Integrating contextual information with per-pixel classification for improved land cover classification

A hybrid segmentation procedure to integrate contextual information with pe r-pixel classification in a metropolitan area land cover classification pro ject is described and evaluated. If is presented as a flexible tool within a commercially available image processing environment, allowing components to be adapted or replaced according to the users needs, the image type, and the availability of state-of-the-art algorithms. In the case of the Twin C ities metropolitan area of Minnesota, die combination of die Shen and Casta n edge detection operator with iterative centroid linkage region growing/me rging based on Student's t-tests proved optimal when compared to other more common contextual approaches, such as majority filtering and the Extractio n and Classification of Homogenous geneous Objects classifier. Postclassifi cation sorting Further improved the results by reducing residual confusion between urban and bare soil categories. Overall accuracy of the optimal cla ssification technique was 91.4% for a level II classification (10 classes) with a K-e of 90.5%. The incorporation of contextual information in the cla ssification. process improved accuracy by 5.8% and K-e by 6.5%. As expected , classification accuracy for a simplified level I classification (five cla sses) was higher with 95.4% and 94.3% for K-e. A second important advantage of the technique is the reduced occurrence of smaller mapping units, resul ting in a more attractive classification map compared to traditional per-pi xel maximum likelihood classification results. (C) Elsevier Science Inc., 2 000.