Implementing Kohonen´s SOM with missing data in OTB

This paper focuses on the implementation of the Kohonen´s Self Organizing Map (SOM) through the Orfeo Toolbox. It makes the link between the theoretical background and the strategic choices in the implementation: (1) Generic choice of the learning functions that ensures convergence of the training process; (2) Generic definition of the neuron that allows flexible use on multicomponent data, time series and also heterogeneous data; (3) Flexible class inheritance to redefine the update function to make the SOM deal with missing data. Example are been given with the preprocessing of low resolution time series contaminated by clouds and shadows due to weather conditions during the acquisitions. The technique is capable to reconstruct a complete time series free of clouds from MODIS data. It will also focus on the interest of a sparse distance criteria to reconstruct the time series on presence of haze that is difficult to detect.