From Local to Global: A Transfer Learning-Based Approach for Mapping Poplar Plantations at Large Scale

Within the current context of availability of Earth Observation satellites at high spatial and temporal resolutions, mapping large areas become doable. To this end, supervised classification of remote sensing images is the commonly adopted approach. Having a high-quality and representative training set is always the key to a successful classification result. However, this is often a tedious task that involves samples gathering from field surveys or photointerpretation. The larger the area to map, the more challenging this exercise becomes. In this letter we present an active learning-based technique to address this issue by optimizing the training set required for classification while providing a generic classifier suitable for large scale. Experiments were carried out to identify poplar plantations in France using Sentine1-2 time series. The results are promising and show the good capacities of the proposed approach to be adapted at the national scale.