Accuracy assessment of remote sensing classification techniques

Remote sensing data are expressed as digital numbers per pixel which means that they are not object oriented, thus leading to some limitations for the automatic interpretation of the satellite images. The main objective of a classification process is to solve the problem of feature selection in order to obtain a thematic map from multi-spectral image. This research analyses the classified images to assess the accuracy of various classification processes, present their mathematical models and compare their performance. Many experiments have been performed using Landsat TM images with multi-bands observations. The best classifier deduced from the previous analysis was subsequently applied to Spot XS data. The results are verified by ground truth data and the accuracy of each classification technique is evaluated according to statistical accuracy's assessment rules.