Speaker
Description
Adult emerging aquatic insects provide ecosystem services at the landscape scale. Their emergence from water areas represents a transfer of energy from aquatic to terrestrial ecosystems. In particular, their carcasses and excreta subsidize terrestrial systems with organic matter. Agricultural intensification is reshaping aquatic and terrestrial ecosystems, impacting sensitive taxa, such as Plecoptera and Trichoptera, and the services they provide. Mapping their terrestrial distribution is therefore a major challenge, both for environmental management and for ecosystem services mapping.
Species Distribution Models (SDMs) correlate field occurrence data with environmental data to predict species distribution. DeepSDMs, a recent branch of SDMs relying on deep learning, are increasingly used as the quality of their predictions often outperforms that of traditional SDMs (e.g., GLM, Random Forest, MaxEnt). These DeepSDMs often use raw remote sensing products, like satellite or LiDAR imagery, as input data. A key methodological choice is the grain size (i.e., the spatial resolution) of the environmental data. While the impact of grain size is well studied for traditional SDMs, it remains underexplored for DeepSDMs.
In this study, we develop DeepSDMs for 2 emerging aquatic insects with contrasting dispersal capacities (Plecoptera and Trichoptera). We compared predicted distributions across 6 remote sensing products with different grain sizes: drone multispectral data (32 cm), drone LiDAR (40 cm), public LiDAR (50 cm), drone multispectral data resampled to 2.5 m to mimic high-resolution satellite, super-resolved multispectral data (2.5 m), and medium-resolution satellite multispectral data (10 m). We also varied the spatial extent of the environmental context using image patches centered on each species sampling location. Patch sizes were 40×40 m, 70×70 m, and 100×100 m.
The best results for Plecoptera were achieved at 40 m using 2.5 m multispectral data (resampled from drone imagery), while adding LiDAR degraded performance. The best results for Trichoptera were achieved at 100 m using 2.5 m multispectral data combined with public LiDAR. These results suggest that the choice of sensors, extent, and grain size should be tailored to each species. Moreover, the optimal combination of these parameters is not obvious a priori and only emerges after testing several of them. This approach provides practical guidance for implementing DeepSDMs across multiple taxa.
| Are you participating to the "AnaEE Environmental Rising Star Award "? | No |
|---|