| 385 | |
| 386 | == Parameterization and Evaluation (chronological order) == |
| 387 | |
| 388 | === Workflow === |
| 389 | The following outlines the strategy for parameterizing and evaluating the performance of ORCHIDEE-CN-CAN in simulating several key forest ecosystem processes including tree growth dynamics, energy exchange, C-N cycling and plant hydrology. |
| 390 | |
| 391 | The work will likely require several iteration containing all or just a coupled of the following: 1) running the model, 2) tuning key parameters, 3) re-running the model 4) spin-up and 5) evaluating to model, to discern parameter values that allow us to reproduce as closely as possible the global patterns and trends for several key processes. The tests will move across scales starting at pixel scale moving to longitudinal band, European scale and ending at global scale. Although the longitudinal bands are of little use in the evaluation itself they can be considered pre-tests before global tests are run. It is hoped that longitudinal bands would speed up the tests and reduce the computational cost. |
| 392 | |
| 393 | Table 0. Workflow of the parameterization and evaluation of ORCHIDEE-CN-CAN. |
| 394 | Phase Work to be done |
| 395 | Prepare and check Check whether the model runs and spin-up are apparently bug free (Table 1). Check whether the scripts, tools and data are available and still working (Table 2). |
| 396 | Initial evaluation Run longitudinal spin-up and check order of magnitude in soil carbon pools. Use the same runs to check response gradients to temperature, N-deposition, precipitation and management (§4.4). |
| 397 | Parameterization Use FLUXNET and some additional data sources to parameterize the model at the site level (Table 3). |
| 398 | Final evaluation Run a spin-up + transient simulation over Europe compare the simulation against the spatially explicit data (Table 4). |
| 399 | |
| 400 | |
| 401 | === Apparent bug-free === |
| 402 | Before starting the work proposed in Table 0 it needs to be confirmed that the model is technically capable of the tasks presented in Table 1. |
| 403 | |
| 404 | Table 1. Essential technical capabilities before evaluating ORCHIDEE-CN-CAN. |
| 405 | Description of the task Status |
| 406 | Is the model stable? Can it be run for 500 - 1000 years? Can it be run for all PFTs? Problems with grasslands |
| 407 | Is the N-cycle working? Can we run simulations with IMPOSE_CN=n and IMPOSE_CN=y? OK |
| 408 | Is land cover change working? Can we run simulations with (changing) land cover maps (impose_veg=no)? OK |
| 409 | Is forest management working? Can we read forest management maps, are the temporal trends in biomass what is expected for unmanaged, high-stand and coppice management? Still need to run idealized set-ups |
| 410 | Is litter raking working? Can we read litter raking maps and is the litter pool in forest decreasing and the litter pool in croplands increasing? Still needs to be checked |
| 411 | Does the model run in parallel? OK |
| 412 | Is the analytical spin-up working? OK |
| 413 | Do we have all the driver maps for the years 1600 to 2000 for European simulations? Climate, land cover change, forest management, litter raking, and N-deposition? OK |
| 414 | Do we have all the driver maps for the years 1600 to 2000 for global simulations? Still need litter raking and FM |
| 415 | Read N-deposition maps through COMP ??? |
| 416 | |
| 417 | |
| 418 | === Data availability === |
| 419 | Observational data products for model-data comparison can be found at: /home/satellites5/maignan/ORCHIDEE/EVALUATION. |
| 420 | • phenology |
| 421 | ◦ GIMMS (LAI and FPAR3g) |
| 422 | • Forest structure |
| 423 | ◦ Remote sensing products of biomass (temperate and boreal maps, i.e., Turner) |
| 424 | ◦ Biomass of EU forest from JRC (Europe) |
| 425 | ◦ Global 1-degree Maps of Forest Area, Carbon Stocks, and Biomass, 1950-2010 (ORNL DAAC) |
| 426 | ◦ Avitabile product (Global forest biomass) |
| 427 | ◦ Forest basal area (Europe) |
| 428 | ◦ Canopy height (Global) |
| 429 | • NPP |
| 430 | ◦ Site-level NPP database Luyssaert et al 2007 |
| 431 | • NEP |
| 432 | ◦ FLUXNET site-level data |
| 433 | • TER |
| 434 | ◦ FLUXNET site-level data |
| 435 | • GPP |
| 436 | ◦ FLUXNET site-level data |
| 437 | ◦ EC-based upscaled GPP, i.e., Jung |
| 438 | • NPP/GPP |
| 439 | ◦ site-level data and regional patterns, i.e., Campioli et al 2015 |
| 440 | • Soil hydrology |
| 441 | ◦ ESA CCI ECV |
| 442 | ◦ measurements from Brazil (ABRACOS product) |
| 443 | ◦ River discharge records from selected gauges (Global coverage) |
| 444 | • Albedo |
| 445 | ◦ MODIS or GlobAlbedo for albedo |
| 446 | • Energy (sensible and latent heat) |
| 447 | ◦ GLEAM for evapotranspiration |
| 448 | ◦ EC-based upscaled evapotranspiration, i.e., Jung |
| 449 | • Tree ring width (not on /home/satellites5/) |
| 450 | ◦ ITRDB |
| 451 | • LCC |
| 452 | ◦ Luyssaert et al 2014 – FLUXNET changes in Rn, LE, H, G, albedo |
| 453 | ◦ Duveiller et al 2018 – Remote sensing changes in Rn, LE, H+G, albedo |
| 454 | • NFI |
| 455 | ◦ France, Spain, Germany and Sweden (/home/satellites5/) |
| 456 | ◦ EU-wide data through the VERIFY project? |
| 457 | |
| 458 | There are several data tools (ATLAS, Mapper and Jerome’s) to help compare model outputs with observation, which we might be able to use. If we will not be using the available tools for comparison, we need to preprocess the observational data products to produce global means, time series, decadal averages of spatial patterns etc. The analyses presented in Naudts et al 2015 are a good starting point for the evaluation of ORCHIDEE-CN-CAN. The following scripts are available and have been tested and re-activated |
| 459 | |
| 460 | Table 2. Model evaluations scripts available per November 2018 |
| 461 | Description Status Path |
| 462 | Extract species-level productivity from French NFI data and compare against simulated productivity – looking to replace the French data by EU-wide data through the VERIFY project OK dofoco/dofoco/SCRIPTS/XXX |
| 463 | Extract Jung GPP and compare against spatially explicit simulations for 8 regions in Europe ??? dofoco/dofoco/ SCRIPTS/XXX |
| 464 | Extract Jung evapotranspiration and compare against spatially explicit simulations for 8 regions in Europe ??? dofoco/dofoco/ SCRIPTS/XXX |
| 465 | Extract MODIS albedo (NIR & VIS) and compare against spatially explicit simulations for 8 regions in Europe ??? dofoco/dofoco/ SCRIPTS/XXX |
| 466 | Extract height and compare against spatially explicit simulations for 8 regions in Europe ??? dofoco/dofoco/ SCRIPTS/XXX |
| 467 | Extract effective LAI and compare against spatially explicit simulations for 8 regions in Europe ??? dofoco/dofoco/ SCRIPTS/XXX |
| 468 | Extract BA and compare against spatially explicit simulations for 8 regions in Europe ??? dofoco/dofoco/ SCRIPTS/XXX |
| 469 | Extract half-hourly EC observations and compare with half-hourly site level simulations for NEP, GPP, TER, and evapotranspiration ??? ENSEMBLE/FLUXNET |
| 470 | Extract half-hourly EC observations and compare with half-hourly site level simulations for albedo ??? ENSEMBLE/FLUXNET |
| 471 | |
| 472 | |
| 473 | === Parameterization === |
| 474 | As we will compare the simulations to observations, the simulations need to be our best shot to resemble reality. Hence, all site-level simulations made for the parameterization will have the following configuration: |
| 475 | • start from a spin-up |
| 476 | • impose_cn = no (i.e., accounting for N-deposition ) |
| 477 | • Forest management = 2 (i.e., accounting for a thin and fell type of forest management) |
| 478 | |
| 479 | European simulations will have the following configuration: |
| 480 | • start from a spin-up |
| 481 | • impose_cn = no (i.e., accounting for N-deposition ) |
| 482 | • impose_veg = no (i.e., accounting for PFT distribution) |
| 483 | • Forest management from map |
| 484 | |
| 485 | The following parameterization approach – making use of parameters that were already shown to be sensitive to tuning – is proposed: |
| 486 | |
| 487 | Table 3. Proposed order, parameters and tools to parameterize ORCHIDEE-CN-CAN. Once all steps has been performed, the NPP to GPP ratio should be re-evaluated, possibly resulting in another tuning cycle for some of the PFTs. |
| 488 | Process Parameter(s) Tool |
| 489 | Onset of growing season and start of senescence thresholds for phenology FLUXNET |
| 490 | NPP/GPP ratio coeff for maintenance respiration FLUXNET + Campioli |
| 491 | Magnitude of LAI k_latosa_min and k_latosa_max NFI + Luyssaert et al 2007 |
| 492 | Magnitude of GPP LL_alpha, Vcmax, J_max FLUXNET |
| 493 | Magnitude of NPP Implicit through NPP/GPP and GPP Luyssaert et al 2007 |
| 494 | Evapotranspiration LAI_top, water stress FLUXNET |
| 495 | Water stress To be discovered FLUXNET |
| 496 | Magnitude of Rh To be checked - Rh Luyssaert et al 2007 |
| 497 | Magnitude of TER To be checked - Rh FLUXNET |
| 498 | Diameter, height, and biomass form factor NFI + Luyssaert et al 2007 |
| 499 | density and biomass self-thinning parameters NFI + Luyssaert et al 2007 |
| 500 | Harvest and biomass Self-thinning and RDI parameters NFI + ??? |
| 501 | Albedo Implicit through LAI and forest structure FLUXNET? |
| 502 | Tree ring width Self-thinning, recruitment ITRDB or VERIFY Fig 4 |
| 503 | |
| 504 | |
| 505 | ==== Settings for the FLUXNET comparison ==== |
| 506 | The parameterization starts with several 1-pixel test cases coinciding with long-term flux-net sites to test whether the model captures the growth dynamics such as phenology, max LAI, GPP, etc. These tests require a spin-up. The 1-pixel test cases will allow for both parameter tuning and changes in the code to improve the model behavior. The majority of the data represent mature forests, hence, the modelled forests should be mature as well. The model will be run for 80 years, before any output will be compared to the FLUXNET measurements. An iterative process is be planned: |
| 507 | |
| 508 | • 80 years to reach mature forest → parameterize |
| 509 | • Re-run the 80 years to reach mature forest with the new parameters → parameterize |
| 510 | • Re-run spin-up and 80 year simulation to reach mature forest with the new parameters → parameterize |
| 511 | • Continue until satisfied |
| 512 | |
| 513 | Only if we experience too many difficulties with manual tuning (if there are too many non-linearities in the model), we will use the multi-site optimization tool developed by Vlad . When the simulated growth dynamics are satisfying, 140 years long tests will be performed to check cumulative variables such as basal area, tree height, tree diameter, stand density, standing biomass, and harvest. To evaluate net ecosystem exchange of carbon and soil carbon and nitrogen pools a spin-up is required. Note that the spin-up depends on the parameters used in ORCHIDEE and that the sensitivity of parameters in ORCHIDEE depends on the spin-up. There is no easy way to break this dependency. We should avoid to ‘over-tune’ the 1-pixel FLUXNET comparisons. Instead, we will continue evaluating the model over longitudinal bands. |
| 514 | |
| 515 | |
| 516 | ==== Settings for longitudinal bands ==== |
| 517 | Test with longitudinal bands will be used to ensure that we have the expected gradients across regions (response to: climate, precipitation, management, and nitrogen). Longitudinal bands will not be (formally) compared against data, they should be considered as intermediate tests. If the expected responses are not present in the longitudinal bands, there is no reason to expect that they will be present in the European/global runs. Therefore, the issue should be investigated/fixed before launching large scale simulations. |
| 518 | |
| 519 | Three longitudinal tests will be conducted covering 2 pixels wide bands from N-Canada to Southern Chili, N-Norway to S-Africa and one band N-E Russia to Australia (including W-Australia → only place where something grows in Australia). If the evaluation is limited to checking whether response gradients are indeed present in the simulations, the configuration could be limited to running the analytical spin-up until equilibrium. Once we moved on to the longitudinal bands the focus will be on tuning of parameters. Any problems with the model and its functionality should have been caught during the 1-pixel and longitudinal tests. |
| 520 | |
| 521 | Longitudinal bands can be run largely independent from the 1-pixel tests. After concluding the 1-pixels tests it probably makes sense to run the longitudinal bands to check whether the latest parameters still produce the expected response gradients. |
| 522 | |
| 523 | |
| 524 | ==== Settings for European tests ==== |
| 525 | Once happy with the pixel and longitudinal tests we will move on to the European scale. Ultimately, the aim is to produce the control run for future simulation experiments. The model should be spin-up for 1600 by making use of the 1901-1920 climatology (because those are the coldest decades in the climate forcing). During the spin-up the forest management map and N-deposition maps for 1600 should be used. As a consequence, equilibrium soil carbon should be tested at the regional scale rather than the individual pixels. |
| 526 | |
| 527 | Subsequently, a transient spin-up will be run starting from the 1600 spin-up. During this spin-up, changes in forest management, CO2, N-deposition, litter-raking, and land cover change will be accounted for. The climate data will still have to cycle over 1901-1920. The transient simulation will run until 1750 and, pending on successful evaluation, extended until 1950 and 2015. 1750, 1950 and 2015 are common starting points for simulation experiments. |
| 528 | |
| 529 | Whether we take 20 or 30 years climate-cycles depends on the exact length of the simulation; the target is that by the time the simulation reaches the year 1901, a cycle has been completed and so the cycle in the climate forcing is synchronized with the simulation years. |
| 530 | |
| 531 | The European control should be run from the year 1601 to 2015 includes: |
| 532 | |
| 533 | • A spin-up as initial condition to make a European 1° x 1° CONTROL simulation |
| 534 | • 64 PFTs (some with 4 age classes) |
| 535 | • With forest management |
| 536 | • Litter raking |
| 537 | • Increasing atmospheric CO2 concentrations read from file |
| 538 | • LCC |
| 539 | • CRU-NCEP meteorological forcing, cyclic meteorology (1901-1920) until 1900, then use the corresponding year |
| 540 | • A dynamic N-cycle (impose_cn=n) |
| 541 | • N-deposition, we have N-deposition files from approximately 1860 until present. |
| 542 | |
| 543 | Table 4. Observational data products to evaluate European control simulation |
| 544 | Observations Tool |
| 545 | EU-wide NFI data (Fig. 3) through the VERIFY project ??? |
| 546 | Jung GPP – relies on FLUXNET thus partly circular XXX |
| 547 | Jung evapotranspiration – relies on FLUXNET thus partly circular XXX |
| 548 | MODIS albedo NIR & VIS XXX |
| 549 | Remote sensing tree height XXX |
| 550 | Extract effective LAI from MODIS XXX |
| 551 | BA from JRC – relies on NFI thus partly circular XXX |
| 552 | Biomass from JRC – relies on NFI thus partly circular ??? |
| 553 | Biomass from Turner – Remote sensing? ??? |
| 554 | Harvest from Schelhaas - data through the VERIFY project ??? |
| 555 | LCC – Luyssaert et al 2014 XXX |
| 556 | LCC – Duveiller et al 2018 ??? |
| 557 | |
| 558 | |
| 559 | ==== Settings for spin-up and re-parameterization ==== |
| 560 | As a spin-up is costly in both time and computer resources, we need a strategy to avoid wasting these resources. Thus, the spin-up will like the parameterization, be done across scale moving from pixel to global scale. The spin-up will be done in parallel with the parameterization. Often the problems with the spin-up have a technical characters and show up for the pixels with extreme climate conditions. Before launching a longitudinal, regional or global spin-up, we should agree on the model version to use, because structural changes to the code will necessitate re-running the spin-up. The model version to use, will most likely be the version ready once the parameterization at the 1-pixel level is satisfying, and no more changes to the code need be added. |
| 561 | |
| 562 | • Identify the variables that are targeted by the spin-up (such as NEP, heterotrophic respiration, decomposition etc). The spin-up will reveal whether parameters affecting these variables need to be tuned. The spin-up itself is also an interesting test case that could be loosely compared against data. |
| 563 | • We could compare the spin-ups to maps of soil carbon stocks to check the order of magnitude. Soil carbon maps should only be formally compared with the control run (spin-up + transient simulation) for the year 2000 because that run includes the simulated effects of N-deposition, management, litter raking and land cover change. The more simple configurations of the spin-up do not account for these processes or do not account for the right sequence of processes |