Cavazzini, Armando 
  
Nine land cover classes were distinguished: forest, grass (crops), bare soil, water, asphalt (concrete), tiles, rocks, roofs. 
Shadows were listed as unclassified. Because of the high degree of intermixing between residential, commercial and 
industrial uses, the number of anthropic landcover classes is preminent. 
3. RESULTS 
Table 1 summarizes the error matrix and figure 2 exhibits the classification image. 
The classification was found to be of high quality, with 96.8 per cent of cases correctly allocated. It is also evident that 
much of the misclassification arose from the commission of cases from the asphalt-concrete and the roofs, being the roofs 
mainly composed by the same endmembers. It is interesting that the overall accuracy index and the k coefficient (Congalton 
and Mead, 1983) exhibit a large agreement even if the k computation incorporates the off-diagonal elements as a product of 
the row and column marginals. The high accuracy of the per pixel classification is indicative of land cover classification 
performance using MIVIS data. 
Figure 1. A portion of the area that has been classified (Sulmona, Italy) 
The classification image has been georeferenced using an image to map (georeferenced orthophoto) registration with a 
second order polinomial warping algorithm. The root mean square (RMS) error for this trasformation was 0.3 pixels (0.9 
meters). As a map, a precise georeferenced orthophoto has been used. 
For the purposes of this paper, spectral imagery, orthopotos and classified images have been incorporated in an ArcView 
GIS? project as raster layers on which vector information is displayed, this making data more significant. The methodology 
for the GIS implementation of MIVIS images involves four stages: 1) data extraction (MIVIS data scanning) 2) information 
extraction (classification algorithms and other spectrals computations) 3) data integration (use of external ground data) 
4) data analysis (GIS integration, land cover and land suitability analysis, vegetation stress and water turbidity 
quantification, edaphic parameters mapping...). 
The forest thematic layer has been isolated and a NDVI (Normalized Difference Vegetation Index) has been calculated. The 
association of the vegetation index with a thermal band has been useful to check and geographically locate stressed 
vegetation. The nearness to heavy traffic and industries was found to be the main reason for vegetation stress. 
Streams and canals were isolated as a thematic layer and surface water turbidity was quantified by a blue and 
  
238 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 
pam