Random Forest climate downscaling

Knowledge graph centered on Random Forest climate downscaling with 22 nodes and 48 connections. Top connected: Crested Butte, Daymet: Daily Surface Weather Data on a 1-km Grid , Multiscale Model to Model Cloud Compare method, Daymet: Daily Surface Weather Data on a 1-km Grid , Upper East River Beaver Pond.

Method synopsis

Machine learning approach using Random Forest algorithms to downscale coarse resolution climate data to higher spatial resolution using topographic predictors. Models are trained on relationships between climate variables and geographic features, then applied to generate fine-scale climate surfaces.

Synthesized from method descriptions across 1 paper using this protocol.

Procedure as described in the canonical source

Steps below were extracted from the paper that introduces this protocol — Downscaled hyper-resolution (400 m) gridded datasets of daily precipitation and temperature (2008-2019) for East Taylor subbasin (western United States) (2022), Earth System Science Data Discussions. Implementations in other papers (listed below) may differ.

1. Generate nearest neighbor weather station maps
   For each grid point, identify the ten most correlated weather stations using entire 12-year time series. Create nearest neighbor maps showing spatial distribution of most correlated stations for each climate variable.
   Quantities: Ten nearest neighbors per grid point across entire study areaDuration: Computed once using full 12-year time seriesConditions: Based on correlation analysis across all years
   Equipment: Weather station correlation algorithms
2. Train Random Forest downscaling models
   Use longitude, latitude, elevation, and ten nearest neighbor weather station values as predictors. Train separate Random Forest models for each day to predict precipitation and mean temperature. Use 500 trees with p/3 features per split and minimum node size of 1 for spatial downscaling.
   Quantities: 13 predictor variables (3 topographic + 10 weather stations), 500 trees per model, separate models for each dayDuration: Models trained for each day of 12-year periodConditions: Separate training for wet days (>1 mm precipitation) for precipitation models
   Equipment: Python scikit-learn module, Random Forest algorithm
3. Apply spatial filtering
   Subject downscaled precipitation grids to variable filter performing distance-weighted average of surrounding grid cells to ensure smooth precipitation field in low-gradient areas without affecting high-gradient areas.

Typical Equipment

• PRISM climate dataset
• Daymet dataset
• Random Forest algorithm
• Digital elevation model

Output Measurements

• daily precipitation
• daily temperature
• 400 m resolution climate grids

Papers Using This Protocol (1)

Places Where Used