From tree cover to tree species - Leveraging deep learning to map Swedish forests using Sentinel-1 and -2 observations

Forests provide several ecosystem services that are directly beneficial to human health and wellbeing such as sequestering carbon, regulating water quality and temperature, supplying timber, stabilizing soil, and enhancing biodiversity. However not all forests are the same.

Forests with a high diversity of tree species provide more ecosystem services and are more resistant to pests than forests with fewer tree species.

The main aim of this proposal is to bridge the gap in remote sensing research in terms of the large-scale mapping of forest tree species composition at a finer spatial resolution than previously done using a modern approach.

I would like to investigate the applicability of state-of-the-art deep learning algorithms to map tree species composition in a predominantly forest landscape covering 12% of Sweden (52,016 km2) using both synthetic aperture radar backscatter from the Sentinel-1 satellite and multispectral reflectance from the Sentinel-2 satellite.

To this end, I hypothesize that (1) the complementary combination of multitemporal optical and radar satellite observations is able to capture spectral, spatial, and structural representation of tree species in Swedish forests that enable their identification, and (2) the proposed deep learning architecture provides higher classification accuracies in identifying tree species using these complementary data relative to conventional shallow learning algorithms.

By addressing these hypotheses through five research questions, I hope to improve the current state of tree species mapping and in the process create data products that increase our understanding of the ecology of tree species such as fine-scale community dynamics and the contribution of trees to ecosystem functions and services.