RPC-Based Orthorectification for Satellite Images Using FPGA

• Published in: Sensors (Basel, Switzerland) Switzerland), 2018-08, Vol.18 (8), p.2511
• Main Authors: Zhang, Rongting, Zhou, Guoqing, Zhang, Guangyun, Zhou, Xiang, Huang, Jingjin
• Format: Article
• Language: English
• Subjects: Coordinate transformations; Field programmable gate arrays; field-programmable gate array (FPGA); Fixed point arithmetic; Geometric rectification (imagery); Interpolation; Matlab; orthorectification; Personal computers; Pixels; Processing speed; rational polynomial coefficient (RPC); Remote procedure calls; Satellite imagery
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s18082511

Conventional rational polynomial coefficients (RPC)-based orthorectification methods are unable to satisfy the demands of timely responses to terrorist attacks and disaster rescue. To accelerate the orthorectification processing speed, we propose an on-board orthorectification method, i.e., a field-programmable gate array (FPGA)-based fixed-point (FP)-RPC orthorectification method. The proposed RPC algorithm is first modified using fixed-point arithmetic. Then, the FP-RPC algorithm is implemented using an FPGA chip. The proposed method is divided into three main modules: a reading parameters module, a coordinate transformation module, and an interpolation module. Two datasets are applied to validate the processing speed and accuracy that are achievable. Compared to the RPC method implemented using Matlab on a personal computer, the throughputs from the proposed method and the Matlab-based RPC method are 675.67 Mpixels/s and 61,070.24 pixels/s, respectively. This means that the proposed method is approximately 11,000 times faster than the Matlab-based RPC method to process the same satellite images. Moreover, the root-mean-square errors (RMSEs) of the row coordinate (Δ ), column coordinate (Δ ), and the distance Δ are 0.35 pixels, 0.30 pixels, and 0.46 pixels, respectively, for the first study area; and, for the second study area, they are 0.27 pixels, 0.36 pixels, and 0.44 pixels, respectively, which satisfies the correction accuracy requirements in practice.