Land cover classification from satellite imagery, and its applications in cellular network planning

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] To further increase the classification accuracies, radar image processing techniques were investigated to preprocess the Radarsat data before classification. Eight processing techniques were applied to Radarsat data at various windows from 3 x 3 to 25 x 25 pixels. For a single radar feature, the Entropy processing at window size 13 x 13 provides the best overall land cover classification accuracy improvement when fused with the Landsat imagery. For multiple radar features, a higher accuracy improvement was found when combining the features (i.e., 13 x 13 Entropy, 9 x 9 data range, 19 x 19 mean) with the Landsat data. This study introduces an approach of fusing Landsat data with multiple Radarsat features to the land cover classification practice. Post-classification techniques were studied for land cover classification maps. Several weighted kernels were developed for the majority filtering process. The method evaluates the correlation between neighbor pixels according to the distance and further improves the classification accuracy. For the St. Louis study area, the Gaussian weighted kernel increases the overall land classification accuracy compared to the Landsat images. Post-classification smoothing of the sensor fusion result (Landsat and radar feature combination) further increases the accuracy. A decadal change analysis was also conducted for the St. Louis, Missouri area using Landsat imagery and census population data. This study proposes a methodology to integrate remotely sensed and census data in urban change analysis. The assessment can provide information that can highlight priority urban growth regions. The analysis shows strong correlation between population and land cover changes, which indicates the potential of satellite imagery to generate the physical feature input for tele-traffic forecasting of a cellular network.