Processing Workflows for Airborne Remote Sensing
Transcription
Processing Workflows for Airborne Remote Sensing
08/07/2011 Processing Workflows for Airborne Remote Sensing Koen Meuleman, Jan Biesemans, Tim Deroose, Kristin Vreys, Walter Horsten, Sindy Sterckx, Dries Raymaeckers Content Introduction: motivation Processing workflows: functional flows Processing workflows: middleware Processing workflows: hardware Archiving workflow (Level0 – Level1) Processing workflow (Level1 – Level2/3/4) Algorithms Level0 – Level1: • Interest point selection • Co-registration • Block bundle adjustment Algorithms Level1 – Level2: • ortho-rectification • DEM/DSM generation • MODTRAN4 • water vapor, visibility • haze removal, shadow removal • atmospheric BRDF, topographic BRDF, target BRDF Algorithms Level2 – Level3: mosaic generation Algorithms Level2/3 – Level4: classification algorithms, change detection, … Algorithmic components: Software development requirements 08/07/2011 © 2009, VITO NV – All rights reserved 2 Introduction: motivation Video Pushbroom whiskbroom Frame Camera The software development process of the processing workflows for airborne remote sensing started at July 2004 and was triggered by: • PEGASUS, MERCATOR-1 (Flemish Government): High-Altitude Long Endurance Solar Powered UAV (operations between 14 and 20 km) • MEDUSA Phase C/D (ESA PRODEX): design & assembly of an ultra-light RGB camera (< 2 kg, data throughput is around 20 MBit/s) for UAV platforms (i.e. the Mercator-1 platform) in support of photogrammetric and disaster management applications. • AGIV (Flemish Government): prototyping photogrammetric workflows for GRB anomaly detection. The GRB is the large scale geographic reference database of Flanders containing a set of GIS and CAD layers for roads, road elements, buildings, railways, waterways, and parcels • APEX (ESA PRODEX): 312 spectral rows in the VNIR and 195 spectral rows in the SWIR. Data throughput is around 500 MBit/s. • BELSPO funded hyperspectral campaigns: CASI-SASI in 2002, CASI-ATM in 2003, HYMAP in 2004, AHS160 and CASI550 campaign in 2005, AHS160 & CASI550 & AISA & UltraCAM campaigns in 2007. • FP6 OSIRIS: Sensor Web: to enable the user to deploy, task and query sensors in support of disaster management applications which require near-real-time information (to be demonstrated in 2009 in a forest fire application). • RIMS (BELSPO): Generating image mosaics from video in support of disaster management or situation awareness services. • ISYSS (IBBT): sensor positioning, event detection, hardware implementations 08/07/2011 © 2009, VITO NV – All rights reserved 3 Introduction: development strategy Project-driven: user requirements and system requirements are not always constant (especially in prototyping research projects) evolutionary life-cycle model (i.e. build a larger prototype system by pooling single projects to serve a common objective) 08/07/2011 © 2009, VITO NV – All rights reserved 4 Introduction: development strategy Archive: raw data + metadata & data products APEX Operational platform V1 AGIV Operational platform V1 V2 V3 SLA (Service Level Agreement) 1 SLA 2 SLA 3 2011 2012 Research, assembly and integration test platform Innovation by means of classical “project work” (FP7, BELSPO, IWT, …): AGIV, VITO, universities, companies, … 08/07/2011 © 2009, VITO NV – All rights reserved 5 Processing Workflows: functional flow “classical” airborne missions 08/07/2011 © 2009, VITO NV – All rights reserved 6 Processing Workflows: functional flow “classical” airborne missions Archiving workflow “Classical” airborne applications: • Very high resolution and high-precision mapping using photogrammetric camera systems (DMC, Ultracam, …) for civil structure mapping and civil structure change mapping. • Calibration Support Tools: Support tools for checking the interior and exterior orientation parameters and the spectral and radiometric calibration. • Archiving: The generation of standardized Level1 HDF5 files (i.e. “archive objects”) from the raw incoming image data (Level0) and image metadata in the framework of long-term archiving strategies. Not all data arrives at Level0. Sometimes, only Level2, Level3 or intermediate image products are available. After the calibration checks, the incoming data is transformed to self-descriptive Level2 or Level3 HDF5 archive objects. 3D workflow The automated image-based generation of area-wide DEM and DSM models at a user specified spatial resolution. Datafusion aspects with LIDAR are not yet included. Processing workflow This workflow is subdivided in three sub-systems: • Level1 to Level2 processing workflow: the automated generation of orthorectified and atmospheric corrected single images. • mapping of the quality physical environment (air quality, water quality, land cover quality mapping) using multispectral and hyperspectral camera systems (APEX, HYMAP, AHS160, HYSPEX) • Level2 to Level3 workflow: the generation of image mosaic products. automated Level2/3 to Level4 change detection workflow: a. the generation of a soft classification based on Level2/3 image data, auxiliary data (DSM, texture parameters, …) and a “field-truth” data set. b. Modules for interpreting the soft classification and the creation of a change-vector-layer HDF5 (www.hdfgroup.org) is used for all product packaging and archiving. • Main user requirement is “high-precision” both in georeferencing and radiometric/spectral calibration 08/07/2011 8/07/2 011 © 2009, VITO NV – All rights reserved 7 Processing Workflows: functional flow near-real-time situation awareness services Sensor Web Enablement (SWE): Mercator-Low (1/4 scale model of the Mercator-1) equipped with a PAL video camera (Military test range, Limburg, Belgium, 6 June 2006) 08/07/2011 © 2009, VITO NV – All rights reserved 8 08/07/2011 © 2009, VITO NV – All rights reserved 9 Processing Workflows: middleware for distributed computing Middleware is computer software that connects software components or applications. The software consists of a set of enabling services that allow multiple processes running on one or more machines to interact across a network. • Airborne missions generate thousands of images need for distributed computing need to chose patterns for parallelism: Master/Worker: Master application constructs a job-list and maintains the jobdependency. Worker applications ask the master for a job and execute this job. Task/Data Decomposition: algorithmic module is executed on smaller subsets of data. Master applications implements the task and data decomposition. • Parallelism is implemented in the middleware, NOT in the applications (this keeps the C++/C/Fortran/Java/IDL code of the applications as simple as possible) 08/07/2011 © 2009, VITO NV – All rights reserved 10 Processing Workflows: middleware for distributed computing Middleware for cluster/distributed computing is developed by VITO: • • • • • Message passing (over reliable TCP/IP sockets) and job-pulling Master-Worker pattern developed in Java (VITO, Dept. Remote Sensing) Multiple Masters can run next to each other. Multiple masters can run on one machine. Masters can be configured to take only processing jobs from specific registered users/operators. One worker per machine. Workers keep multiple threads alive which invoke the running of applications. The number of threads can be altered on-the-fly. Workflow Monitoring Software: Java GUI application for monitoring and configuring the processing cluster 08/07/2011 © 2009, VITO NV – All rights reserved 11 Processing Workflows: Hardware Components » » » » » Database Server Master Nodes (are assigned to specific user and workflow types) Worker Nodes (are assigned to a specific Master) Web Server Storage » Long Term Data Archive: » optimized for growth and expansion » contains L1/L2/L3/L4 archived products » Short Term Storage: » optimized for throughput » Contains temporary data between processing steps 08/07/2011 © 2009, VITO NV – All rights reserved 12 Processing Workflows: Hardware Components Prototype Research and Development Platform Operational Platform Total 172 cores on 19 machines 08/07/2011 © 2009, VITO NV – All rights reserved 13 Archiving workflow: geometric calibration Quality control and transformation to internal standards: • Interior orientation (FOV, focal length, CCD properties, …) • Exterior orientation 1. dGPS correction (POSPAC) 2. GPS X, Y, Z & IMU roll, pitch, yaw 3. Block bundle adjustment 4. Boresight angles • Coordinate transformations (internal standard for linesensors: native GPS coordinate system; for framecameras: in the coordinate system of the block bundle adjustment but with the full specification for transformation towards GPS coordinate system). Geometric calibration using GCP’s via a Monte Carlo simulation (10000 runs). APEX test-flight example (October 2008). 08/07/2011 © 2009, VITO NV – All rights reserved 14 APEX PAF - level 0 -1 Processing RSL development VITO development Archiving workflow -) Level 1 X data -) Applanix data -) Config XML’s APEX level 1 standard product (HDF5) Can be delivered to the users (ftp, harddisk) PosPac Processing 08/07/2011 © 2009, VITO NV – All rights reserved 15 Archiving workflow: Spectral calibration 14 Spectral calibration, i.e. the determination of shifts in central wavelength and bandwidth, only possible if spectral resolution is fine enough (i.e. to see the effect in absorption regions) 12 at-sensor radiance 10 O2 8 H20 H20 6 4 Fraunhofer line O2 2 0 400 H20 450 500 550 600 650 700 750 800 850 900 950 1000 wavelength (nm) CASI reflectance spectra of a sand pixel. Left: original with wrong wavelength calibration file. Right: after recalibration of the sensor by the subcontractor: reflectance spectra in blue region are correct and less spikes are observed near the absorption bands (however still some spikes are visible near the O2 band). 08/07/2011 © 2009, VITO NV – All rights reserved 16 Archiving workflow: file format and unit standardization Function: After quality control, geometric calibration, spectral calibration and radiometric calibration, (a) the packaging of image data and image metadata in one single “archive object”, i.e. a Level1 HDF5 file, (b) the production of an orthorectified quicklook. All products are distributed in the HDF5 self-descriptive format to ensure that data and meta-data are kept together. HDF5 is the standard image product distribution format of USGS and NASA. 17 April 2009 Processing Workflows 08/07/2011 Airborne Remote Sensing © 2009, VITO NV – All rights reserved 17 17 Processing workflow (Level1 to Level2/3/4) Subdivided in 3 workflows grouped in one single WWW user interface: • • • Level1 (raw) to Level2 (geometric and atmospheric corrected block of images) Level2 to Level3 (mosaic of Level2) Level2/3 to Level4 (e.g. change detection products, soft classifications, …) This allows for: (a) Different processing “entry points”. (b) The option to forward Level2, 3 or 4 products in the archive system 08/07/2011 © 2009, VITO NV – All rights reserved 18 Algorithms Level1-Level2: Orthorectification For: • • • A better interfacing with Modtran4, and Given the requirement that all Level2 algorithms have to work on the raw sensor geometry and resampling has to be done at the very end of the workflow, and To support frame sensors, whiskbroom and pushbroom sensors It was decided to develop an in-house C++ module. This C++ module was fully validated against the Inpho OrthoVista package using UltracamD imagery and Parge using AHS data 08/07/2011 © 2009, VITO NV – All rights reserved 19 Algorithms Level1-Level2: DEM/DSM generation VITO currently uses Correlator3D (www.simactive.com) for DEM/DSM extraction (very easily operated and ultra-fast due to GPU aided processing). : C++ implementation of the paper: C. Strecha and L. Van Gool. 2008. A generative model for true orthorectification. ISPRS 2008 (Beijing) 08/07/2011 © 2009, VITO NV – All rights reserved 20 Algorithms Level1-Level2: On-the-fly configuration of MODTRAN4 (AFRL: US Air Force Research Laboratory) Atmospheric Correction = Determining the at-target radiance (LT) by correcting the measured at-sensor spectral radiance for the “path radiance” LP (Haze) and “background radiance” (LB). LP and LT are influenced by: earth-sun distance, solar incident angle (i), solar azimuth, sensor zenith angle (e), sensor viewing azimuth, atmospheric composition (water vapor, O3, CO2, CO, …) and cloud cover. MODTRAN4 has 176 configurable parameters, as such, pre-calculated look-up tables (e.g. ATCOR 6D LUT) are non-generic (a) An XML configuration file can be uploaded to fully customize the MODTRAN4 processing; (b) image based parameter estimation (water vapor, visibility, illumination geometry). Modtran 5 will be implemented in course of 2011 08/07/2011 © 2009, VITO NV – All rights reserved 21 Algorithms Level1-Level2: Overview Atmospheric Correction MODTRAN4: from at-sensor radiance to at-surface reflectance Estimation of the atmospheric composition Elimination of “contaminations” Image-based haze correction (supervised method, not suitable for chain-processing) Image-based Shadow correction (supervised method, not suitable for chain-processing) Correction for BRDF effects Image-based visibility (Aerosol Optical Depth) estimation (only if no spectal GCPs available) Image-based Water Vapor estimation Atmospheric BRDF (via MODTRAN4 viewing-geometry dependent simulations) Topographic BRDF (supervised method, not suitable for chain-processing) Target BRDF (supervised method, not suitable for chain-processing) Color balancing multiple images Multi-resolution spline blending Radiometric triangulation (20XX: under development) Legend: Empirical based on between-band relations Empirical single band based Mixture empirical/physical Stochastic Atmospheric correction is no “exact” science 08/07/2011 © 2009, VITO NV – All rights reserved 22 Algorithms Level1-Level2: Water Vapor, Visibility Water vapor estimation: Rodger, A. and M.J. Lynch, 2001. Determining atmospheric column water vapour in the 0.4-2.5 µm spectral region. Proceedings of the AVIRIS Workshop 2001, Pasadena, California Visibility (aerosol optical depth) estimation: Richter, R., D. Schläpfer and A. Müller. 2006. An automatic atmospheric correction algorithm for visible/NIR imagery. International Journal of Remote Sensing, 27(10), pp. 2077-2085. 08/07/2011 © 2009, VITO NV – All rights reserved 23 Algorithms Level1-Level2: Haze removal Haze removal: Richter, R., 2005. Atmospheric/Topographic Correction for Airborne Imagery. ATCOR-4 User Guide Version 4.0. DLR, Wessling, Germany, 104p. UltracamD 08/07/2011 © 2009, VITO NV – All rights reserved 24 Algorithms Level1-Level2: Shadow removal Shadow removal: Richter, R., 2005. Atmospheric/Topographic Correction for Airborne Imagery. ATCOR-4 User Guide Version 4.0. DLR, Wessling, Germany, 104p. Hymap 08/07/2011 © 2009, VITO NV – All rights reserved 25 Algorithms Level1-Level2: Atmospheric BRDF Atmospheric BRDF: Viewing the same object from various angles may lead to significant variations of the path scattered radiance component, an effect known as atmospheric BRDF. Can be taken into account by detailed MODTRAN4 simulations during the atmospheric correction. Difference images: MODTRAN nadir only simulation minus MODTRAN view and illumination specific simulations for a blue, green and red band. The lower image is the geometric corrected digital number. In that image, flight direction is from left to right (East). Solar zenith was 29.4°and Solar azimuth was 161.5°. Remark that in some areas, the AHS160 sensor completely saturated. MODTRAN4 rather effective to remove “blue haze” 08/07/2011 © 2009, VITO NV – All rights reserved 26 Algorithms Level1-Level2: Topographic BRDF • The VITO orthorectification module determines the complete viewing and illumination geometry. • Richter, R., 2005. Atmospheric/Topographic Correction for Airborne Imagery. ATCOR-4 User Guide Version 4.0. DLR, Wessling, Germany, 104p. 08/07/2011 © 2009, VITO NV – All rights reserved 27 Algorithms Level1-Level2: Target BRDF (Kernel BRDF models) Target BRDF correction according: Jupp, D. L. B., 2000: A compendium of kernel & other (semi-)empirical BRDF Models. Office of Space Science Applications - Earth Observation Centre, available only as online document (May 2002): www.cossa.csiro.au/tasks/brdf/k_summ.pdf Ideal surface: No BRDF effects AMBRALS (the Algorithm for Modeling[MODIS] Bidirectional Reflectance Anisotropies of the Land Surface) allows a variety of kernel-driven semiempirical BRDF models to be explored. Kernel BRDF methods are empirical models based on linear combinations of “kernels”:ρ = fiso + fgeokgeo + fvolkvol which represent surface reflectance (ρ) as a function of component reflectances (fx) and the kernels (kx) which are mathematical functions that depend on sun (or incident) and view (or observer) angles. The subscripts “geo” and “vol” refer to the physical bases for some kernels in which there is an identification of a “geometric” or hotspot factor and a “volume” or path length and scattering factor Grass vegetation: BRDF effects @ 600 nm UltracamD (raw) UltracamD (Kernel BRDF corrected) 08/07/2011 © 2009, VITO NV – All rights reserved 28 Configurable tilling option 08/07/2011 © 2009, VITO NV – All rights reserved 29 Algorithms Level2-Level3: Mosaic generation & automated optimal seamline estimation C++ application: 1. cost grid based on the combination of similarity within image and between images 2. Iterative cost grid masking 3. Ford-Fulkerson Graph-Cut method on masked cost grid. Iterative masked cost grid = searching for the best possible solution = just before connectivity gets lost. 08/07/2011 © 2009, VITO NV – All rights reserved 30 Algorithms Level2/3-Level4 Level1/2/3 – Level4 C++ implementation of following algorithms: • Unsupervised K-means classification • Spectral angular mapper • Maximum likelihood classification • Linear discriminant analysis (LDA) • Quadratic discriminant analysis (QDA) • … 08/07/2011 © 2009, VITO NV – All rights reserved 31 Processing Workflows: functional flow “classical” airborne missions Example: civil structure change detection as operational service for AGIV (www.agiv.be). Processing steps: 1. All pixels visible in multiple images determine pixel elevation from parallax information. 2. Geometric correction of every image. 3. Radiometric correction of every image. 4. Automatic classification with “field truth” extracted from the AGIV vector database of civil infrastructure (i.e. the GRB). 5. Post-classification logics to construct a vector file with new buildings, renovations and building demolition. 08/07/2011 8/07/2 011 © 2009, VITO NV – All rights reserved 32 32 Algorithms Level2/3-Level4 Example: GRB mutation and anomaly detection. GRB (Grootschalig Referentiebestand) change detection: blue polygons are already mapped buildings, green polygons are the result of an automatic building detection process (combined K-means, Quadratic Discriminant Analysis and postclassification logic) 08/07/2011 © 2009, VITO NV – All rights reserved 33