I am setting up an experiment to estimate the accuracy of different interpolation algorithms for generating a spatially continuous rainfall data (gird) for a certain area. The data density (number of points versus the area) and the spatial arrangement of the data (i.e., random versus grid-based ), will vary for each run (attempt)
The objective is to understand how each algorithm performs under varying data density and spatial configuration.
Typically, different studies have done this using station data of varying data density and spatial configuration. In the current context, there are limited stations (just about 2)and the intent is to execute this experiment using data sampled (extracted) from existing regional rainfall grid, but varying the data density as well as the spatial configuration.
Note that I cannot generate random values because the kriging is to be implemented as a (multivariate stuff ) using some rainfall covariates....random values will mess up the relevance of such covariates.
I did a rapid test and found that, despite a wide difference in the density and configuration, there was no significant difference in the accuracy result, based on cross validation results. What's going on? It's not intuitive to me!!
Please, can you identify something potentially not correct with this design? Theoretically, is there anything about dependency that may affect the result negatively? Generally, what may not be fine with the design? how can we explain this in your view??
Thanks for your thoughts.
Sorry for many text..
Can you maybe put some pictures as examples? And also some information on how you are using that grid further into your steps?
From what I can understand, since you used some real life data and used it to make the grid the results from different methods are similar? That does make sense, even though there are different methods for interpolation to make raster from vector of points, they all are designed to make the continuous values equivalent to the discrete points. There will be differences between the values you get for cross validation from each of them, but unless there is effects of some outlier values, most of them should give similar values.
The effectiveness/suitability of one algorithm over others only come into play when you need to use the grid for something else, and how the further calculation steps use the interpolated grid.
You don't supply any information that would allow someone to assist. Like the Area of Interest, maybe with the sample points, the number of sample points, the distributions of the sample points, etc. Since you are using weather data, the AOI and points are critical because the algorithms aren't magic, if the data is too sparse, all of them will give basically the same answer. The ones that rely on an appropriate kernel density especially. The is an essential assumption underneath any interpolation, that there is some continuous manifold ( surface ) - if there is orography, and your sampling rate is insufficient, that essential precipitation influence may fail to be accounted for or have an extreme effect. Points on the boundary of the point distribution are not treated the same.If you sample points are extremely stretched out in one axis or another affects the points influence.
Hi Elijah A Njoku,
I am attaching a paper which can help with your problem.
In it, the authors make a spatial interpolation of a limited number of points (observations of a climatic variable) using several spatial interpolation techniques. They used gridded data as auxiliary variables to improve the interpolation process via a Random Forest algorithm. Analysis of the error in the results shows that this type of interpolation can be very effective.
Hope it helps, cheers,
Augusto
Article Spatial Interpolation of McArthur's Forest Fire Danger Index...
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts