Properties2
| Type | Concept |
| Note created | Feb 17, 2025 |
A downscaling model is used to change the resolution of spatially-indexed data from larger-sized cells to smaller ones. While naive methods like areal interpolation and weighted averaging can lay good results in the simplest use cases, downscaling uses a regression model that uses proxy variables (that are known in the smallest desired scale) as independent variables and the large-scale data as the target variable.
To train such model, the large-scale is used to find a relation between the proxy variables and the data at hand. Once trained, it is applied to the small scale cells, laying an initial estimation of the small scale data. After that, a reconciliation method is used to make sure that these small-scale estimations add up to the known, large-scale data.
It is important to find an adequate model for this task: decision trees are a bad choice since the cuts will be fit to the large scale samples and will have issues generalising to the small scale. Instead, linear regressions or other linear methods are a better approach.