High-Performance Computing for Processing Earth Observation Data By

High-Performance Computing for Processing Earth Observation Data By Dr Ashok Kaushal Senior Divisional Director Enterprise Geospatial & Defense Solutions Rolta India Limited ashok. kaushal@rolta. com Innovative Technologies for Insightful Impact

Agenda • Trends • Needs • Process Automation • Geo. Imaging Accelerators/ GXL • Job Processing Systems/ JPS • Conclusions

Trends • 230 EO [versus 107 in last decade] satellites projected over next decade for use of satellite imagery – Emerging markets expected to account for 75 satellites four-fold increase over last decade – 41 Nations [currently 26] to have own satellites • Commercial sale of EO data expected to double – Commercial EO data from satellites expect CAGR of 15% over next 10 years, reaching $4 billion by 2019 – Optical data will represent 79% of overall sales – Number of high resolution satellites offering commercial data are expected to double from currently 24 ‘Satellites to be Built & Launched by 2019, World Market Survey’, Euroconsult

Trends • Exponential increase of volumes of satellite EO data • Increasing value of EO data with applications in – Agriculture, Environment, Urban Development, Disaster Management, Surveillance and others • Increasing value of up-to-date info – Rapid. Eye, Geo. Eye, Digital Globe, IRS/ Cartosat • Significant growth of awareness in EO data – Google Earth, Microsoft Bing Maps, Bhuvan • Increasing importance of collaboration and sharing of current data/information for Situational Awareness

Needs • Satellite Programming • Timely Data Acquisition • Process Automation • Data Pre-Processing • Data Management • Data Dissemination • Information Sharing • Geo Collaboration

Process Automation For Production Move from this To This

Process Automation Incoming raw image Extract raw Image to native format Collect GCP Using Master Image Refine collected GCP Compute Math Model Orthorectify Raw Image Load Image to Oracle Database

Geoimaging Accelerator (GXL) Geoimaging Accelerators are automated workflows created from linking together of any number of pluggable image processing functions

Geoimaging Accelerator (GXL) Objectives: • Need for large volume image data processing • to reduce image pre-processing bottlenecks • Demand for greater automation & less user interaction • to save money on operator time • Workflows that can scale across multiple processors • to add capacity as and when needed • Plug & Play architecture • to add new components or functions to expand • Cost Effective Solution to Remain Competitive • to run 24/7 with zero or little operator intervention

Geoimaging Accelerator (GXL) • Distributed processing • Two levels • Basic • Automated CPU • Accelerated • Multi-core CPU • Optimized GPU • Ortho / Ortho XL • Satellite & Airphoto • Pan. Sharp / Pan. Sharp XL • Mosaic/ Mosaic XL Ingest GXL Output Job Processing System (JPS)

Geoimaging Accelerator (GXL) RPC Model Calculation Orthorectification DEM World. View-1 Level 1 b e. g. Ortho Product Orthorectification GXL AP Model Calculation Orthorectification DEM Orthorectification GXL Airphoto / Airphoto XL Format & Tile Ultra. Cam X Imagery e. g. Ingest GPS/INS Ortho / Ortho XL Ortho Product

Geoimaging Accelerator (GXL) Pan Sharpening Pan. Sharp GXL Pan and MS Imagery Colour. Balance Mosaic GXL Orthophotos Epipolar Rotation Stereo Pair Cutline Selection Pan. Sharp Product DEM Extraction GXL Mosaic Product Geocoding Raster DEM

Geoimaging Accelerator (GXL) Accelerated GXL? • A hardware-based, GPU enabled, highperformance image processing system • Design to process large volumes – 40 times faster than desktop product – 2 -4 TB per day for desk-side system – 10 TB + for rack mounted system Orthorectify & Mosaic India in a Day!

Geoimaging Accelerator (GXL) Architecture Layer: Component: Integration: Interface Layer Job Processing System Data / Imagery Level Workflows: GXL Bindings: Python, Java C++ SDK GPU / HW PPFs Formats: BIL, TIFF, etc. Algorithms: Pansharp, Ortho, etc. Processing Layer Architecture Layer Operations / Systems Level HW / Architecture Level

Geoimaging Accelerator (GXL) Highlights • Flexible orthorectification: – Support for several sensors (SPOT, QB, Ikonos, WV, …) – Optional radiometric calibration of SPOT images – Optional GCP collection from multiple reference data types • Flexible mosaicking: – – Mosaics from mixed-resolution raw scenes Optional tie point collection and refinement Various types of color balancing Various tiling schemes • High quality: – Sub-pixel accuracy of GCPs and orthoimages – Nicely color-balanced mosaics

Geoimaging Accelerator (GXL) Processing Metrics Product Type Dataset SPOT 5 - Level 1 A 2. 5 meter 8 U Pan IKONOS - Geo Ortho Kit World. View-1 and Quick. Bird Level 1 B Quick. Bird - Ortho. Ready - 4 channel PS Quick. Bird - Level 1 B 16 U Pan Ikonos 16 U Pan 16 U Multispectral Resolution Volume [m] [TB/Day] 2. 5 1. 0 0. 5 0. 6 2. 4 2. 00 2. 94 3. 26 3. 52 4. 57 Area [km 2/day] 13. 7 Million e. g. Europe: 10. 1 M km 2 1. 62 Million e. g. Mongolia: 1. 56 M km 2 448 k e. g. Sweden: 450 k km 2 174 k e. g. Florida: 170 k km 2 3. 62 Million e. g. India: 3. 17 M km 2

Processing Throughput Geoimaging Accelerator (GXL) Product Type Dataset MB/Sec GB/Min TB/Day SPOT 5 - Level 1 A 2. 5 meter 8 U Pan 24. 23 1. 42 2. 00 IKONOS - Geo Ortho Kit 16 U Pan Ikonos 35. 67 2. 09 2. 94 World. View-1 and Quick. Bird Level 1 B 16 U Pan 39. 59 2. 32 3. 26 Quick. Bird - Ortho. Ready - 4 channel PS 16 U Multispectral 42. 67 2. 50 3. 52 Quick. Bird - Level 1 B 16 U Multispectral 55. 47 3. 25 4. 57

Geoimaging Accelerator (GXL) Cost $1, 000 500 GXL Rack Accelerated GXL Basic 200 Batch Processing 10 100 Orthos per day 50 GB Project Scale 5000 Orthos per day Plus 100 Image Mosaic per day 5 TB Project Scale 2000 Orthos per day 1 TB Project Scale GXL Deskside Accelerated 20 Orthos per day 10 GB Project Scale GB 1 - 5 TB Performance /Day 5 - 10 TB

Applications Geo. Imaging Accelerator • Environmental • Carbon sequestration • Biomass estimation • Agricultural • Crop yield • Crop forecasting • Aerospace & Defense • Border monitoring • Disaster management • Data Supply • Product delivery • Archive re-processing

Job Processing System • Distributed Processing System – Run multiple jobs concurrently on multiple servers JPS Database Computer JPS Processing Server Job Computer JPS Processing Server Job Job

Job Processing System • Job: – An entry in the JPS-DB – A Process started and monitored by a Processing Server • Processing Server – Daemon managing jobs Processing Server Job

Job Processing System • • Distributed Cloud Computing (Autonomous Nodes) Automatic Load Balancing Simple Web Interface Threefold Value: 1. Automation = Increased Throughput (Revenue) 2. Job Tracking = Improved QA (Operational Costs) 3. Multi-Platform, Multi-Language = Sustainability Job JPS-DB Job GXL 1 GXL 2 Job Other Nodes Job 22

Job Processing System

Conclusions • Effective use of voluminous satellite imagery from numerous high-resolution satellites desires automated pre-processing using HPC • Distributed processing using multi-core CPU and GPU with CUDA and Open MP provides an ideal platform for faster turn-around-time during pre-processing of geoimaging

Thank you !