Data Interpolation using XGBoost
Opened this issue · 1 comments
For the NANs in our data, we consider many ways to do interpolation. One of them is XGBoost. Since XGBoost is not adapted to spatial data, it naturally has difficulty in understanding the spatial relationship in data. Simply applying XGBoost on our data would give us interpolation like this.
To further improve the quality of data, we consider using a hybrid way of interpolation. First, we create a list of data that is the XGboost prediction of NaNs. Then using them as a base line, we apply IDW on them and combine both predictions to get a better result. The interpolation using this hybrid way looks like this:
Also tried the new SGS method which enables us to use multi-processing to accelerate the kriging process. But it still suffers from the resolution problem. Currently we are using a 200m resolution, the whole process is about 100s. Once we reduce it to 100m resolution or even higher resolution, the RAM would explode in Variogram computation.
