Effect of image compression and resampling methods on accuracy of land-cover classification

Abstract

High-resolution digital images that are used in remote sensing technologies tend to be of large sizes, especially with satellite imageries that provide better than 5 m spatial resolution such as (QuickBird, IKONOS, etc). These images consume larger storage space, larger transmission bandwidth, and longer transmission times. In order to minimize the storage space and transmission time, the archived images should be compressed before storage and transmission. Most publications generally investigate the quality of reconstructed (compressed) image, only few studies address the influence of the image compression on processing results, i.e. image classification. This paper investigates the effects of the most commonly used compression wavelet-based formats (JPEG2000 and MrSID) on the classification results of the two high-resolution QuickBird and IKONOS images. It is important to mention that the two high-resolution images cover two completely different study areas; therefore the study methodology was divided accordingly into two main cases. Both cases to the end that the impact of previously mentioned two wavelet-based formats was assessed on the classification results. Furthermore, only for the first data set, another effect was investigated. This other effect is the influence of the traditional resampling methods on the classification results. It is important to mention that typically, in the geo-referencing or geometric correction operations, resampling process is the second step following rectification. Moreover, the influence of resampling methods over compressed images was investigated. This paper shows that classification accuracy which derived from MrSID compressed images is better than that accuracy of JPEG2000 compressed images, especially at the high compression ratios of the two data sets. The first data set illustrate that the NN resampling method is the best method, in which the classification accuracy that was obtained was significantly closer to the classification accuracy of original image than any other traditional methods. Moreover, we concluded that CC method is the best in case of JPEG2000 compressed images while BI method is the best in case of MrSID, especially at high compression ratios. Meanwhile, the second data sets revealed that the ANNs classification method introduce classification accuracy higher than MLC method in case of MrSID compressed image, at higher compression ratios, while it was the higher in case of JPEG2000 at all compression ratios.