Abstract:

In hyperspectral images,there exists a huge amount of data with hundreds of spectral bands,which mayresult in heavy computational burden of the pixel purity index (PPI).In order to solve this problem,an architec-ture for FPGA (Field Programmable Gate Array) implementation is proposed.In this architecture,the matrix com-putation based on the parallel processing of projection vectors is adopted to reduce the interface data accesses,aninner product circuit with parallel bands is proposed to improve the projection computation parallelism,and simpli-fied projection vectors are employed to avoid the floating- point multiplication in hardware implementation.The re-sults of the experiments on the AVIＲIS Cuprite hyperspectral images with 350 ×350 resolution and 224 bands showthat the proposed architecture outperforms the ENVI software and some other existing architectures in terms of end-member extraction accuracy,and that the PPI computation can be completed within 10 seconds in the correspondingFPGA implementation,which meets the requirement of real- time processing in spaceborne/airborne fields.

Key words: hyperspectral images, endmember extraction, pixel purity index, random projection, field programma-ble gate arrays