fish starvation issues during parameterisation
94 views
Skip to first unread message
Atlantis ecosystem model
May 19, 2015, 7:48:52 AM5/19/15
to atlantis-eco...@googlegroups.com
[Re-posted with permission from Marc and Alex questions to Beth and Bec]
Hey Beth and Bec,
Marc and I are working hard to finalise the model calibration. Unfortunately we are still having issues with fish starvation. We tested a huge set of parameter settings to solve this issue focusing mainly on clearance (C), growth (MUM) and availabilities (avail). Here is a summary of the stuff we tested:
1. Multiplying our initial values for C and MUM by 2 - 100. Both parameters are based on weight@age data. For each simulation all groups and ageclasses were changed simultaneously. Changes to C and MUM were tested separately.
2. Same settings as in (1) using only planktivorous fish species to test the effect on lower trophic levels. In addition we increased C and MUM simultaneously.
3. To exclude the posibility of growth limitation, we increased the default MUMs by 100 and tested the effect of increased C and avail seperately and in combination.
4. Same settings as in (3) but with default MUM values.
5. Same settings as in (3) using default MUMs * 2 and age-specific correction factors (age_mult = 3, 2.5, 2.2, 1.7, 1.35, 1.15, 1.05, 1, 1, 1 for ages 1:10).
6. Same settings as in (3). To exclude a spatial mismatch of predator and prey we based the movement patterns on polygon size.
7. Same settings as in (3). In addition to harmonised movement patterns we also distributed the initial numbers proportional to polygon size.
Despite these various settings the majority of age-structured groups (especially fish groups) were starving throughout the simulation period (ResN/ResN_initial and StructN/StructN_initial <0.5, generally <0.1). Surprisingly in most cases relative individual weight sharply drops at the beginning of the simulation and remains at low levels afterwards. A stable individual weight and even an increase (which is not desirable but should be possible in theory) was only achieved using parameter setting (3) with high MUMs, C and avail. However, model results were highly variable in regards to agestructure which was composed of a small number of ages (sometimes only agecl 10 was present). In addition, increasing MUM by a factor of 100 to achieve stable relative weights over time would lead to a discrepancy as MUMs were no longer based on actual data (weight@age).
Summing up our key questions are:
1. How can we fix our persistent starvation issues?
2. Is it possible to achieve stable weights over time using MUM values which are based on weight@age data?
3. Are MUM, C and avail mathematically interconnected (in our base version we used the "American-approach" with C = MUM/10) or do we need to parametersie them idenpendantly?
4. Regarding biomass timeseries we sometimes had to implement linear and quadratic mortalities to compensate for high MUM, C and avail ranging from 10-2 to 10-6. What is your advice here? Is it fine to use those mortality terms or do you aim to avoid them?
At the moment we are running some sort of sensitivity analysis to minimise the parameter space. While MUM values are kept at default values C is chosen from three different normal distributions for each group. The distributions represent a small (mean = 10), medium (mean = 50) and high perturbation (mean = 200). In addition the availability for each predator prey interaction is randomly choosen from 0.0001, 0.005, 0.01, 0.1, 0.5 to 0.9. Unfortunately we are having only minor sucess with this approach as 75% of simulations crash at time step 1 due to negativ fluxes of specific groups. We are not sure if this is a sign of an unstable base model or a poor randomisation of parameters which inevitably results in unrealistic scenarios.
## Added by Marc
As Alex wrote at the moment mums are about 10 x the clearance rate and that lifts some of the group to better starvation values where biomass values more or less fit as well. However, the groups with the high mum increase very fast and cannot be eaten away by the other groups not even on the long term.
Increasing avaiability solves some of that but values are in some cases already quite high and thus we had to include linear and quadr. mortality terms for many groups while some groups like meiofauna or small infauna do not build up biomasses as high as we would expect (based on EwE).
As there are so many combinations and not to many ways to use observations to ground some of parameters it feels a bit like trial and error. So if there is a possible "suitable range" that we should not leave like all lin. mortalities below 0.001 are fine, MUM should always be higher/lower than C etc. would help us a lot to obtimize the balancing.
Marc & Alex
Hey Beth and Bec,
Marc and I are working hard to finalise the model calibration. Unfortunately we are still having issues with fish starvation. We tested a huge set of parameter settings to solve this issue focusing mainly on clearance (C), growth (MUM) and availabilities (avail). Here is a summary of the stuff we tested:
1. Multiplying our initial values for C and MUM by 2 - 100. Both parameters are based on weight@age data. For each simulation all groups and ageclasses were changed simultaneously. Changes to C and MUM were tested separately.
2. Same settings as in (1) using only planktivorous fish species to test the effect on lower trophic levels. In addition we increased C and MUM simultaneously.
3. To exclude the posibility of growth limitation, we increased the default MUMs by 100 and tested the effect of increased C and avail seperately and in combination.
4. Same settings as in (3) but with default MUM values.
5. Same settings as in (3) using default MUMs * 2 and age-specific correction factors (age_mult = 3, 2.5, 2.2, 1.7, 1.35, 1.15, 1.05, 1, 1, 1 for ages 1:10).
6. Same settings as in (3). To exclude a spatial mismatch of predator and prey we based the movement patterns on polygon size.
7. Same settings as in (3). In addition to harmonised movement patterns we also distributed the initial numbers proportional to polygon size.
Despite these various settings the majority of age-structured groups (especially fish groups) were starving throughout the simulation period (ResN/ResN_initial and StructN/StructN_initial <0.5, generally <0.1). Surprisingly in most cases relative individual weight sharply drops at the beginning of the simulation and remains at low levels afterwards. A stable individual weight and even an increase (which is not desirable but should be possible in theory) was only achieved using parameter setting (3) with high MUMs, C and avail. However, model results were highly variable in regards to agestructure which was composed of a small number of ages (sometimes only agecl 10 was present). In addition, increasing MUM by a factor of 100 to achieve stable relative weights over time would lead to a discrepancy as MUMs were no longer based on actual data (weight@age).
Summing up our key questions are:
1. How can we fix our persistent starvation issues?
2. Is it possible to achieve stable weights over time using MUM values which are based on weight@age data?
3. Are MUM, C and avail mathematically interconnected (in our base version we used the "American-approach" with C = MUM/10) or do we need to parametersie them idenpendantly?
4. Regarding biomass timeseries we sometimes had to implement linear and quadratic mortalities to compensate for high MUM, C and avail ranging from 10-2 to 10-6. What is your advice here? Is it fine to use those mortality terms or do you aim to avoid them?
At the moment we are running some sort of sensitivity analysis to minimise the parameter space. While MUM values are kept at default values C is chosen from three different normal distributions for each group. The distributions represent a small (mean = 10), medium (mean = 50) and high perturbation (mean = 200). In addition the availability for each predator prey interaction is randomly choosen from 0.0001, 0.005, 0.01, 0.1, 0.5 to 0.9. Unfortunately we are having only minor sucess with this approach as 75% of simulations crash at time step 1 due to negativ fluxes of specific groups. We are not sure if this is a sign of an unstable base model or a poor randomisation of parameters which inevitably results in unrealistic scenarios.
## Added by Marc
As Alex wrote at the moment mums are about 10 x the clearance rate and that lifts some of the group to better starvation values where biomass values more or less fit as well. However, the groups with the high mum increase very fast and cannot be eaten away by the other groups not even on the long term.
Increasing avaiability solves some of that but values are in some cases already quite high and thus we had to include linear and quadr. mortality terms for many groups while some groups like meiofauna or small infauna do not build up biomasses as high as we would expect (based on EwE).
As there are so many combinations and not to many ways to use observations to ground some of parameters it feels a bit like trial and error. So if there is a possible "suitable range" that we should not leave like all lin. mortalities below 0.001 are fine, MUM should always be higher/lower than C etc. would help us a lot to obtimize the balancing.
Marc & Alex
asta...@gmail.com
May 19, 2015, 7:51:33 AM5/19/15
to atlantis-eco...@googlegroups.com, atlantis-eco...@googlegroups.com
Response by Beth:
Response by Asta:
Hello Marc and Alex,
From my experience of the SE Australian model (and a bit of the Baltic Sea model), the mum_ and C_ values are actually quite similar in many fish species. In the Baltic model some fish mum values are ca 1.5-2 times larger than C_ values, but for others C_ are 1.5 times larger. I have not done detailed variations for the different combinations of these parameters (we are doing some tests where both mum and C are defined through a three-parameter function, but I do not have results yet).
I have played a bit with the availability values, and in my opinion they should be based on some external biological knowledge or best guess, i.e. be independent of C_ and mum_ values.
Also it seems that many parameterisations do need linear and quadratic mortalities in the end, so I do not see why they should not be included. It seems that smaller linear and quadratic mortality values will make the system more dynamic and responsive to bottom-up and top-down control. I am now playing with the Baltic Sea model and trying to set them to below 1e-8, but higher values seem to be needed for mammal and birds that are not eaten by anything (or else they should have high starvation mortality, or some other control).
So perhaps, as Beth said, the best way may be to play with the numbers - adjust linear, starvation and quadratic mortalities and limit their recruitment, or make it more density dependent (beta parameter in BH recruitment relationship, if you use it).
Also, what about the min and max gape limits (KUP_ and KLP_ params)? Is the feeding window set to smooth or knife-edged? It does seem to make a big difference on some functional groups.
Finally, one can also ask how starvation should be defined. Currently in Atlantis the optimum weight is set as (2.65+1)*StN. I found that for some species this led to unrealistic length-fecundity relationship (fecundity drops rapidly with size, because fish is considered starved). Changing this to (2.65+0.01)*StN gave better relationship, but I suppose ideally one should go through the literature and explore this in a variety of species (which I am sure Beth did when setting this up)
We are doing a similar set of random parameter sampling, and yes, in 90% of cases runs crash due to negative fluxes. But if you can setup large batch runs in parallel, you can still get a fair number of runs from those successful 10% :)
All of these suggestions above may be useless if there are some important issues and too low biomasses in the lower trophic levels.
Do you have good tools to look at the who is eating what? Does predation mortality change when you modify C_ or mum_ values?
The solution will likely be a combination of things. For instance if you are "fitting" to biomass and have it about right but only because you have lots of individuals who are individually too small then the solution will lie in killing things off with predation (and fixed morality terms if absolutely necessary) not the mum and C.
Response by Asta:
Hello Marc and Alex,
From my experience of the SE Australian model (and a bit of the Baltic Sea model), the mum_ and C_ values are actually quite similar in many fish species. In the Baltic model some fish mum values are ca 1.5-2 times larger than C_ values, but for others C_ are 1.5 times larger. I have not done detailed variations for the different combinations of these parameters (we are doing some tests where both mum and C are defined through a three-parameter function, but I do not have results yet).
I have played a bit with the availability values, and in my opinion they should be based on some external biological knowledge or best guess, i.e. be independent of C_ and mum_ values.
Also it seems that many parameterisations do need linear and quadratic mortalities in the end, so I do not see why they should not be included. It seems that smaller linear and quadratic mortality values will make the system more dynamic and responsive to bottom-up and top-down control. I am now playing with the Baltic Sea model and trying to set them to below 1e-8, but higher values seem to be needed for mammal and birds that are not eaten by anything (or else they should have high starvation mortality, or some other control).
So perhaps, as Beth said, the best way may be to play with the numbers - adjust linear, starvation and quadratic mortalities and limit their recruitment, or make it more density dependent (beta parameter in BH recruitment relationship, if you use it).
Also, what about the min and max gape limits (KUP_ and KLP_ params)? Is the feeding window set to smooth or knife-edged? It does seem to make a big difference on some functional groups.
Finally, one can also ask how starvation should be defined. Currently in Atlantis the optimum weight is set as (2.65+1)*StN. I found that for some species this led to unrealistic length-fecundity relationship (fecundity drops rapidly with size, because fish is considered starved). Changing this to (2.65+0.01)*StN gave better relationship, but I suppose ideally one should go through the literature and explore this in a variety of species (which I am sure Beth did when setting this up)
We are doing a similar set of random parameter sampling, and yes, in 90% of cases runs crash due to negative fluxes. But if you can setup large batch runs in parallel, you can still get a fair number of runs from those successful 10% :)
All of these suggestions above may be useless if there are some important issues and too low biomasses in the lower trophic levels.
Do you have good tools to look at the who is eating what? Does predation mortality change when you modify C_ or mum_ values?
beth.fulton
May 28, 2015, 7:39:15 PM5/28/15
to atlantis-eco...@googlegroups.com, atlantis-eco...@googlegroups.com
G'day guys,
Sorry to take so long to get back to you.
Asta has given great advice which covers much of what I would have said.
After having looked over the files I'd say a bunch of your starvation comes from competition (so high numbers leading to competition for food and so size decreases as a result). I'd try dialling up availability of prey species that are getting exploding numbers to try and get their mortality rates up. Also make sure you're letting species access invertebrate prey at appropriate points in their lives (this is often an under appreciated supplementary prey set).
You can also try decreasing recruitment rates and increasing predation rates (though later perhaps more for top predators as you don't want to make the model dependent on "external mortality" rather than predation mortality). It will be a complicated mix of things most likely but this is where I'd start. Perhaps if we tick tack a bit more often for a few weeks (so I don't get behind and take so long to answer) we can work our way through the combination of potential influences.
I'm not 100% happy with any of the mortality reporting files, but the ones I use most are SpecificMort and SpecificPredMort (it is supposedly mort per year, but because of the way I stored the values you can get values >> 1 when in reality in the model caps are imposed so you can't get >1 - I have to record things better but its a bit difficult to do well given current structure of code so I it won't happen for a while)
Cheers
Beth
Reply all
Reply to author
Forward
0 new messages