Multilayer Perceptron for Change Detection of Remotely Sensed Images

Multilayer Perceptrons (MLPs) have been proven to be an effective way to solve classification tasks. A major concern in their use is the difficulty to define the proper network for a specific application, due to the sensitivity to the initial conditions and overfitting and underfitting problems which limit their generalization capability. Moreover, time and hardware constraints may seriously reduce the degrees of freedom in the search for a single optimal network. A very promising way to partially overcome such drawbacks is the use of MLP ensembles. In this thesis the focus is on the analysis and development of Ensemble of Neural Networks for change detection in supervised context. An ensemble is a system in which a set of heterogeneous artificial neural networks are generated in order to outperform the single-network based classifiers (especially when all are weak). Supervised change detection techniques generally represent the most accurate methodological solution for mapping land-cover changes while identifying the associated land-cover transitions between two different dates [1-4]. However, the application of these techniques depends on the availability of exhaustive ground-truth information for all the land-cover classes present in the area of interest at the times under investigation. Such a requirement is seldom satisfied since gathering a reliable ground truth for all the classes characterizing the considered scenes at the two dates under analysis presents several practical drawbacks and limitations (both in terms of time and economic cost) that may render this task almost impossible in most real-life cases. Nevertheless, to solve these specific types of problems, it would be highly beneficial for an operator to rely on a robust automatic technique that may allow an effective detection of the “targeted” land-cover transitions by taking into account only ground-truth information for the few classes of interest at each date (thus, avoiding the burden and cost associated to the collection of a full and exhaustive ground-truth data set at both times). When a few labeled patterns are present, then the problem can be approached in two different ways: either using the concept of semi-supervised learning (used since 1970 [5]) or using the concept of ensemble learning (started since 1965 [6]). In the present study, we are concentrating only on the ensemble learning. In the present work each classifier in the ensemble is trained using all labeled samples available. Then the ensemble helps in generating a label of unknown patterns using some of the non trainable combiners. Experiments are carried out on two different multitemporal remotely sensed images. Experimental results reveal that the ensemble learning approach reduces classification error over the weakest classifier.