| Reference: | EARL LSTM |
| Publication: | Klotz et al. (2024): A runoff reconstruction dataset for Europe |
| Dataset: | TK |
| Name (contact): | Daniel Klotz (daniel.klotz@ufz.de) |
| Model type: | LSTM |
| Streamflow source: | Based on data references from EStreams (do Nascimento et al., 2024). |
| Dynamic inputs | |
| 1. Source: | E-OBS (Klein Tank et al., 2002; Haylock et al., 2008). |
| 2. Inputs: | Daily precipitation and mean temperature. |
| 3. Comment: | Aggregated at basin level. |
| Static inputs | |
| 1. Source: | HydroATLAS (Linke et al., 2019). |
| 2. Inputs: | 10 inputs: Average fraction of sand and silt in the soil, basin area, minimum/maximum/average elevation, average precipitation, mean potential evaporation, and the basin elongation ratio. |
| 3. Comment: | Follows the procedure of Caravan (Kratzert et al., 2023). |
| Model specs | Hidden size of 256 CMAL head with a single component (Klotz et al., 2022) Trained for 30 epochs, validated on a random hold-out of 500 basins Dropout rate: 0.4 Optimizer: AdamW Learning rates (epoch:lr): {0:0.001, 5:0.0005, 20:0.0001} |
| Intended use: | Simulate catchments in Europe |
| Caveats & Recommendations: | The model is trained with a single forcing product (E-OBS) only. Furthermore, the overall quality of the reconstructions is restricted by (a) the streamflow observation quality, (b) the input quality, and (c) the capabilities of the EARLS LSTM (see: Klotz et al. 2024). |