A Machine Learning Approach to Waterbody Segmentation in Thermal Infrared Imagery in Support of Tactical Wildfire Mapping

Wildfire research is working toward near real-time tactical wildfire mapping through the application of computer vision techniques to airborne thermal infrared (IR) imagery. One issue hindering automation is the potential for waterbodies to be marked as areas of combustion due to their relative warmth in nighttime thermal imagery. Segmentation and masking of waterbodies could help resolve this issue, but the reliance on data captured exclusively in the thermal IR and the presence of real areas of combustion in some of the images introduces unique challenges. This study explores the use of the random forest (RF) classifier for the segmentation of waterbodies in thermal IR images containing a heterogenous wildfire. Features for classification are generated through the application of contextual and textural filters, as well as normalization techniques. The classifier's outputs are compared against static GIS-based data on waterbody extent as well as the outputs of two unsupervised segmentation techniques, based on entropy and variance, respectively. Our results show that the RF classifier achieves very high balanced accuracy (>98.6%) for thermal imagery with and without wildfire pixels, with an overall F1 score of 0.98. The RF method surpassed the accuracy of all others tested, even with heterogenous training sets as small as 20 images. In addition to assisting automation of wildfire mapping, the efficiency and accuracy of this approach to segmentation can facilitate the creation of larger training data sets, which are necessary for invoking more complex deep learning approaches.