PI: Yang Cai, Carnegie Mellon University

Member: Yongxiang Hu, NASA Langley Research Center


The objective of this study is to build an embedded data mining processor for onboard physical property retrieving, which would significantly reduce the cost of physical inversion for combined observations.

Given physical properties P of the atmosphere, most multi-spectral observations O can be simulated
O = f(P), or so-called “forward models”. Inverse Physics is the process of retrieving physical properties from observations P = f-¹(O).



Using Generalized Non-Linear Regression, the wind speed W can be derived from microwave B:


where the W(i) and B(j,i) are wind speed and microwave brightness temperatures from forward model simulations and previous retrieval results. rho is a correlation factor between the brightness temperature of channel j and the wind speed, and s is measurement error of B(j);


A prototype of the physical inversion models is constructed on the NI PXI-7831R FPGA prototyping board. The FPGA Vertex II 1000 contains 11,520 logic cells, 720 Kbits Block RAM, and 40 embedded 18x18 multipliers. The following shows a basic design for GNR on FPGA. We also implemented Radial Basis Function to improve the performance.


Implementation of GNR with FPGA Vertex II 1000


Memory management on the FPGA chip


We found that FPGA out-performances Pentium at least two to three orders of magnitude in terms of speed. For example, for GNR model, the FPGA uses 39 us (with 10 MHz clock speed). Pentium uses between 1000 us and 2000 ns (with 1 GHz clock speed). The bottlenecks of the FPGA-based computing is the data I/O. How to get data in and out of the FPGA is on a critical path in terms of speed.