Slide 1

A LEXANDER K OLTUNOV GOES Early Fire Detection Project An Overview and Update akoltunov@ucdavis.edu Center for Spatial Technologies and Remote Sensing (CSTARS), University of California, Davis (Principal Investigator) TFRSAC-Spring 2013, NASA Ames, Monterrey Bay CA, 4/17/2013

Slide 2

GOES Early Fire Detection (GOES-EFD) System Detect new incidents consistently within first 1-2 hours, preferably before they are reported by conventional sources A low-cost and reliable capacity for systematic rapid detection and initial confirmation of new ignitions at regional level as a component of the USFS Active Fire Mapping program run by the FS RSAC complement existing related products (WF- ABBA, MODIS/VIIRS Active Fire ) 2

Slide 3

GOES-EFD system integration within USFS Active Fire Mapping program Provide standardized geospatial EFD products seasonal regional Integrate EFD into existing decision-making environment at dispatch centers and GACCs EFD to improve situational awareness, response planning, confirmation for fires reported via other sources 3

Slide 4

GOES-EFD Algorithm, Software, and Utility Programs GOES-EFD Algorithm, Software, and Utility Programs GOES-EFD effort: Structure and Participants GOES Images Landsat Images CAL FIRE Los Angeles County Fire Department (LACoFD) San-Bernardino National Forest Federal Interagency Communications Center (FICC) Supporting End-User Partners UC Davis FS RSAC CSUMB Developers New ignition locations Detection confidence Metadata per incident Performance stats over previous fire seasons New ignition locations Detection confidence Metadata per incident Performance stats over previous fire seasons Output Products Run @ RSAC and NASA Ames Feedback, Requirements, Evaluations, Integration R&D, Implementation, Validation, Integration USDA FS UC Davis DHS Sponsors RSAC: Format, Merge, Distribute to First Responders Developers Team: UCD: Alex Koltunov, Wei Li, Elaine Prins, Susan Ustin RSAC: Brad Quayle, Brian Schwind CSUMB: Vince Ambrosia 4

Slide 5

Where are we today? Alpha-version (GOES-EFD ver. 0.2) readyAlpha-version (GOES-EFD ver. 0.2) ready simulated real-time mode, simulated real-time mode, mainly Matlab implementation; mainly Matlab implementation; Case studies in CA indicate:Case studies in CA indicate: faster initial detection than the operational satellite algorithm faster initial detection than the operational satellite algorithm potential to provide the earliest alarm, time and again potential to provide the earliest alarm, time and again Algorithm optimizations and tests underway (as resources permit)Algorithm optimizations and tests underway (as resources permit) GOES-EFD ver.03 near completion (July 2013)GOES-EFD ver.03 near completion (July 2013) with ~35% fewer false alarms with ~35% fewer false alarms 5

Slide 6

2013-15: Major development-test iterations, implementation, and integration: complete the GOES-EFD -version2013-15: Major development-test iterations, implementation, and integration: complete the GOES-EFD -version 2015: Deployment of GOES-EFD- at USFS RSAC as a component of the FS Active Fire Mapping Program2015: Deployment of GOES-EFD- at USFS RSAC as a component of the FS Active Fire Mapping Program 2015-2016: Near-real-time delivery to participating users;2015-2016: Near-real-time delivery to participating users; initial training and evaluation by participating users user feedback and performance documentation follow-up optimizations 2017: Post-Deployment system maintenance and enhancement2017: Post-Deployment system maintenance and enhancement 2016-2017: Adaptation to GOES-R Advance Baseline Imager2016-2017: Adaptation to GOES-R Advance Baseline Imager GOES-EFD: Intended Schedule 6

Slide 7

Geostationary Satellites (GOES East/West): Frequent, Low-Cost Imaging of Vast Territories GOES Imager (NOAA): Viewing geometry fixed Viewing geometry fixed Visible + Thermal Infrared (TIR) images Visible + Thermal Infrared (TIR) images @ ~15-30 min step (5-min during Rapid Scan) @ ~15-30 min step (5-min during Rapid Scan) Effective TIR pixel size ~ 6 x 4 km over CA Effective TIR pixel size ~ 6 x 4 km over CA GOES-WestGOES-East 7

Slide 8

Active Fire Monitoring vs. Early Fire Detection *Wildfire Automated Biomass Burning Algorithm Active Fire MonitoringEarly Fire Detection Maximize % of detected burning pixelsMaximize % of detected new fire incidents (ignitions) Minimize % of false fire pixelsMinimize the number of false new incidents (alarms) Estimate flaming area, temperature, etc.Minimize time to initial detection of an incident Perform consistently all year-round globally (e.g. for comparative studies) Optimize for fire season and a chosen surveyed scene GOES WF-ABBA, MODIS / VIIRS Active Fire, AVHRR Primary Objectives are Related but Rather Different: GOES-EFD and So Are the Optimal Algorithms 8

Slide 9

GOES-EFD is a tool specifically optimized for the objectives of Early Detection 9

Slide 10

4 km background pixel fire pixel Burning Area Physical Basis for Infrared Fire Detection T=700 K T=600 K T=500 K Radiance Thermal IR 4 m 12 m Blackbody Spectrum T=320 K 4 km Plancks Law: Radiance ( ) = B ( , T ) wavelength temperature Primary regions used for detection: Primary regions used for detection: Short-wave TIR (3 - 5 m) Long-wave TIR(10 - 12 m) fire R>>R fire R 4 m >> R 11 m soil RR soil R 4 m ~ R 11 m 10

Slide 11

Annette fire: GOES 4- m Band vs. 11- m Band low high 4 m 11 m 11

Slide 12

GOES-EFD algorithmic principle: merge Temporal + Contextual information Database of Optimal Basis Images Training Stage Detection Stage Inspection Image Select Images Compute Parameters Combine Past Images compare Pre-processing Post-processing Reference Image Koltunov, Ben-Dor, & Ustin (2009) Int J of Rem Sens Koltunov & Ustin S.L. (2007) Rem Sens Environ Multitemporal background prediction by Dynamic Detection Model: 12

Slide 13

GOES-EFD ver. 0.2: Training and Preprocessing 13

Slide 14

GOES-EFD ver. 0.2: Detection Stage 14

Slide 15

Pixel-wise Unfiltered Fire Mask Fire Pixels Rejected Thermal Anomalies Cloud Snow 15

Slide 16

From Pixels to Events (potential incidents) GOES-EFD target objects are New Incidents (multi-pixel, multi-frame objects) past frame current frame high low Fire Confidence An old event: do not report/ report as re-detected A new event: report this event Event Tracker: Analyze the temporal evolution of spatially connected groups of fire pixels 16

Slide 17

Retrospective Assessment of Incident Detection Timeliness and Accuracy Official wildfire incident records + Landsat-based burn detection Koltunov A., Ustin, S. L., Prins, E (2012) On timeliness and accuracy of wildfire detection by the GOES WF-ABBA algorithm over California during the 2006 fire season, Remote Sensing of Environment, v.127: 194-209 Not a trivial problem: Official records are incomplete High resolution imagery is infrequent Very different from validating an Active Fire Product While any kind of error in the database is possible, not all kinds of errors are equally probable Derive useful and reliable performance measures, despite uncertainties and biases in truth data 17

Slide 18

Initial Experiment with fire season 2006 Central California Central California Detection Period: 40 days; 2852 images: Aug 3 Oct 1 at ~20-min time step on average. -- Substantial Cloud Cover Wildfire Incidents* Used: Large (>2 ha final size) wildfires; CA only CA only * Used wildfire incident databases from: California Department of Forestry and Fire Protection (CAL FIRE) Geospatial Multi-Agency Coordination (GeoMAC) group Include wildfire incident reports from: USFS, BLM, NPS, CAL FIRE, et al. Sample #1: 13 fires with known initial report HOUR Sample #2: 25 fires with known initial report DATE 18

Slide 19

Detection latency statistics: GOES-EFD ver.02 Detection latency relative to initial report from conventional sources perfect algorithm 19

Slide 20

Performance statistics: GOES-EFD ver.02 Performance statistics: GOES-EFD ver.02 GOES-EFD rapid GOES-EFD regular GOES-EFD @30min WFABBA @30min Detected incidents for 13 res with recorded report hour Detected in < 1 hour11/1310/13 7/13 Detected before reported4/13 3/132/13 Total latency reduction 216 Min 142 min 105 min 45 min for 25 res with recorded report date Detected in < 12 hours16/2515/25 11/25 False / non-wildfire incidents up to 784up to 7938 to 5339 to 55 GOES-EFD tends to detect fires earlier than WF-ABBA GOES-EFD can provide the earliest alarm 20

Slide 21

Recent and ongoing activities toward GOES-EFD ver.03 Targeting false positives due to:Targeting false positives due to: Motion of undetected cloud pixels Sub-optimal image registration across time Analyzing spatial gradients and spatial context of temporal changeAnalyzing spatial gradients and spatial context of temporal change Correlation between spatial connectivity, size, and magnitude of thermal anomaly pixels? Spatial filtering of the residuals after multitemporal anomaly detection of Long-wave TIR? Second-dominant image motion and local misalignment tests Preparing a large-area testPreparing a large-area test 21

Slide 22

Incidents with wrong combinations of Spatial, Spectral, and Temporal features eliminated ~ 21% of false new incidents (at error prob.< 0.1%)eliminated ~ 21% of false new incidents (at error prob.< 0.1%) Rejected False Alarm True (wildfire) Event BT-4m DDM Background Prediction Residual BT-4m DDM Background Prediction Residual 22

Slide 23

Laplacian Spatio-Temporal Filters of Long-wave Thermal Band eliminated ~ 17% of false new incidents (at error prob.< 0.1%)eliminated ~ 17% of false new incidents (at error prob.< 0.1%) 23

Slide 24

Improved Performance statistics: GOES-EFD ver.03 GOES-EFD regular WFABBA @30min Detected incidents for 13 res with recorded report hour Detected in < 1 hour10/137/13 Detected before reported4/132/13 Total latency reduction 142 min 45 min for 25 res with recorded report date Detected in < 12 hours15/2511/25 False / non-wildfire incidents 79 55 51 eliminated ~ 35% of false new incidents (at error prob.< 0.2%)eliminated ~ 35% of false new incidents (at error prob.< 0.2%) 24

Slide 25

GOES-R Advanced Baseline Imager (2016) Full disk coverage: every 15 minutesFull disk coverage: every 15 minutes Continental US coverage: every 5 minutes.Continental US coverage: every 5 minutes. Spatial resolution : 2 km in TIRSpatial resolution : 2 km in TIR A new channel at: 10.3 m.A new channel at: 10.3 m. Fewer saturated pixelsFewer saturated pixels When GOES-R is available: Mature, well tested GOES- EFD systemMature, well tested GOES- EFD system EFD-prepared, EFD-friendly user communityEFD-prepared, EFD-friendly user community acceptance by scientific communityacceptance by scientific community 25

Slide 26

Conclusions We are moving forward, as resources permitWe are moving forward, as resources permit Prospective users:Prospective users: when you say you need Earlier Detection you are helping speed up the development ! Stay tuned 26

Slide 27

2013-15: Major development-test iterations, implementation, and integration: complete the GOES-EFD -version2013-15: Major development-test iterations, implementation, and integration: complete the GOES-EFD -version 2015: Deployment of GOES-EFD- at USFS RSAC as a component of the FS Active Fire Mapping Program2015: Deployment of GOES-EFD- at USFS RSAC as a component of the FS Active Fire Mapping Program 2015-2016: Near-real-time delivery to participating users;2015-2016: Near-real-time delivery to participating users; initial training and evaluation by participating users user feedback and performance documentation follow-up optimizations 2017: Post-Deployment system maintenance and enhancement2017: Post-Deployment system maintenance and enhancement 2016-2017: Adaptation to GOES-R Advance Baseline Imager2016-2017: Adaptation to GOES-R Advance Baseline Imager GOES-EFD: Intended Schedule 27

Slide 28

We gratefully acknowledge Kevin Guerrero, Mark Rosenberg CAL FIRE Jim CrawfordLA County Fire Dept. Shawna Legarza, FICC Quinn Hart,UC Davis Mui Lay, George SheerUC Davis Dedicated Support by : and personally thank 19 UC Davis GOES Receiver infrastructure and data are provided by CIMIS (California Irrigation Management Information System) program CIMIS (California Irrigation Management Information System) program http://wwwcimis.water.ca.gov/cimis 28

Slide 29

29

Slide 30

30

Slide 31

31

Slide 32

New burn detection in Landsat pairs June 25 Zoom 1 July 27 dNBR dNBRA no true burns Zoom 2: dNBR dNBRA July 27 June 25 one true burn dNBR dNBRA RGB= bands (7,4,3) Path-Row: 43-34 zooms What if, there is a new burn near suspected false positive? What if, no new burns found near suspected false positive? 32

Slide 33

Example: Rejecting 1 confident fire pixel of the 4-pixel Annette fire (2006/8/3) BT-4m DDM Background Prediction Residual fr.671 t=2006.08.01.1645 cold anomaly hot anomaly dBT4-11 DDM Background Prediction Residual fr.671 t=2006.08.01.1645 33

Slide 34

Automatic Thermal Image Registration Automatic Thermal Image Registration Radiance ~4 m original registered 34

Slide 35

Radiance ~4 m originalregistered Automatic Thermal Image Registration-2 Automatic Thermal Image Registration-2 35

Slide 36

GOES Short-wave Thermal Band Video 36