Next Generation Satellite Lidar Sensing, Neural Compression, and Accelerated Neural Inference

This project improves the ability of satellites to map Earth’s surface, vegetation, and atmosphere using light detection and ranging, or lidar, a technique that measures distance by timing reflected laser light. Current satellite lidar systems face three related limits: they receive only small numbers of reflected light particles, called photons; they have limited capacity to send data to the ground; and they require substantial computing time to convert measurements into useful maps. These limitations restrict the speed, coverage, and detail of Earth observations. The research addresses all three barriers as a connected system, allowing satellites to produce more detailed and timely three-dimensional information while using less communication bandwidth and computing power. The resulting capabilities support environmental monitoring, disaster response, water-resource assessment, agriculture, forestry, infrastructure planning, risk analysis, public safety, and national security. The project also advances education by training graduate and undergraduate researchers at the intersection of artificial intelligence, communication systems, and satellite-based Earth observation. Open-source software, datasets, and evaluation benchmarks broaden access, support reproducibility, and enable use by academic, government, and industry communities. Workshops and educational activities broaden participation and help prepare a workforce able to develop reliable and efficient sensing systems for scientific and societal needs. The scientific problem is how to preserve and recover the greatest amount of useful information from satellite lidar measurements containing very few detected photons when sensing, data transmission, and reconstruction are all constrained. The goal is to create a learning framework that incorporates physical models, spans the full processing chain, and jointly designs how measurements are acquired, compressed, protected against transmission errors, and converted into three-dimensional products. The scope includes three connected components. First, the research studies coded patterns of emitted light and sampling below conventional rates while explicitly modeling the random arrival of photons. Second, it develops neural joint source-channel coding, a learned method that combines data compression and error protection, to provide reliable high-resolution reconstruction under limited bandwidth and changing satellite communication conditions. Third, it accelerates computation through diffusion-based reconstruction that refines an estimate in only a small number of steps, and through direct estimation of selected percentiles of detected heights needed to create digital terrain models, vegetation height percentiles, and canopy height maps. The methods optimize the balance between reconstruction accuracy and data rate, measure how much useful information is retained across the processing chain, and train neural models to form compact internal representations using randomized simulations of the sensing and communication processes. Evaluation spans different photon levels, noise conditions, and communication-channel variations and measures reconstruction accuracy, robustness, data volume, computing cost, and retained information. The research connects theoretical information limits with practical sensing, communication, and computing constraints for efficient and accurate three-dimensional Earth observation. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.