Threshold Optimization of Contextual Fire Detection Algorithm using Fuzzy Clustering

A contextual algorithm that is widely used for identifying forest fire pixels uses a threshold derived from statistical examinations of temperature distributions of background pixels. In general, about 3σ (standard deviation) above the mean of these background temperatures is used for the threshold. In case where land cover types are multifaceted in the background pixels, whose distributions of surface temperatures are clearly diverse, the σ value becomes overestimated resulting in increased threshold. This is a typical edge problem encountered in the current contextual algorithm which explains why relatively small fires are often omitted. Therefore, in this paper, a new algorithm to optimize the threshold is proposed to overcome the above problem. In this algorithm, statistical inferences of various land covers in the background pixels as well as its center pixel are used. For this, a fuzzy clustering is applied to the background pixels and corresponding statistical distribution of temperatures of each cluster is examined to derive a series of thresholds of the clusters. Then the characteristics of the center pixel are analyzed based on memberships of different clusters. Lastly, an optimum threshold is calculated throughout some arithmetic operations to the derived thresholds and memberships information. In this study, the proposed algorithm was applied to the MODIS imageries that were attained during several fire seasons. The results were compared with those from the current algorithm developed by NASA MODIS science team. The proposed algorithm showed relatively high accuracy by detecting several pixels that the current algorithm failed to sense. For a ground truth, forest fire data provided by the Korea Forest Service were used. Key-Words: Forest fire detection, Contextual algorithms, Threshold optimization, Fuzzy clustering, MODIS