User Tools

Site Tools


wiki:contest_data_fusion

Stack bands in feature fusion approach

LiDAR=band 0
NDVI=band 1
LINEAR FEATURES=band 2 and 3
CASI1-144=band 4 … band 147
    pkcrop -i 2013_IEEE_GRSS_DF_Contest_LiDAR.tif -i 2013_IEEE_GRSS_DF_Contest_NDVI.tif -i 2013_IEEE_GRSS_DF_Contest_LINEAR.tif -i 2013_IEEE_GRSS_DF_Contest_CASI.tif -o 2013_IEEE_GRSS_DF_Contest_FEATURE_FUSION.tif

Feature fusion

Create training vector with stacked features

    pkextract -i 2013_IEEE_GRSS_DF_Contest_FEATURE_FUSION.tif -s 2013_IEEE_GRSS_DF_Contest_Samples_TR_26915.shp -o training_feature_fusion.shp

Optimize SVM parameters for these features

    pkopt_svm -t training_feature_fusion.shp -cc 0.1 -cc 10000 -g 0.001 -g 10 -step 10
–ccost 7657.41 –gamma 0.186167

Redo the feature selection based on this new feature set (we can stop already after 10 features)

    pkfs_svm -t training_feature_fusion.shp -cc 100 -g 1 -n 10 -v 1 -cv 2
-b 140 -b 2 -b 48 -b 0 -b 3 -b 23

Image classification

Based on the optimal feature set, we can create a new land cover map

    pkclassify_svm -i 2013_IEEE_GRSS_DF_Contest_FEATURE_FUSION.tif -t training_feature_fusion.shp -o testmap_feature_fusion.tif -ct ct.txt --ccost 100 --gamma 1 -b 140 -b 2 -b 48 -b 0 -b 3 -b 23
wiki/contest_data_fusion.txt · Last modified: 2017/12/05 22:53 (external edit)