I have 1851 soil samples data on pH covering a study area of 7482sq.km in northern Ghana and I am using 52 environmental long-term average variables (Relief, Climate, MODIS Reflectances and derived products) to fit a model in order to explain the variability for pH prediction. So far, all models tested have shown low explained variance and sometimes even gives negative.
How may I improve the Explained Variance below?
Kindly see attached, the spatial distribution of points, and metadata excel file showing details about the covariates used.
Below is the summary of my models explained variance.
Multiple Linear Regression: 0.03
Step-wise Multiple Linear Regression: 0.04
RandomForest: -7.02
ExtremeGradient Boosting: 0.03
Support Vector Machines with Polynomial Kernel: 0.03
Additional information about the sample data
My assessment so far
I would be most grateful for insights into any techniques that could help improve the model explained variance.
Let me recommend this work to improve the evaluation part:
Conference Paper Data Formats and Visual Tools for Forecast Evaluation in Cyb...
As for the modelling part, you need to consider variable transformations. You can use this guide:
https://xdplot.blogspot.com/p/data-transformations-guidelines.html
But you'll need to back-transform the results.
To evaluate your models after back-transformations, you need to use the RelMAE (or AvgRelMAE), as noted here (see page 3):
Chapter Forecast Error Measures: Critical Review and Practical Recommendations
Hi, Isaac,
Thank you for describing your data, analysis objective, and results thus far.
Perhaps the solution you seek evades detection vis-à-vis parametric linear methods.
I’ll simply recommend that you conduct what is known as a novometric analysis, treating pH as the ordered class (“dependent”) measure, and the 52 environmental variables as attributes (“independent variables”).
Explained/illustrated in the brief linked article, this analysis will identify the (non)linear model which most accurately predicts pH score. Described in the article, confirmatory and/or exploratory hypotheses may be evaluated, no distributional assumptions are required (p is exact), classification accuracy is assessed in validity analysis, and accuracy is adjusted for chance and for model complexity.
- https://odajournal.com/2020/06/28/what-is-novometric-data-analysis/
The only software which conducts this analysis is available at no cost--a PC with a Pentium chip running Windows(tm) is needed:
- https://odajournal.com/2020/10/06/megaoda-software-within-r-package-now-available-on-github/
Lots of other examples of novometric analysis are available online (type novometric in the search box):
- https://ODAJournal.com
You describe an interesting analysis! If you try using maximum-accuracy methods, please contact me if I may be helpful in that regard--I'd love to read your findings.
Happy 2021 (to the World)
I suggest that you consider three references: (PDF) To Explain or to Predict? (researchgate.net) and Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating | Ewout W. Steyerberg | download (1lib.us) as well as An Introduction to Statistical Learning with Applications in R | Gareth James, Daniela Witten, Trevor Hastie, Robert Tibshirani | download (1lib.us)
I noticed that you used predict and explain in the same sentence. The first thing you should do is write a clear research question. Do you wish to predict or explain.? The methods of doing these are quite different as are the model building methods. There are many methods to chose from once you are clear on your goals. When you have your research question, ask again and we will be glad to help you choose. The above references should assist you. Best wishes, David Booth
Dear all,
Many thanks for helping me to think through my Digital Soil Mapping Analysis. Your references were also valuable.
In the end, Ordinary Kriging was the best estimator as the machine learning models were not able to capture enough variability in the observation data which was due to poor spatial dependence (i mean more points closer to each other have dissimilar pH values).
I reckon that if I have to use machine learning, I may need to use many hundreds of covariates where each covariate tries to explain a small portion of the variability and finally scale them down to the best predictors. This process is iterative and highly time-consuming but would be considered later.
I attach here, the Ordinary Kriging pH prediction results and a 90% prediction interval for quantifying uncertainty.
Thanks a lot!
Are you using 52 predictors to explain pH spatial distribution? This is way too much. Select your predictors using a stepwise procedure (as in multiple regression) or perform a redundancy test. If no predictor significantly explains the pH distribution, then just apply ordinary kriging to interpolate pH values. Alternatively, you may try a local regression interpolation, like GWR.
Dear @Christian,
Thanks for your input. I have tried multiple regression(please see my very first post on the results which was so poor)
Yes I did use O.K to interpolate. I would consider GWR as well to see the accuracy.
Many thanks
- Badges
- Science topic
- Similar topics
- Agricultural Science
- Soil Science
- Soil Analysis