Evaluation of Deconvolution Methods for PRISM Images
(1) German Aerospace Center, Muenchnerstr. 20, D-82234 Wessling, Germany
The Applied Remote Sensing Cluster (CAF) at the German Aerospace Center (DLR) has vast experience in processing of high-resolution panchromatic and multispectral airborne and space borne images. Within the scope of a project by the European Space Agency (ESA), DLR is responsible for the establishment of processors for ALOS/AVNIR-2 and ALOS/PRISM data. This processing chain not only includes radiometric and geometric correction for ALOS/AVNIR-2 and ALOS/PRISM but also atmospheric correction for ALOS/AVNIR-2. This paper gives a short introduction into the processing chain as a whole and a more in-depth look into the deconvolution strategies taken into consideration for ALOS/PRISM images. The geometric and atmospheric corrections are precisely analysed in other abstracts, which are also submitted to this symposium.
Within the first stage of the processing chain, the processor imports the JAXA Level 1a product and performs systematic as well as radiometric corrections followed by an optional image deconvolution for ALOS/PRISM. The next processing stage - namely the geometric correction - creates ortho-images with the help of a digital elevation model. Finally, based on the radiometric and optionally also geometric corrected ALOS/AVNIR-2 data, a conclusive atmospheric correction is performed for the final image product generation.
Remote sensing images are subject to distortions, such as motion blur, noise, atmospheric turbulences, etc. Image restoration techniques such as deconvolution and denoising can significantly improve the image quality and thus greatly help in further visual interpretation and analysis/processing of the ALOS/PRISM images. Deconvolution is the process of reverting the distortion of a signal (in this case an image) to compensate for an undesired convolution. To find out an optimal deconvolution approach for ALOS/PRISM images, several known methods such as Wiener inverse filtering, FFT filtering and statistical methods were analysed and compared to each other for selected sample data. Most of these deconvolution methods require an additional denoising step to work accurately, thus avoiding amplification of the noise present in an image. Therefore more advanced methods combining the deconvolution and denoising steps, e.g. with the aid of complex wavelets packets, were also taken into consideration. The pros and cons of these evaluated methods, which led to the final selection of a deconvolution algorithm for ALOS/PRISM images, have been described in more details in this paper.