Part 6/12:
He shared insights into optimization techniques: using Numba (a Python-to-machine code compiler) combined with parallel processing, engineers made significant speedups—up to 25 times faster inference times compared to traditional CPU methods. This enabled the deployment of applications like building detectors and road network extractors directly on satellites.
For example, a scenario involved analyzing a 5 km by 32 km scene, initially comprising around 950MB of raw data. After onboard processing, only a few megabytes containing critical features like roads or flood zones were transmitted—more than 90% reduction in data volume. Such efficiencies are instrumental in reducing transmission costs and latency.