Advances in UXO classification

Size: px

Start display at page:

Download "Advances in UXO classification"

Vivien Wells
5 years ago
Views:

1 Advances in UXO classification Stephen Billings, Laurens Beran, Leonard Pasion and Nicolas Lhomme NSGG UXO213 Conference

2 Outline A. Why classification? UXO contamination ESTCP Pilot Discrimination Studies B. Key advances in classification Improved data quality and fidelity Improved processing Dedicated software C. Case studies Understanding the limits of the data A reliable QC process Classification with dynamic data

3 Why classification? The UXO problem in the USA Over 15 million acres of UXO contaminated land due to military training Eminent and substantial health risk with remediation cost estimated to be over 14 billion US$ (EPA 22)

4 Why classification? The UXO problem in the USA Over 15 million acres of UXO contaminated land due to military training Eminent and substantial health risk with remediation cost estimated to be over 14 billion US$ (EPA 22) 23: The Defense Science Board suggested use of classification technology to reduce costs 26: The ESTCP received funding to stimulate development of advanced classification technologies for UXO cleanup ESTCP: The Environmental Security Technology Certification Program (DoD) SERDP: The Strategic Environmental Research and Development Program (DoD/DoE/EPA)

New Boston, NH George West, CO SW Proving Ground, AR Fort McClellan, AL Camp Beale, CA Spencer Range, TN Camp

5 A. Why classification? Evolution of technologies through the Demonstration ESTCP Program Man-portable Mag and EM-61 Multi-static EMI New Boston, NH George West, CO SW Proving Ground, AR Fort McClellan, AL Camp Beale, CA Spencer Range, TN Camp Ellis, IL Helena Valley, MT Camp Lejeune, NC Former Lowry BGR, CO Camp Sibert, AL San Luis Obispo, CA Camp Butner, NC Pole Mountain, WY Mass. Military Res., MA Bellows, HI Waikoloa, HI

6 A. Why classification? Helena New Boston Beale San Luis Obispo George West Yuma Pole Mtn. Lowry Spencer Sibert McClellan Butner Lejeune ESTCP live-site demonstrations (25-213)

classification RxX-TxZ RxY-TxZ RxZ-TxZ Data Acquisition Feature extraction Classification RxX-TxY

7 A. Why classification? Rx1 Rx2 Rx3 Rx4 Rx5.8 RxX-TxX RxY-TxX RxZ-TxX RxX-TxY RxY-TxY RxZ-TxY Processing flow for UXO detection and classification RxX-TxZ RxY-TxZ RxZ-TxZ Data Acquisition Feature extraction Classification RxX-TxY RxY-TxY RxZ-TxY RxX-TxZ RxY-TxZ RxZ-TxZ Ty, Rx Ty, Ry Ty, Rz Tz, Rx Tz, Ry Tz, Rz

8 Key advances in classification Improved data quality and fidelity

Mapper Geometry 3 Tx 7 cubes with 3D Rx Time.16.

9 B. Advances in classification: data A new generation of EMI Sensors Sensor EM61 -MK2 Metal Mapper Geometry 3 Tx 7 cubes with 3D Rx Time milliseconds milliseconds

combinations of Tx/Rx Transverse target excitation is achieved from a single sensor location Practical application -

10 B. Advances in classification: data Advanced EMI Sensors Multi-axis sensors (eg., MetalMapper) - 3 orthogonal transmitters (Tx) - Several 3D receivers (Rx) - Multiple looks at target through combinations of Tx/Rx Transverse target excitation is achieved from a single sensor location Practical application - First pass: DGM with EM61 - Second pass: Cued interrogation with advanced EMI sensor to collect high quality data in static mode (sensor is parked on top of detected anomaly)

11 Key advances in classification Improved data processing

B. Advances in classification: Processing Extraction of target polarizabilities by inversion Geophysical inversion: Point-dipole model can represent a compact metallic object

12 B. Advances in classification: Processing Extraction of target polarizabilities by inversion Geophysical inversion: Point-dipole model can represent a compact metallic object L 2 =L 3 L 1 UXO generally have: A large amplitude, slow-decaying primary polarizability (L 1 ) Equal secondary polarizabilities (L 2 =L 3 ). Predicted data match observed data

13 B. Advances in classification: Processing Extraction of target polarizabilities by inversion Geophysical inversion: Point-dipole model can represent a compact metallic object L 2 =L 3 L 1 Classification: UXO generally have A large amplitude, slow-decaying primary polarizability (L 1 ) Equal secondary polarizabilities (L 2 =L 3 ). L 1 L 2 L 3

B. Advances in classification: Processing Identification of multiple sources Nearby targets must be inverted simultaneously Single Source Two Source Inversion Cell 82 - Target 2136 - Model 1 (SOI)

14 B. Advances in classification: Processing Identification of multiple sources Nearby targets must be inverted simultaneously Single Source Two Source Inversion Cell 82 - Target Model 1 (SOI) Cell 82 - Target Model 2 (2OI-1) Cell 82 - Target Model 3 (2OI-2) 6mm_IVS 6mm_IVS 6mm_IVS 1 Recovered parameters 1 UXO 1 Frag Song et al. Time Nonlinear (ms) Inversion for Multiple Objects Time in Transient (ms) Electromagnetic Induction Time Sensing (ms) of UXO: Technique and Applications IEEE Transactions on Geoscience and Remote Sensing, Issue 99, 1-14

15 Number of UXO digs B. Advances in classification: Processing Statistical classification L 2 =L 3 L 1 UXO features: UXO have slow time-decay Geonics EM61-Mark 2 Number of non-uxo digs

16 B. Advances in classification: Processing Statistical classification L 2 =L 3 L 1 UXO features: UXO have slow time-decay UXO are axis-symmetric MetalMapper Better separation between UXO and most of the clutter

17 B. Advances in classification: Processing Statistical classification L 2 =L 3 L 1 UXO features: UXO have slow time-decay UXO are axis-symmetric Similar early-time L1 UXO have unique polarizability decays EM-61 time range ISO: Industry Standard Object (4 steel pipe)

18 Amplitude Amplitude B. Advances in classification: Processing Statistical classification L 2 =L 3 L 1 UXO features: UXO have unique polarizability decays Library matching Time Time Time

19 B. Advances in classification: Processing Classification with advanced EMI sensors EM-61 Size/Decay Advanced EMI Polarizability matching

20 B. Advances in classification: Software Dedicated software tools for efficient, reliable classification Geosoft UX-Analyze UXOLab

21 Percent of UXOs Correctly Classified (%) B. Advances in classification Performance of advanced classification E.g.: Classification at Pole Mountain, WY (211) PoleMtn1 Sky PolFitAgg None MetalMapper Custom s1 v1 UXO Advanced EMI Pol. matching Classification Number of Clutter Items Incorrectly Classified Advanced classification significantly reduces remediation costs Classification Digging MetalMapper+Classif. Detection Survey

22 Case studies: ESCTP live site demonstrations Site selection: Contamination with live munitions from military practice (from 2 mm and fuzes to 155 mm bombs) Hundreds of anomalies per acres Study over 1-1 acres with 5-3 anomalies Different conditions to test potential technologies (vehicle, cart or man portable) Validation process Additional munitions are seeded for Quality Assurance and Control All detected anomalies are excavated for verification Technology demonstration Sensor data are collected for detection (dynamic survey) and classification (reacquisition to collect the highest data quality with static sensor on top of anomaly) Detection: Submit detection list of potential targets that require further investigation Classification: Submit prioritized dig list and identify potential munitions of concern

23 Case study 1: Understanding the limits of the data ESTCP demonstration at former Camp Beale, CA (211)

Understanding local data limitations The quality of recovered polarizabilities for targets of interest can vary between sites ISO: Industry Standard Object (4 steel pipe) 1 2 1 ISO: MetalMapper L1 L2

24 Understanding local data limitations The quality of recovered polarizabilities for targets of interest can vary between sites ISO: Industry Standard Object (4 steel pipe) ISO: MetalMapper L1 L2 L3 Pole Mountain 211 L1 L2 L ISO: MetalMapper L1 L2 L3 Spencer 212 L1 L2 L ISO: MetalMapper L1 L2 L3 Camp Beale 211 L1 L2 L Ltot Ltot Median Ltot

25 Percent of UXOs Correctly Classified (%) Example: MetalMapper Processing at Camp Beale Data analyst #1: Method: Aggressive library matching Result: 98% correct classification BUT two ISO were missed _fuzesasclutter BealeOpen Sky PolFit None MetalMapperP Custom s1 v1 UXO Two seeded ISO Missed Difficult UXOs: #BE-2532 (ISO stats seed) #BE-1965 (ISO stats seed) Number of Clutter Items Incorrectly Classified ISO: Industry Standard Object (4 steel pipe)

Understanding local data limitations Example: Difficult targets

Model 3 (2OI-2) BE-2532 ISO statistical seed BE-1965 ISO

5 (SOI) Time (ms) 1 ISO IVS ISO IVS All 3 polarizabilities are

26 Understanding local data limitations Example: Difficult targets in MetalMapper Processing at Camp Beale Cell Target Model 3 (2OI-2) BE-2532 ISO statistical seed BE-1965 ISO statistical seed 1 Cell Target Model 1.5 (SOI) Time (ms) 1 ISO IVS ISO IVS All 3 polarizabilities are not always well constrained Switch to using only primary polarizability.1.5 Time (ms)

27 Percent of UXOs Correctly Classified (%) Percent of UXOs Correctly Classified (%) Example: MetalMapper Processing at Camp Beale Data analyst #1: Method: Aggressive library matching Result: 98% correct classification BUT two ISO were missed _fuzesasclutter BealeOpen Sky PolFit None MetalMapperP Custom s1 v1 UXO 1 Data analyst #2: Method: Conservative approach, using all polarizabilities, then primary only Result: 1% correct classification _fuzesasclutter BealeOpen Sky SVM None MetalMapperP Custom s2 v2 UXO Two seeded ISO Missed 6 Difficult UXOs: #BE-2532 (ISO stats seed) #BE-1965 (ISO stats seed) 4 NO UXO Missed No Diffic Number of Clutter Items Incorrectly Classified ISO: Industry Standard Object (4 steel pipe) Number of Clutter Items Incorrectly Classified

28 Case study 2: The importance of a reliable QC process

29 Rx7 Rx7 Rx6 Rx6 Rx5 Rx5 Rx4 Rx3 Rx2 Rx1 Reliable data quality control (QC) process Identify problem soundings (e.g. MetalMapper) TransZ, RecY RxX-TxX RxX-TxX RxY-TxX RxY-TxX RxZ-TxX RxZ-TxX RxX-TxY RxY-TxY RxZ-TxY RxX-TxY RxY-TxY RxZ-TxY RxX-TxZ RxY-TxZ RxZ-TxZ RxX-TxZ RxY-TxZ RxZ-TxZ Defective receiver Inver. # Status CC Misfit Pol Ord Pol 1 Pass.93Inver # Status37mm CC 1 Pass 1. 2 Fail mm 2 Fail M48 1 INVERT 37mm 1 2 M RxX-TxX RxY-TxX RxZ-TxX RxX-TxX RxY-TxX RxZ-TxX Sensor data RxX-TxY RxY-TxY RxZ-TxY RxX-TxY RxY-TxY RxZ-TxY 1-2 Predicted data 1 2 Sphere 15mm 1 2 Sphere RxX-TxZ RxY-TxZ RxZ-TxZ 1 RxX-TxZ RxY-TxZ RxZ-TxZ Recovered Polarizabilities

30 Rx1 Rx2 Rx3 Rx4 Rx5 Rx5 Rx6 Rx6 Rx7 Rx7 Reliable data quality control (QC) process Identify problem soundings (e.g. MetalMapper) RxX-TxX RxX-TxX RxY-TxX RxY-TxX RxZ-TxX RxZ-TxX RxX-TxY RxY-TxY RxZ-TxY RxX-TxY RxY-TxY RxZ-TxY RxX-TxZ RxY-TxZ RxZ-TxZ RxX-TxZ RxY-TxZ RxZ-TxZ Inver. # Status CC Inver. # Misfit Status Pol OrdCCPo 1 Pass Pass37mm 1. 2 Fail Fail - 1. Inversion without x defective receiver 1 2 M48 37mm 1 2 M INVERT RxX-TxX RxY-TxX RxZ-TxX Sensor data RxX-TxY RxY-TxY RxZ-TxY RxX-TxX RxY-TxX RxZ-TxX RxX-TxY RxY-TxY RxZ-TxY Predicted data Sphere 15mm 1 2 Sphere RxX-TxZ RxY-TxZ RxZ-TxZ RxX-TxZ RxY-TxZ RxZ-TxZ 1 Recovered Polarizabilities 1 1 1

31 Case study 3: Classification of dynamic data for EMI sensors ESTCP demonstration at Spencer Range, TN (212)

32 UXO Classification at Spencer Range Test vehicular and man-portable EMI sensors Test classification on dynamically collected EMI data

Decay (t 14 /t 2 ) dynamic MM - Ch1: X-comp 1 1 Cued Interrogation.5.5 Limitations Cued -.5 interrogation survey -.5 requires additional time and cost -1-1 -.5.5 1 Need to rely on EM61 data map Difficulties resolving multiple targets in close proximity dynamic MM - Ch1: horiz.

78 dynamic MM - Ch1: Y-comp dynamic MM -- Ch1: X-comp Z-comp dynamic MM - Ch1: Y-comp 1 1 212-1-23 13:49:15 Model 5 of 6 (Inv #3 / 3 = 3OI: 2 / 3) -1-1 -.5.5 1 dynamic MM.6 - Ch1: vert. comp 1 1.

33 Decay (t 14 /t 2 ) dynamic MM - Ch1: X-comp 1 1 Cued Interrogation.5.5 Limitations Cued -.5 interrogation survey -.5 requires additional time and cost Need to rely on EM61 data map Difficulties resolving multiple targets in close proximity dynamic MM - Ch1: horiz. comp Geologic 1 noise.5 Inaccuracy of EM61 Mark2.5 picks lead.4 to multiple recollects of cued data CC:.955 GOF:.23 Misfit: 6.71 SNF:.78 dynamic MM - Ch1: Y-comp dynamic MM -- Ch1: X-comp Z-comp dynamic MM - Ch1: Y-comp :49:15 Model 5 of 6 (Inv #3 / 3 = 3OI: 2 / 3) dynamic MM.6 - Ch1: vert. comp Spencer Range Example: dynamic EM61 MM MKII - Ch1: - Channel horiz. comp Geonics EM61-MK2 Channel 1 24 feature vectors x x x Dynamic MetalMapper X=.8 Y=.14 coverage Target location Dynamic MetalMapper Channel 5, Vector Sum x x 63 Target depth Z=.13 dynamic MM - Ch1: vert. comp x Is it possible to directly apply.1 classification -1 to dynamically -1 collected data (one-stage approach)? =13 =115 = Size (t 2 ) p/f inv mod msnr zmospunc 1 p 1 / NaN NaN 2 f 2 / NaN NaN -

Collecting dynamic data in difficult terrain with sensors designed to be deployed in a cued

34 Combined Detection and Classification with Advanced EMI Sensors Production rates? How does increased sensitivity to near surface clutter affect processing approach? Collecting dynamic data in difficult terrain with sensors designed to be deployed in a cued mode Reduction of data quality relative to cued surveys: positioning, smaller time window, fewer looks at target (dynamic MM)

35 Classification using Dynamic Data Only TEMTADS 2x2 Result: Dynamic data at Spencer Range was successively used for classification Note: Relatively simple classification problem (low target density, etc.) at Spencer Range site MetalMapper

36 Cost Excavations Advanced Classification at Spencer Range: Reduction in Excavations and Cost 339 Spencer Range MetalMapper Performance 8% fewer excavations with all MEC recovered Non-MEC MEC DGM One Stage Two Stage $24K Spencer Range MetalMapper Cost $11K Digging $133K CUED MM DGM DGM One-Stage Two-Stage

37 Conclusions Advanced UXO classification has been enabled by better sensors, better processing and better software Classification can significantly reduce remediation efforts at live munition sites There are a number key components to successful classification A reliable data Quality Control (QC) process Understanding the limits of the data Selecting an appropriate classification strategy Future Adoption by industry and regulators Finding the limits of classification Underwater applications

38 Acknowledgements The work in this presentation was funded by the Strategic Environmental Research and Development Program (SERDP) and the Environmental Security Technology Certification Program (ESTCP) Contributors: Kevin Kingdon BTG, Barry Zelt BTG, David Lutes BTG, Lin-Ping Song UBC and Doug Oldenburg UBC The University of British Columbia - Geophysical Inversion Facility

39 Black Tusk Geophysics Vancouver Office #41, 1755 West Broadway Vancouver, BC Canada, V6J 4S5 Phone Brisbane Office 131 Ernest Street Manly, QLD AUSTRALIA, 4179 Phone

40 Back up slides

41 Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Percent of TOIs Correctly Classified (%) Cued MetalMapper results from Spencer Range Results by industry geophysicists 1 MM SpencerOpen Shaw Library None MetalMapperU Custom s1 v1 1 MM SpencerOpen NAEVA WithGT None MetalMapperN Custom s2 v MM SpencerOpen USACE LibMatch None MetalMapperN Custom s3 v Number of Non-TOIs Incorrectly Classified No Difficult TOIs MM SpencerOpen URS URSANN1 None MetalMapperU Custom s2 v Number of Non-TOIs Incorrectly Classified MM SpencerOpen 1 2 SIG 3 Intermediate 4 35 None MetalMapperN 6 7 Custom 8 s2 9 v2 1 Number of Non-TOIs Incorrectly Classified MM SpencerOpen SIG Intermediate 7 None MetalMapperN Custom s2 v2 1 2 MM SpencerOpen USACE LibraryMatch None MetalMapperU 7 Custom s3 v3 1 Number of Non-TOIs Incorrectly Classified Difficult TOIs: 5No Difficult TOIs 5 8 #SR-181 (37mm QA seed) Difficult TOIs: #SR-181 (37mm QA seed) Number of Non-TOIs Incorrectly Classified 2 8 MM SpencerOpen NAEVA WithGT None 7 MetalMapperN Custom s2 v2 6 5No Difficult TOIs MM SpencerOpen USACE LibraryMatch None MetalMapperU Custom s3 v3 1 Number of Non-TOIs Incorrectly Classified Difficult TOIs: #SR-194 (Small ISO QA seed) Difficult TOIs: #SR-194 (Small ISO QA seed) #SR-38 (37mm QA seed) Number of Non-TOIs Incorrectly Classified Difficult TOIs #SR-194 (Sm Difficult TOIs #SR-194 (Sm #SR-38 (37

42 EM61 MKII - Channel EM61 MKII - Channel EM61 MKII - Channel MM - Channel MM - Channel MM - Channel TEMTADS - Channel TEMTADS - Channel TEMTADS - Channel Dynamic Data Collected at Spencer Range Advanced EMI instruments produce higher resolution maps that result in more accurate target picking EM61-Mark2 Channel 2 MetalMapper Channel 5 TEMTADS 2x2 Channel Northing Local coordinates (m) I - Channel EM61 2 MKII - Channel EM61 2 MM MKII - Channel MM - Channel 5 TEMTADS MM - Channel - Channel 5 TEMTADS

43 Classification Processing Flow in UXOLab Estimate target parameters Data /Inversion QC Model Selection Request training data Create Ranked Anomaly List Determine a Stop Digging Point Single & multi-source inversions Location, orientation Target polarizabilities

44 Classification Processing Flow Estimate target parameters Data /Inversion QC Model Selection Request training data Create Ranked Anomaly List Determine a Stop Digging Point Identify data problems Identify poor fits to the data Determine if any anomalies cannot be analyzed

45 Classification Processing Flow Estimate target parameters Data /Inversion QC Model Selection Request training data Create Ranked Anomaly List Determine a Stop Digging Point Determine which of the models should be used in the classifier

46 Classification Processing Flow Estimate target parameters Data /Inversion QC Model Selection Request training data Create Ranked Anomaly List Determine a Stop Digging Point Semi-supervised approach Find clusters of similar objects Find outstanding objects

47 Classification Processing Flow Estimate target parameters Data /Inversion QC Model Selection Request training data Create Ranked Anomaly List Determine a Stop Digging Point Library matching or statistical classifier

48 Classification Processing Flow Pg 3 of 31 (shading = L123 misfit) Estimate target parameters T11 /.77 Small ISO 2 T119 /.78 Small ISO 2 T967 /.88 Small ISO 2 T78 /.189 Small ISO 2 T115 /.88 75mm T152 /.178 Small ISO 2 Data /Inversion QC T54 / T295 / T828 / T418 / T943 / T628 / Small ISO 2 Small ISO 2 75mm 75mm Small ISO 2 75mm Model Selection T272 / T183 / T122 / T127 / T1169 / T16 / Request training data Small ISO mm-d 92 Med. ISO-b 93 6mm mortar 94 37mm-b 95 Med. ISO-a 96 T1223 /.134 T241 /.2 T1133 /.162 T79 /.165 T49 /.228 T945 /.214 Create Ranked Anomaly List 75mm 97 Med. ISO-b 98 75mm 99 75mm 1 37mm-d 11 37mm-b 12 Determine a Stop Digging Point T232 /.27 37mm-d 13 T17 / mm-d 14 T27 /.187 Small ISO 2 15 T462 / mm 16 T991 /.26 Small ISO 2 17 T888 /.22 Small ISO 2 18 T133 /.221 T763 /.217 T78 /.219 T1191 /.236 T588 /.265 T359 /.26 37mm-b 37mm-d 75mm Small ISO 2 37mm-d 37mm-a Determine which targets are to be dug

Quality Management for Advanced Classification. David Wright Senior Munitions Response Geophysicist CH2M HILL

Quality Management for Advanced Classification David Wright Senior Munitions Response Geophysicist CH2M HILL Goals of Presentation Define Quality Management, Quality Assurance, and Quality Control in the