CHAPTER 6. WILDFIRE DETECTION 87The decision function D 2 (x, n) takes values between 1 and −1 dependingon the values of the Y (x, n), U(x, n) and V (x, n) channel values. The decisionfunction D 2 (x, n) is defined as:D 2 (x, n) ={1 −|U(x,n)−128|+|V (x,n)−128|,128if Y (x, n) > T I−1, otherwise(6.3)where Y (x, n), U(x, n) and V (x, n) are the luminance and chrominance values ofthe pixel at location x of the input image frame at time step n, respectively. Theluminance component Y takes real values in the range [0, 255] in an image andthe mean values of chrominance channels, U and V are increased to 128 so thatthey also take values between 0 and 255. The threshold T I is an experimentallydetermined value and taken as 100 on the luminance (Y) component in thiswork. The confidence level of D 2 (x, n) is −1 if Y (x, n) is below T I . The reasonthat we have the threshold T I is to eliminate dark regions which also have lowchrominance values. Since wildfire smoke regions are mostly colorless, having verylow chrominance values, the decision value approaches to 1 as the chrominancevalues U(x, n) and V (x, n) are around the mean value of 128 for pixels whoseluminance values are greater than T I . Confidence value drops down to −1 forpixels with high chrominance values.6.2.3 <strong>Detection</strong> of Rising RegionsWildfire smoke regions tend to rise up into the sky at the early stages of the fire.This characteristic behavior of smoke plumes is modeled with three-state HiddenMarkov Models (HMM) in this chapter. Temporal variation in row number of theupper-most pixel belonging to a slow moving region is used as a one dimensional(1-D) feature signal, F = f(n), and fed to the Markov models shown in Fig.6.2.One of the models (λ 1 ) corresponds to genuine wildfire smoke regions and theother one (λ 2 ) corresponds to regions with clouds and cloud shadows. Transitionprobabilities of these models are estimated off-line from actual wildfires and testfires, and clouds. The state S1 is attained, if the row value of the upper-most pixelin the current image frame is smaller than that of the previous frame (rise-up).If the row value of the upper-most pixel in the current image frame is larger than
CHAPTER 6. WILDFIRE DETECTION 88that of the previous frame, then S2 is attained and this means that the regionmoves-down. No change in the row value corresponds to S3.Figure 6.2: Markov model λ 1 corresponding to wildfire smoke (left) and theMarkov model λ 2 of clouds (right). Transition probabilities a ij and b ij are estimatedoff-line.A slow moving region is classified as a rising region when the probability ofobtaining the observed feature signal F = f(n) given the probability model λ 1is greater than the probability of obtaining the observed feature signal F = f(n)given the probability model λ 2 , i.e., when the upper-most pixel belonging to aslow moving region tends to exhibit a rising characteristic:p 1 = P (F |λ 1 ) > p 2 = P (F |λ 2 ) (6.4)where F is the observed feature signal, λ 1 and λ 2 represent the Markov modelsfor wildfire smoke and clouds, respectively.As the probability p 1 (p 2 ) gets a larger value than p 2 (p 1 ), the confidence levelof this sub-algorithm increases (decreases).Therefore, the zero-mean decisionfunction D 3 (x, n) is determined by the normalized difference of these probabilities:D 3 (x, n) = p 1 − p 2p 1 + p 2(6.5)