We propose a landmine detection algorithm that uses ensemble discrete hidden Markov models with context
dependent training schemes. We hypothesize that the data are generated by K models. These different models
reflect the fact that mines and clutter objects have different characteristics depending on the mine type, soil
and weather conditions, and burial depth. Model identification is based on clustering in the log-likelihood space.
First, one HMM is fit to each of the N individual sequence. For each fitted model, we evaluate the log-likelihood
of each sequence. This will result in an N x N log-likelihood distance matrix that will be partitioned into K
groups. In the second step, we learn the parameters of one discrete HMM per group. We propose using and
optimizing various training approaches for the different K groups depending on their size and homogeneity. In
particular, we will investigate the maximum likelihood, and the MCE-based discriminative training approaches.
Results on large and diverse Ground Penetrating Radar data collections show that the proposed method can
identify meaningful and coherent HMM models that describe different properties of the data. Each HMM models
a group of alarm signatures that share common attributes such as clutter, mine type, and burial depth. Our
initial experiments have also indicated that the proposed mixture model outperform the baseline HMM that uses
one model for the mine and one model for the background.