Continual Learning for Semantic Segmentation of Remote Sensing Images

The gradually availability over time of new training data for Neural Networks (NNs) causes problems for traditional Machine Learning approaches: models are affected by catastrophic forgetting of previously learned information; moreover, samples representing dynamic environments have shifts in the distribution; the big amount of samples makes impossible to manage the entire dataset simultaneously, in terms of excessive computational cost and memory required. Continual Learning (CL) algorithms are therefore introduced to train NNs, gradually updating the model with new available samples. Remote Sensing Earth Observation Segmentation Task suffers from all these issues and CL algorithms can be applied to preserve old classes. Three approaches proposed in literature have been considered, exploiting CL regularization-based algorithms and heterogeneous datasets (with different sources, classes, point of view, resolution) provided sequentially over time. Results show that they reach better performances in the Segmentation Task, overcoming the catastrophic forgetting and improving the adaptation of the model to data shifts.