Exploring Various Applicable Techniques to Detect Smoke on the Satellite Images

Every year, the wildfires ravage broad areas of natural forest and nearby regions, causing substantial financial and life losses and deteriorating the air quality. More air hazards are emitted to the atmosphere reaching as high as the stratosphere propagating through the air currents. With the aggravation of climate change, wildfires of either human or natural cause could become more ferocious and devastating. A feasible solution is to detect the wildfire and respond early before the fire spread becomes irreversible. Satellite imagery serves as a cost-effective means to update near-real-time holistic landscape views of land and sea over extended periods. Such an advantage makes early fire detection and warning even in remote areas possible. The rendered images provided by the satellites' various instruments incorporate various channels to provide real and artificial colors to reveal landscape details imperceptible to the naked eyes. This imagery dataset discussed in this paper derives from NASA's Aqua and Terra satellites and make available at the NASA-IMPACT data share repository [1]. The dataset totals 704 images of cropped frames and their labeled images taken during both satellites' extensive flyby observations. The images also contain spatial-temporal information serving as relevant metadata for analysis. This paper provides a survey of the recent advances in neural network-based object detection techniques followed by machine learning and deep learning-based methods to detect and localize smoke. A comprehensive elaboration of the datasets follows the method overview.