Energy-efficient on-board AI for early fire-smoke detection

This research aims to provide a solution for energy-efficient AI-based on-board processing of hyperspectral imagery supporting automated early detection of fire smoke. We propose using modified and resampled MODIS imagery data that emulates the swath as well as spectral, spatial, and radiometric resolution of HyperScout-2 channel 1 hyperspectral imagery. In doing so, we intend to provide a solution that meets on-board processing limitations and up/downlink data transfer restrictions of the Kanyini – HyperScout-2/ with Intel’s Myriad X VPU chip.

Based on a semi-automatically created fire smoke training dataset, our proposed AI processing approach is performed at two levels: a) on-board feature and band selection, and b) ground AI neural network tasks – in order to optimize on-board processing and downlink data transfer. Expected outputs include on-board and ground AI algorithms for fire smoke detection, applicable for various hyperspectral imagery datasets.


Project Leader:
Dr Stefan Peters, University of South Australia