Stability Analysis and Case Study for Fire Smoke Removal of Sentilel-2 Satellite Images Using Fuzzy Classification Yi-Hsin Chung(a*), Yu-Wen Li(b) and Li-Yu Chang (c)
a) Engineering Assistant, Taiwan Space Agency
8F, 9 Prosperity 1st Road, Hsinchu Science Park, HsinChu City 300, Taiwan
*yihsinchung[at]tasa.org.tw
b) Engineering Assistant, Taiwan Space Agency
c) Engineer, Taiwan Space Agency
Abstract
Satellite imagery allows for the identification of wildfire locations and potential spread. However, smoke generated by fires often blocks the penetration of visible and near-infrared (VNIR) radiance, making it difficult to recognize surface features by satellite images. In contrast, shortwave infrared (SWIR) radiance can penetrate smoke particles, thereby improving the visibility of surface information and providing the possiblity of wildfire monitoring for the smoke affected areas. A previous study, ^Using Sentinel-2 SWIR to Remove Forest Fire Smoke^ (ACRS), primarily employed fuzzy classification-developed based on fuzzy set theory-offers a soft classfier approach to combine multiple linear functions in the SWIR to VNIR mapping and improves upon the results obtained using a single linear relationship. This method provides better correspondence between the SWIR and VNIR bands in smoke removal applications. However, in the fuzzy classification process, the number of classes must be defined prior to processing, and this parameter significantly affects both computational efficiency and the quality of results. Using fewer classes (e.g., 3 to 5) may have better efficiency in performance, but it often leads to poor outcomes due to an inability to resolve complex land cover types within the image. Conversely, using too many classes (e.g., more than 30) requires a high number of iterations and substantial processing time to achieve convergence, while yielding only limited improvements in classification accuracy. This study investigates how the number of fuzzy classification classes affects the accuracies across different wildfire case images. Based on the results, the study will offer practical recommendations for optimal class settings in future wildfire monitoring applications using Sentinel-2 satellite images.
Keywords: Sentinel-2- Forest Fire Smoke- SWIR- VNIR- fuzzy classification
