DECOL: DEep COntinual Learning from Satellite Image Time Series – DECOL

Modern imaging sensors frequently capture satellite images of all land surfaces, called satellite image time series (SITS). Over the last decade, automatic analysis of SITS has advanced with the development of scalable and accurate deep neural network architectures capable of handling the temporal structure of SITS. However, all these approaches are trained in a static fashion: a new model is learned for each dataset without taking advantage of the knowledge accumulated over time. In addition to being computationally and temporally expensive, waiting for a full series of satellite images is also suboptimal because satellite images are usually processed in a few days. A dynamic learning solution consists in updating a model from a data stream by accumulating knowledge over time without forgetting relevant information from the past. In the era of deep learning, this strategy is known as continuous learning. Another problem is the scarcity of labeled data to continuously train these models. Most of the existing solutions (e.g., domain adaptation) are also trained statically and cannot thus be updated when new data arrives.

The DECOL project aims to design novel continual learning techniques for SITS. Besides, DECOL will study the robustness of the proposed methods to artifacts in SITS data (cloud cover, changes in acquisition conditions, etc.). It will also investigate continual learning of time series representations in scenarios where labeled data are scarce including continuous domain adaptation and prior-knowledge integration. The proposed techniques will be applied to the flagship application of forest monitoring in the Amazon.