Onboard FPGA AI/ML processing on Landsat-8 satellite images: a case study of wildfires detection

Using satellite-acquired data often requires transferring information to control stations for processing, which degrades the quality of service. Rapid information usage is crucial in scenarios like early warning systems for critical structures or environments. Using artificial intelligence and machine learning (AI/ML) onboard satellites can enhance autonomy and control. This paper aims to identify the key requirements for onboard FPGA AI/ML processing without significantly increasing payload complexity and to define a reference architecture based on initial study scenarios. A fire detection data set, consisting of images from Landsat-8 satellites, includes wildfire and control images from Uruguay and Brazil. Four network architectures were evaluated for wildfire detection, with Resnet-18 being the best tradeoff. All networks effectively generalized the task. The proposed system is capable of detecting fires in an area of 51,984 km2 in 176 s.