Learning-Based Smoke Detection for Unmanned Aerial Vehicles Applied to Forest Fire Surveillance

Abstract

Forests are potentially and seriously threatened by fires which have caused huge damages and losses of life and properties every year. In general, it is easier to detect smoke than fire in its early stage. Developing an effective and safe smoke detection method is thereby critical for early forest fire fighting and preventing the fire developing into uncontrollable. This paper presents a learning-based fuzzy smoke detection approach intended to achieve an effective and early forest fire detection, while greatly reduce the negative impacts from clouds in the sky, illumination variations, and changes of forest features. First, a fuzzy-logic based smoke detection rule is designed for detecting and segmenting smoke regions in the visual images captured by the camera onboard an unmanned aerial vehicle (UAV). The differences of each two components of red, green, and blue (RGB) model and intensity in hue, saturation, and intensity (HSI) model of images are chosen as inputs of a fuzzy logic rule, while the smoke likelihood is selected as its output. Then, an extended Kalman filter (EKF) is further employed for reshaping the inputs and output of the fuzzy smoke detection rule on-line. It is expected to provide the smoke detection method with additional regulating flexibility adapting to variations of environmental conditions and reliable automatic detection performance. Next, the morphological operation is also adopted to remove imperfections induced by noises and textures distorted nonconvex/concave segments. Finally, extensive studies on several sets of images containing smoke under distinct environmental conditions are conducted to validate the proposed methodology.