Deep learning for environmental monitoring by transforming earth observation data into geo-indicators

The recent Sentinel satellite constellations deliver open earth observation data with suitable characteristics – regarding spatial, spectral and temporal resolution – for environmental monitoring. Unfortunately, the uptake of remote sensing-based information for environmental policy support remains limited as it cannot fully replace information obtained from ground-based measurement campaigns. Recently, deep learning algorithms are gaining popularity in a broad range of application domains within the remote sensing community. The fact that these models are data-driven could raise policy makers’ confidence in the resulting output. However, the large quantities of labelled samples, crucial to effectively train these models, are lacking for many environmental applications. Existing methods to increase the effectiveness of limited training datasets include active learning and transfer learning. The aim of this PhD research is to apply deep learning to satellite imagery and test these augmenting techniques on real applications within the environmental policy sector.