Using neural networks for change detection and classification of COSMO-SkyMed Images

In this paper advanced neural networks (NNs) based techniques are presented to detect and classify changes in multi-temporal SAR COSMO-SkyMed images. The core and more innovative algorithm of the methodology is based on PCNNs (Pulse Couple Neural Networks), which perform the change detection task, while a standard Multi-Layer Perceptron (MLP) topology is considered to identify the type of changes occurred. The paper aims at providing a complete automatic processing chain usable also for extended areas. In fact, the PCNNs, being unsupervised networks, are inherently automatic. On the other hand, MLPs, which are supervised networks, are assumed to be learned off-line and plugged into the workflow without further training. The experimental results, which have been obtained over various urban areas in the city of Rome, are encouraging and confirm the validity of the considered approach.