Evaluation of supervised classification by class and classification characteristics

摘要

As acquisition technology progresses, remote sensing data contains an ever increasing amount of information: optical and radar images, low, high and very high-resolution, multitemporal hyperspectral images, derived images, and physical or ancillary data (databases, Digital Elevation Model (D.E.M), Geographical Information System (G.I.S.)). Future projects in remote sensing will give high repeatability of acquisition like Venμs (CNES) which may provide data every 2 days with a resolution of 5.3 meters on 12 bands (420nm-900nm) and Sentinel-2 (ESA) 13 bands, 10-60m resolution and 5 days. With such data, supervised classification gives excellent results in term of accuracy indices (like Overall Accuracy, Kappa coefficient). In this paper, we present advantages and disadvantages of existing indices and propose a new index to evaluate supervised classification using all the information available from the confusion matrix. In addition to accuracy, a new feature is introduced in this index: fidelity. For example, a class could have a high accuracy (low omission error) but could be over-represented with other classes (high commission error). The new index reflects accuracy and correct representation of classes (fidelity) using commission and omission errors. Environment applications are in land cover and land use and the goal is to have the best classification for all classes, whether the biggest (corn, trees) or the lightest (rivers, hedges). The tests are performed on Formosat-2 images (every 2 days, 8 meters resolution on 4 bands) in the area of Toulouse (France). Tests used to validate the new index by demonstrating benefits of its use through various thematical studies.