I am trying to predict/discriminate macrophyte groups in rivers (https://snwikaij.shinyapps.io/RF_macrophyte_models). I have run out of viable options to predict groups with an "acceptable" model performance. Does anyone has some suggestions/experience on groups of macrophytes that can be clearly predicted/discriminated?
These groups do not have to be linked to the dataset I use, but it has to be linked to literature and logic(?). I have provided the included species in the added Excel file, in which suggestion can be added. The the predictors in the models are substrate type (fine and coarse, categorical), depth, flow velocity, alkalinity, total phosphorus and conductivity. Based on trial-and-error and literature, I find an accuracy of >50% and Cohen's kappa of > 0.3 acceptable. This amounts to maximum 3-5 groups. Currently the models predict each species individually, growth forms, river types, bryophyte and vascular plants, and HCO3 and CO2-only-users.
hi
this is for rivers in plains, isn't it?
What about the choice of species?
May be that the Ranunculus subgen. Batrachium deserves to be better represented that only by penicillatus.
No Cinclidotus?
What about grouping freely floating vs rooted plants? Ok there are some intermediates.
best regards
marc
Marc Philippe Thank you for responding. Yes, the models are based on rivers although they also seem to perform similar for other linear waterbodies that are not flowing (e.g. channels/ditches). I guess some rivers are very slow flowing or located in agricultural areas sharing some resemblance.
Sadly, for now, only R. fluitans and R. penicillatus are included. Similar only C. hamulata and C. platycarpa are included. Many more "important" species are missing, such as Ranunculus spp., P. polygonifolius, P. alpinus, multiple Myriophyllum spp. and Lemna spp. While these species seem to have very specific locations in rivers. I wish, I could also include these. But, I set some arbitrary rule that I only include species that occur >= 50 times (this is already relatively low). By trial-and-error I noticed that at +/- < 50-100 the models get a bit wobly. I am unfortunately restricted to the data I currently have.
For freely floating vs rooted plants I have some loosely based grouping in the app. In the left bottom corner you can find "Choose grouping" and select "Growth forms". Then moving to the top of the app you can select the tab "Group barplot", showing the predictions of the models based on the factors selected in the left panel. You can also select "Group partial dependency plots" at the top which show the the partial dependency plots for each group. However, I do not know how to define the bryophyte growth form, so this is included as bryophyte.
Hey! I know nothing about water plants, but I know that there are methods in SPSS modeler that can help you find Correlations between variables (https://www.youtube.com/watch?v=L9yFGtscyL8&t=1s) or Auto cluster node (https://www.ibm.com/support/producthub/icpdata/docs/content/SSQNUZ_latest/wsd/nodes/autocluster.html). Maybe this helps.
Laura Dzelzkaleja Thank you for responding. I have tried so many grouping methods, from spin-glass and random-walk algorithms, to clustering co-occurrence matrix and clustering presence-absence matrix. The most simplest solution presence-absence matrix seems to work "best". I have also seen people clustering positions of species within multivariate species (e.g. CCA). This often results in large number of groups, but these cannot be clearly discriminate/predicted. Additionally, the later one this circle reasoning since you first group species in the dataset based on the predictors and then try the predict them based on the same predictors. Also, everything is correlated with everything, which is not necessarily a big issue for the models, but more for the interpretation. For the last point I have not yet found a meaningful solution.
Ok, so maybe its time to start clustering by some other parameters? Like mineral composition in those organisms? Or some behaviour patterns (when they go dormant, growth rate etc.)? Or size. Or microscopic structures like cell spatial distribution patterns in organisms. If any of those are available and what I'm saying makes sense :D Maybe also a way to go could be exclude some of the parameters, or analyse why there are little or no correlations between some of them. Wish you good luck and a spark of inspiration ;)
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis