9 0 0 14 30 ΔZ ΔY0 16.1 0 15 ΔX0 0 30 ΔY ΔR 0 0 15 5.6 30 12.1 0 0 1.3 0 8 0 0 0 Assumptions: h = 400 m (except last column h = 1000 m), ω = ϕ = κ =0, Δω = Δϕ = 0.03°, Δκ = 0.04°, Δβ = 0.02°, ΔR = 5 cm, ΔX0 = ΔY0 = ΔZ0 = 8 cm 8 cm 0 0-5 5-10 10-15 15+ E. Baltsavias due to error in 51 Error budget: conclusions β Δω Δϕ Δκ 0 20.9 7.5 0 15 ΔX Total Total @h=1000 22.4 53.0 23.6 56.2 27.6 66.6 26.4 63.5 26.4 63.5 2.5 26.5 63.6 0 5 9.4 9.4 4 5 11.7 19.1 8 4 17.0 37.3 Δβ ΔR ΔX0 ΔY0 ΔZ0 0 0 8 0 0 0 30 16.1 0 15 ΔY 0 20.9 0 0 14 30 ΔZ 0 0 15 5.6 30 12.1 0 0 1.3 0 0 8 0 0 8 • ΔY slightly larger than ΔX for small β (due to Δβ error) • Above relation changes for larger β (due to Δκ error) E. Baltsavias 52 26 due to error in Error budget: conclusions β Δω Δϕ 0 15 ΔX Δκ Δβ 0 20.9 7.5 0 ΔZ0 0 8 0 0 0 20.9 0 0 14 30 ΔZ ΔY0 16.1 0 15 ΔX0 0 30 ΔY ΔR 1.3 0 8 0 2.5 0 0 15 5.6 30 12.1 0 0 0 5 4 5 8 4 0 0 8 Total Total @h=1000 22.4 53.0 23.6 56.2 27.6 66.6 26.4 63.5 26.4 63.5 26.5 63.6 9.4 9.4 11.7 19.1 17.0 37.3 • ΔR has only marginal influence on ΔZ • And almost no influence on ΔX, ΔY E. Baltsavias due to error in 53 Error budget: conclusions β Δω Δϕ Δκ 0 20.9 7.5 0 15 ΔX Total Total @h=1000 22.4 53.0 23.6 56.2 27.6 66.6 26.4 63.5 26.4 63.5 2.5 26.5 63.6 0 5 9.4 9.4 4 5 11.7 19.1 8 4 17.0 37.3 Δβ ΔR ΔX0 ΔY0 ΔZ0 0 0 8 0 0 0 30 16.1 0 15 ΔY 0 20.9 0 0 14 30 ΔZ 0 0 15 5.6 30 12.1 0 0 1.3 0 0 8 0 0 8 • ΔZ smaller than ΔX, ΔY and less dependent on h Reason: for ΔZ, ΔR, ΔZ0 dominate and are nearly independent of h • ΔZ mainly depends on ΔZ0 (GNSS!) (and ΔR) for small β E. Baltsavias 54 27 due to error in Error budget: conclusions β Δω Δϕ 0 15 ΔX Δκ Δβ 0 20.9 7.5 0 ΔZ0 0 8 0 0 0 20.9 0 0 14 30 ΔZ ΔY0 16.1 0 15 ΔX0 0 30 ΔY ΔR 1.3 0 8 0 2.5 0 0 15 5.6 30 12.1 0 0 0 5 4 5 8 4 0 0 8 Total Total @h=1000 22.4 53.0 23.6 56.2 27.6 66.6 ΔZ 26.4 63.5 26.4 ΔX, ΔY 26.5 63.5 63.6 9.4 9.4 11.7 19.1 17.0 37.3 • ΔZ given is too optimistic • Especially for sloped terrain, ΔX, ΔY dominate and cause also height errors E. Baltsavias 55 PiA = Pulse in the Air 1PiA: one pulse in the air 2PiA: two pulses in the air - Z- accuracy better than XY (espec. as height increases). Both deteriorate with height, Z only a little. - Accuracy worse at FOV edge than at nadir. Accuracy deteriorates with higher PRF. - 2PiA more accurate than 1PiA for same height and PRF. Or provides same accuracy for higher height. E. Baltsavias 56 28 Accuracy of ALS Main products: - raw laser point cloud (how close is laser point to „true“ position/footprint) - classification of point clouds (degree of correctness (e.g. probabiliy) of classification) Derived products - geometric (DTM, DSM etc.) - thematic/semantic (e.g. object detection and classification) DTM accuracy - Difference of DTM to „true“ surface (differences between measured check points and their interpolated height in DTM). Difficulties: surface modelling errors, surface roughness E. Baltsavias 57 Accuracy of ALS DTM accuracy Empirical accuracy will deteriorate when surface is poorly defined (rough) and laser point density is decreased (bigger interpolation errors). E. Baltsavias 58 29 Precision of ALS A priori estimation of point precision based on error assumptions and error propagation. Can help decide on optimum data acquisition parameters. E. Baltsavias 59 Empirical accuracy - Summary - Accuracy given often as 1 sigma - Companies give own empirical formulas and/or graphs - Z- accuracy better than XY (espec. as flying height increases ; better by a factor 2-5). Both deteriorate with flying height, Z only a little. - Accuracy worse at strip edge than at nadir. - Accuracy (DTM) worse with smaller tree penetration rate. - Z-accuracy worse with increasing terrain slope. And at abrupt surface discontinuties (building borders). - Accuracy deteriorates slightly with higher PRF. - For grid intepolation, Z depends on density of raw data, grid spacing and interpolation quality. Grid spacing should be ideally larger (≥ 2) than average point distance of raw points. - Typical Z-accuracies on undulating bare terrain for up to 3,000 flying height: 5-20 cm - For some detailed investigations on DEM quality and accuracy espec. related to ALS see Kraus et al. (2004, 2006), Karel et al. (2006), Karel and Kraus (2007). E. Baltsavias 60 30 ALS Data Acquisition and Processing Workflow E. Baltsavias 61 Coarse Processing Flow (first stages to produce a point cloud) E. Baltsavias 62 31 Processing flow (example Leica and Terrasolid software) E. Baltsavias 63 Calibration • Very important for high accuracy. Can include: - Factory tuning and fixing of values Intensity-based range correction (performed by Leica: bright/dark objects reflect faster/ slower, range is shorter/longer), other range offset calibrations Scanner encoder offset / scan angle correction (can include scale correction) Electronic components (several components tested/tuned and their values fixed) - Mounting of sensors on aircraft (misalignment, offsets between GNSS/IMU and laser) - In-flight calibration Misalignment, offsets, range and angle offsets and scale (mostly refined in strip adjustment) Calibration not standardized, often company specific. For some details, see presentation of Yuji Kuwano “System Overview” in the Kanpur Laser Scanning tutorial and P. Friess at the Ljubljana tutorial (see references at the end). E. Baltsavias 64 32 GNSS / IMU data processing • Combined adjustment of GNSS / IMU data - Kalman filtering • Combination with offsets / misalignment to derive orientation for each laser pulse • Some of the above parameters refined during the strip adjustment E. Baltsavias 65 Strip adjustment Create a seamless data set by correcting for systematic errors between strips E. Baltsavias 66 33 Check relative strip orientation Color coded DEM differences in strip overlap before and after strip adjustment E. Baltsavias 67 Absolute orientation check after strip adjustment With respect to a map coordinate system. Easy check by overlaying accurate vector map data Wrong absolute orientation E. Baltsavias 68 34 Absolute orientation check after strip adjustment With respect to a map coordinate system. Easy check by overlaying accurate vector map data Correct absolute orientation E. Baltsavias 69 Filtering – Introduction • Digital terrain model (DTM): “ground” • Digital surface model (DSM): “top visible surface” • Digital elevation model (DEM): here used as both DSM and DTM • Filtering: classification of points into terrain and above-terrain (sometimes with separation to buildings and trees ; few other non-terrain objects may exist) • Basis for DSM and DTM generation, detection of above-terrain objects (mainly buildings and trees) by subtracting (DSM - DTM) = normalised DSM (nDSM). Special case Canopy Height Model (CHM): in nDSM remove buildings and other non tree/bush objects. Other related important issues: - Intelligent data thin-out (huge datasets!) - Automated detection of breaklines E. Baltsavias 70 35 Filtering – Introduction DSM DTM E. Baltsavias 71 Filtering – Introduction DSM E. Baltsavias Normalised DSM (nDSM) = DSM - DTM 72 36 Filtering – Introduction Canopy Height Model (CHM) = nDSM - all non-tree objects (technical constructions, rocks). E. Baltsavias 73 Filtering – Introduction DTM E. Baltsavias DTM improved with breaklines 74 37 Filtering E. Baltsavias 75 Filtering Connection of neighbouring points does not produce a meaningful surface. E. Baltsavias 76 38 Filtering Points that do not belong to DSM or DTM, which have to be eliminated E. Baltsavias 77 Filtering – Introduction Various aims: • Filtering to extract ground points • Filtering to extract ground surface • Classification to label points E. Baltsavias 78 39 Filtering results Left: raw point density 5.6 points/m2 (top), reduced to 1 point/16m2 (bottom). Right (DTM after filtering): Filtering errors increase with lower point density. Filtering: compromise between good elimination of non-terrain objects and preservation of terrain details. E. Baltsavias 79 Filtering results Large building (railway station) Part of Stuttgart (above raw data, below DTM). Errors increase with size of non-terrain objects, terrain slope (when buidlings/trees) and scene complexity (dense or overlapping buildings/trees). E. Baltsavias 80 40 Filtering results Classification sometimes to: ground objects (red), vegetation (yellow) AND other above-terrain objects (cyan), or buildings and other above-terrain objects (e.g. in SCOP++ Lidar, Inpho) E. Baltsavias 81 Filtering results DTM raw data of Swisstopo (Zurich airport), right zoom. Problems to interpolate DTM, when large non-terrain objects are present, and terrain not flat. E. Baltsavias 82 41 ISPRS laser data filtering test Comparison of filter algorithms, 2002-2004 • 8 sites, 8 participants • Qualitative and quantitative evaluation Conclusions: • All filters perform well on smooth terrain with vegetation and buildings. All filters have problems with rough terrain and complex city landscapes. • In general, filters that compare points to locally estimated surfaces performed best. • The problems caused by the scene complexities were larger than those caused by the reduced point density. • Research on segmentation, quality assessment and usage of additional knowledge sources is recommended. • Full report on http://www.itc.nl/isprswgIII-3/filtertest/index.html Commercial software (see below) can have, depending on scene complexity, 90-95% success rate. The rest is corrected partly automatically (for large errors), partly semi-automatically or manually using visualisation techniques, overlay with images and maps etc. Very important: how good / fast are software tools for manual editing! E. Baltsavias 83 More details on filtering Input data: raster grid (simpler processing), Triangular Irregular Network (TIN) (original points are used, avoids interpolation errors). dy dx Raster TIN • Basic approaches: • Mathematic morphology • Slope-based filtering (can be implemented as morphological operation) • Progressive densification • Surface-based (robust interpolation) • Segmentation (can actually be implemented with various filtering methods) • Point(s) / local neighbourhood versus segments • Problems: selection of thresholds, selection or local neighbourhood size, treatment of large objects, mixed objects, low objects on terrain, bridges/overpasses, preservations of discontinuities E. Baltsavias 84 42 Test neighborhood and the number of points filtered at a time Point-to-Point: • Two points are compared at a time. • Based on the positions of the two points. • Only one point can be classified at a time. Point-to-Points: Points-to-Points: Measure of Discontinuity – Most algorithms classify based on some measure of discontinuity. – Some of the measures of discontinuity used are, height difference, slope, shortest distance to TIN facets, and shortest distance to parameterised surfaces. E. Baltsavias 85 Slope Based Filtering with surfaces Block Minimum Surface based E. Baltsavias 86 43 Mathematical Morphology (binary values) – s = 2D structuring element (like filter mask) = - = Erosion operation + = Dilation operation I s + I-s I-s = s Opening = Erosion, then Dilation I o s := ( I − s ) + s E. Baltsavias 87 Mathematical Morphology (grey values) DSM Erosion (local minimum) Structuring Element (SE) Dilation (local maximum) DTM Opening Possible modifications of conventional morphological filtering: variable SE size, use with irregular points, use of rank filters instead of min and max, use of multiple iterations. E. Baltsavias 88 44 Mathematical Morphology - Example of opening Original DSM The larger the SE, the stronger the filtering of above-terrain objects, but also the smoothing of terrain details. 11x11 m2 15x15 m2 21x21 m2 31x31 m2 E. Baltsavias Slope-based Filtering Filter kernel showing the max. permissible slope 89 Slope as max permissible height difference as function of distance d. Δhmax (d ) d Point with high slope is eliminated from DTM Δhmax (d ) d DTM E. Baltsavias Account for measurement noise 90 45 Slope-based Filtering - Example E. Baltsavias 91 TIN densification (implemented in TerraScan software, Terrasolid) • Start using sparse seed points • lowest points in a large grid, based on largest structure, e.g. 50-100 m • Densify iteratively from below • calculate required thresholds from points currently included in the TIN • add points to the TIN, if they are within thresholds • Threshold computation based on median values of surface normal angles and elevation differences. Uses histograms for computation of median. α E. Baltsavias β 92 46 TIN densification • Add one point at a time in each triangle facet • Accept based on distance and angle threshold • Special case for discontinuous surfaces (urban areas -> buildings) • Then, usual threshold values easily exceeded • Use mirroring of examined point at closest point in TIN triangle to compute deviation d γ d α mirror d´ β E. Baltsavias 93 Hierarchical Robust Interpolation (implemented in SCOP++ Lidar software, Inpho) - Filtering -> here hierarchical robust interpolation (based on Kriging) with an eccentric and asymmetrical weight function: 1. Optionally remove buildings (in Inpho software) 2. Remove trend surface (low-order polynomial) 3. Do interpolation and filtering 4. Perform robust interpolation (or if clusters of blunders exist, do it hierarchically). Requires good mixture of terrain and above-terrain points 5. Create DEM pyramids with decreasing resolution (Thin-out) (often 3 levels suffice) 6. Generate a coarse initial DEM (Interpolation) E. Baltsavias 94 47 Hierarchical Robust Interpolation (implemented in SCOP++ Lidar software, Inpho) 7. (Filtering) Classify points ( on and above ground) and apply robust interpolation to generate a DEM, starting at the coarsest level, using an asymmetric weight function: • low weights assigned to points that are significantly above the terrain • high weights assigned to points on or below the approximated DTM 8. (Sort-Out) Accept in the original data only points lying within a band of this interpolated DEM (+/- 3 sigma of height accuracy of input points). Steps 6-8 are repeated for each finer pyramid level -> Output DSM (buildings, other non-terrain points) and DTM Consideration of breaklines, if available, for improving DTM quality. E. Baltsavias 95 Robust interpolation • Generation of a low resolution data pyramid using the original data (e.g. xyz-coordinates of the lowest points in a grid) • Computation of a low resolution DTM using robust interpolation along with blunder detection • Elimination of LIDAR points outside a predefined tolerance band (here +/- 1 m) • Computation of a DTM with full resolution using robust interpolation along with blunder detection E. Baltsavias 96 48 Interpolation and Filtering - Data pyramid - Compute data (point set) at different levels: Select within each cell, lowest or mean point or closest to cell center - Process (filter) different levels from coarse to fine, e.g. 20m grid -> 4m grid -> original data E. Baltsavias 97 Interpolation and Filtering - Principle kkahhmm kkahhmm Left: interpolation. Function passes through values of input data (considered error-free). Right: Interpolation and filtering. Function does not pass through values of input data, because a Z measurement error is considered. E. Baltsavias 98 49 Interpolation and Filtering • Initial interpolation using unit weights (same weight for all points i) : σi2 = σ02, with σ02 a priori accuracy of points Outlier, probably vegetation Interpolation follows outlier Result of interpolation + filtering (Source: I.P.F. TU Vienna) Trend surface E. Baltsavias 99 Robust interpolation • Calculate filter values = oriented distance from measured point to surface • Compute histogram of filter values • Asymmetric (different weight function left and right of origin): many points above surface, few points below surface • Eccentric: origin is not at 0 but at g (Source: I.P.F. TU Vienna) E. Baltsavias 100 50 Robust interpolation • Use asymmetric weight function (a,b different for left & right branch) pi = 1 b 1 + (a ⋅ f i − g ) , σ i2 = σ 02 pi • a, b parameters, g shift determined from histogram • Also remove points which are too far off the surface (both positive and negative): cut-off, tolerance Cut-off (Source: I.P.F. TU Vienna) E. Baltsavias 101 Robust interpolation Initial interpolation Refined (robust) interpolation (Source: I.P.F. TU Vienna) E. Baltsavias 102 51 Robust interpolation Weight function p E. Baltsavias 103 Robust interpolation Weight parameters (and some other parameters) change from pyramid level to level 5m level select mean point in 5m x 5m cell robust filtering weight function half weight @75cm weight function tolerance 1m select Points ±3m of DTM 2m level select lowest point in 2m x 2m cell robust filtering weight function half weight @30cm weight function tolerance 60cm select Points ± 2m of DTM original (0.5m level) robust filtering weight function half weight @20cm weight function tolerance 30cm E. Baltsavias 104 52 Robust interpolation Pyramid level 1 (lowest resolution). In each grid cell, the lowest point is selected. E. Baltsavias 105 Robust interpolation Pyramid level 2 E. Baltsavias 106 53 Filter results in forested area (Source: I.P.F. TU Vienna) E. Baltsavias 107 ALS raw data filtering - Various predefined strategies (SCOP++ Lidar) From left: weak, default, strong strategy. Compare differences regarding above-terrain object filtering, espec. Between strong and other strategies. E. Baltsavias 108 54 ALS raw data filtering - Different filter parameter values (SCOP++ Lidar) The filter parameter for this software determines the compromise between elimination of above-terrain objects and preservation of terrain details. Strategy used is default. From left, filter parameter values: 2, 1, 0.5. E. Baltsavias 109 Segmentation-based Filtering Problem analysis of many filters • Smooth surface assumption does not hold • Lack of context information – Filtering applied only locally – Point-wise filtering New filter approach • Use continuous surfaces instead of smooth surfaces • Filter continuous segments of points instead of points E. Baltsavias 110 55 Segment based filtering • Texture based image segmentation • Point cloud segmentation into continuous surfaces E. Baltsavias Slide provided by George Sithole, George Vosselman 111 Profile segmentation E. Baltsavias Delaunay Triangulation Minimum Spanning Tree Proximity Thresholding Remove Dangling Edges 112 56 Profile classification Raised Lowered Terraced High Low E. Baltsavias 113 Combining profiles to segments E. Baltsavias 114 57 Combining profiles to segments E. Baltsavias 115 Segment classification • Based on majority of segment profile classifications E. Baltsavias 116 58 Bridge detection • Select all terrain profiles • Analyse profile segment classifications E. Baltsavias 117 Conclusions for segment based filtering • • • Segment-based filtering – preserves discontinuities – allows filtering of large objects – can be combined with other filtering methods – could be extended with other attributes (shape, size, colour) Segmentation in areas with low vegetation remains difficult Bridges can be recognised in bare earth segment E. Baltsavias 118 59