Simulation runs deep in the group’s foundations. When ground truth is scarce, we build it: stochastic mathematical InSAR simulators in 2D and 3D time series, a domain-tailored GAN with a U-Net discriminator that generates realistic interferograms, and generative data for water-body segmentation, where generative augmentation delivers 10–28% improvement versus 2–8% for traditional augmentation.
When compute is tight, we shrink it: edge AI for on-device inference near the sensor, drones today and potentially on-board satellite processing tomorrow, powered by frequency-domain structural pruning that compresses networks by up to 1286.55× (UNet/Carvana) while maintaining accuracy.
Together, faithful synthetic worlds and efficient models point toward digital-twin representations of monitored environments: living models that stay in sync with what the sensors see, in collaboration with MatrixLabs, whose programmable 3D world and multi-chain data infrastructure work connects simulation to deployed digital-twin platforms.