A Neural Approach Under Active Learning Mode for Change Detection in Remotely Sensed Images

In this paper, a change detection technique using neural networks in active learning framework is proposed under the scarcity of labeled patterns. In the present investigation, two variants of radial basis function neural networks and a multilayer perceptron are used as learners. Instead of training the network (or ensemble of networks) with randomly collected labeled patterns, in the proposed work, the network (or ensemble of networks) is iteratively trained with label patterns, collected using the query functions. Here, two query selection strategies are used: uncertainty sampling and query-by-committee. In this way, the most informative set of labeled patterns can be iteratively generated by querying. To evaluate the effectiveness of the proposed approach, the experiments are conducted on multi-temporal remotely sensed images. The results obtained using the proposed active learning framework are found to be encouraging.