Band Selection & Algorithm Development for Remote Sensing of Wildland Fires Andy Fordham Chester F. Carlson Center for Imaging Science Rochester Institute of Technology Rochester, New York 14623-5604 ajf8207@cis cis.rit.edu
Outline Overall goal was to select correct bands and design simple algorithms for a fire sensor where the fire occupied only a small fraction of a pixel If correct bands are selected, very simple algorithms will prevail Band Selection based on modeling SW IR MW/LW IR Potassium atomic emission lines Thermal Results MW/LW IR SW IR Potassium line emission analysis Effects of smoke obscuration Conclusions
A wildland fire appears predominantly as a ~1000K blackbody Normalized Radiance & Transmittance 0.00E+00 4.00E+03 8.00E+03 1.20E+04 Wavelength (um) Earth Fire Sun Transmittance
Band selection was performed by Selection Metric DC 2 -DC 1 (s 2 ) 2 +(s 2 ) 2 = Q Select Bands that Maximize Q MW-LW is best Minimal clutter in emissive spectra Large difference between fire & background SWIR is very good SWIR is very good More clutter between 1-2.5mm Easy to detect specular reflectors with 1 mm band
We devised a simple algorithm based on band ratios rather than thresholds or other complex calculations Use Color instead of Threshold do not consider individual band radiance values instead consider relationship between radiances can express this relationship with a ratio Band 1 Band 2 100% water 2% metal 1% steel 1% Fire (1000K) Background (grass) 1000 1500 2000 2500 Wavelength (nm)
800K 900K 1000K 1100K Threshold selection in SWIR Background Water Metal Aluminum Steel SWIR Band 2 (2.1 um) 700K 600K Fires 10% 600K Fire 1% 700K Fire Threshold Specular Metal SWIR Band 1 (1.25 um) Specular Water False Alarms 600K Fire 700K Fire 800K Fire 900K Fire 1000K Fire 1100K Fire
We modeled fire emissions using MODTRAN atmospheric transmittances and Planckian fire emission Radiance 4 3.5 3 2.5 2 1.5 1 0.5 0 MWIR Apparent Temp Comparisons for False Alarms, Fire and Background 3000 3500 4000 4500 5000 5500 Wavelength (nm) MWIR--Very Small Fires Background (Gras 0.1% Fire (1000K Water (Specular) (Warm Ground)
MWIR 1000.00 100.00 10.00 1.00 0.10 MWIR band ratios vs.. %fire in a pixel Fires 1% 600K Fire Threshold 1100K 1000K 900K 800K 700K 600K Water Warm Ground 1.00 10.00 100.00 1000.00 LWIR Steel/Metal False Alarms Background Water (Specular) Metal/Steel (Specula Warm Ground (320K 600K Fire 700K Fire 800K Fire 900K Fire 1000K Fire 1100K Fire
100.00 10.00 MWIR 1.00 LWIR 4-Band Algorithm Fires SWIR Threshold Background Water Metal Warm Ground 600K 700K 800K 900K 1000K 1100K 1400K MW-LW Threshold 0.10 4-band Threshold warm ground metal water 0.01 0.01 0.1 1 1.2mm 10 100 1000 2.1mm
Basics of thermally-excited atomic emission from potassium Extremely narrow lines at 766.5 and 769.9 nm Very high spectral power density compared to the thermal emission from fire or other hot backgrounds Relatively low excitation energy - 4.34 ev Relatively high concentration in plant tissue 0.4 3.4 % by dry weight >10-20% of potassium may be excited at fire temperatures The total emission will be determined by a complex combination of factors in the local environment of a flame
The potassium line emission (fortuitously) passes through the atmosphere 1 0.9 Attenuated K lines Atmospheric transmission (right axis) 0.8 0.7 Relative Intensity 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.6 0.5 0.4 0.3 0.2 Relative Transmission 0.1 0.1 0 0 766.0 767.0 768.0 769.0 770.0 Wavelength, nanometers
Radiance (Gain*W/cm^2nmsr) 25000 20000 15000 10000 5000 0 K-emission Example 600 700 800 900 1000 1100 1200 Wavelength (nm) Actual Aviris Pixel Theoretical Pixel (20%-1215K) Planckian Curve-fitting allows measurement of fire temperature/size We can evaluate K-emission detection using Planckian data.
Initial results from potassium line emission are promising K detection works 50-70% 100m GSD is feasible Requires --High (1-2 nm) ) spectral resolution --High SNR Can only detect high temp fires (>1000 K) Better Data needed for validation Accurate ground-truth is required currently comparing K-data & thermal (e.g. flame detection vice heat detection) errors result since hot ground is not necessarily fire Potassium line emission is more susceptible to smoke than the infrared bands because of particle size distribution of smoke Fires
The effects of heavy smoke must be considered as fire follows smoke follows fire Heavy Smoke obscures all frequencies Effects of Moderate Smoke Vary MWIR/LWIR has best propagation SWIR is adequate NIR (e.g. Potassium) is attenuated strongly Light Smoke is inconsequential in all bands Detailed smoke model is still being constructed
1 Parameterized aerosol transmittance for NIR through LWIR 0.9 0.8 0.7 Transmittance 0.6 0.5 0.4 0.3 0.2 0.1 Normal Aerosols Moderately Heavy Heavier Very Heavy Concentration 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Wavelength (um)
Conclusions Color has good potential for detection Need average 2-4 K NEDT Need relatively high saturation temps Avoids contextual algorithm MWIR/LWIR is best band strategy SWIR is also very feasible (may be better if simpler or cheaper) Potassium detection is inconclusive Need better test data for Potassium Current detection rates are 50-70% These rates compare flame information to heat information (Since heat flame. Results may be better)