damian-oswald / process-rate-estimator Goto Github PK

Create three scatterplots with the d18O on the x-axis and SP on the y-axis, where Nitrification, Denitrification and Reduction color the points according to their magnitude.

For an example and further explanation, see Lewicka-Szczebak et a. (2017).

The hyperparameter visualizations are too long down the page, a panel tabset would solve this issue and allow the reader to manually switch between visualizations.

The standardized regression coefficients $\beta^\ast$ and the semi-partial R² ($\Delta R^2$) are very much related, but they are not exactly the same, as described by Chris Novak (2015). The difference between these is the denominator, which for the $\beta^\ast$ and $\Delta R^2$ contains only the correlation between the predictors. If the correlation among predictors is low, the two values will be close.

The partial R² needs to be calculated from Type-III sum of squares ANOVA; i.e. the default method in R leads to wrong results.

Similar issue: https://stats.stackexchange.com/questions/65919/difference-between-effect-size-partial-r2-and-coefficients

Write a script which runs the model in order to generate the data for a global sensitivity analysis.

Main parameters of interest:

• Isotope end members
• Bulk density

In table 6.1, check the distribution functions of the isotope end member parameters. The range from which these parameters are sampled massively influence their estimated importance to the overall model outcome.

The process rate estimator assumes (for now), that all soil parameters such as bulk density are constant no matter the measurement depth.

However, it is much more realistic that some soil parameters change as a function of the measurement depth.

This could be incorporated by allowing parameters both in scalar form, where they would be repeatedly used for all depths, as well as in vector form, where they are provided for every depth at hand.

This should include the following content:

• Table with the uncertainty values as well as their source.
• Function used to randomly generate uncertainty values (based on runif, rbeta or rnorm).
• Exact process of the sensitivity analysis.
• Results: Which parameters (most importantly isotope end members) are how important?

Additionally to the sensitivity analysis, investigate the sensitivity of the process rate estimator to bulk density (BD), specifically.

Analyze the process rate estimates based on the wheat cultivar. This is one of the main research questions: What is the effect of variety on the process rates?

The results visualization over time don't work that well. They need to:

• Be in a wide layout to better match the interactivity.
• Include x and y axis descriptions with units.
• Text needs to be larger and better cover the available space.

In table 6.1, adjust the parameter range to match the conducted sensitivity analysis. The range from which these parameters are sampled massively influence their estimated importance to the overall model outcome.

Quantify the magnitude of the process estimate uncertainty per column and date.

• Use the covariance matrixes $\Sigma$ of figure 6.2
• Apply the Frobenius Norm for quantifying the magnitude of $\Sigma$. The Frobenius norm for a matrix $A$ is defined as:

The interpolated data for N2O, SP, and d18O are missing values for column 6 at depth 90. This leads to errors in the loops when running the PRE later on.

Here's some code showing the error:

library(PRE)

with(measurements, tapply(d18O, list(column, depth), function(x) sum(!is.na(x))))
with(measurements[measurements$column==6&measurements$depth==90,], plot(SP, date))

data = getMissing()

with(data, tapply(d18O, list(column, depth), function(x) sum(!is.na(x))))

with(measurements[measurements$column==6&measurements$depth==90,], plot(date, SP))
with(data[data$column==6&data$depth==90,], points(date, SP))

This R script should run in one go and write all tables figures in the Decock et al. (2022) -- at least, all figures where the figure values are dependent on the process rate estimator.

Compute overall share of variance explained as:

sort(rowSums(df)/(sum(rowSums(df))/mean(R2)))

• As for now, the sensitivity analysis is conducted by using linear regression models.
• Specifically important: Mixed effects models, where the coefficients are estimated conditionally on the column and depth combinations.
• Emulators such as random forests or XGBoost could be used as well.

• After figure 6.1, add a table of both the coefficients as well as the standardized regression coefficients.

• For process rates, it would be: g N2O-N per ha per depth increment

This R script should run in one go and write all tables present in the Decock et al. (2022).