Relation between spatial resolution and land cover classification accuracy of remotely sensed image

using four images and spatial degradation algorithm

Abstract

Various kinds of high resolution remote sensors are available, such as Landsat TM (spatial resolution 30 mx 30 m), SPOT HRV (20m×20m) and ADEOS AVNIR (16m×16m) launched in 1996. It is significant to quantify the accuracy of interpretation or classification of high resolution remotely sensed images. This paper assesses land cover classification accuracy quantitatively for four remotely sensed images acquired almost simultaneously with various sensors with different spatial resolutions: Landsat TM, SPOT HRV, JERS OPS (18m×24m), Airborne multispectral scanner (AMSS: 6.25m×6.25m). Comparison of the four classified images shows that the AMSS image with the finest resolution has not given remarkable improvement in terms of classification accuracy. Moreover, the authors compare the false remotely sensed images produced by three degradation algorithms. The first algorithm is using arithmetic mean, and the second one is using cubic convolution interpolation, and the third one is using MTF. The algorithm using MTF is recommend as the image degradation algorithm in a strict sense. The false images using MTF have 10-m to 30-m resolution resampled from the 6.25-m AMSS image. Consequently, the 6.25-m AMSS image and the degraded images with finer resolution than 12 m have not given better classification. These consequences that the images with the finest resolution has not given the finest classification accuracy is attributed to the limited number of categories considered. The authors also produce the 16-m false AVNIR image by the degradation algorithm using MTF, and the classification accuracy of ADEOS AVNIR images is predicted.