Channel-Based Attention for LCC Using Sentinel-2 Time Series
This addresses interpretability issues in remote sensing for researchers and practitioners, but it is incremental as it builds on existing attention methods.
The paper tackles the problem of unclear rationale in deep neural network predictions for land cover classification using satellite image time series by proposing an architecture that uses convolutional layers and an attention mechanism to weight channel importance, showing promising results in experiments with Sentinel-2 data.
Deep Neural Networks (DNNs) are getting increasing attention to deal with Land Cover Classification (LCC) relying on Satellite Image Time Series (SITS). Though high performances can be achieved, the rationale of a prediction yielded by a DNN often remains unclear. An architecture expressing predictions with respect to input channels is thus proposed in this paper. It relies on convolutional layers and an attention mechanism weighting the importance of each channel in the final classification decision. The correlation between channels is taken into account to set up shared kernels and lower model complexity. Experiments based on a Sentinel-2 SITS show promising results.