Normalized Mutual Information-Based Ranking of Spatio-Temporal Localization Maps
Résumé
Satellite image time series (SITS) can be described in an unsupervised way by means of spatio-temporal localization maps. These maps are extracted using data mining techniques that spatially and temporally locate pixel evolutions affecting a minimum number of pixels with sufficiently high connectivity. Depending on the parameter settings and on the original data, large numbers of maps may be produced. In order to focus on the most interesting ones,
we propose a method to rank them by computing the normalized mutual information between the spatio-temporal localization maps extracted from the SITS and the ones extracted from the same but randomized SITS. The latter is obtained using a swap-randomization technique. Experimental results on a Landsat 7 SITS covering New Caledonia are presented.