Example GEOtop Calibration Report¶

What follows is an example of GEOtop calibration report.

In this report the GEOtop model has been run in 1D mode solving the water and energy balance for one site DOMEF 2000 located in a mountain grassland in the Italian Alps in South Tyrol. The model was run using in input meteorological parameters for a period of about 4 years, with about 35 months of warm up and then calibrated with respect of soil moisture observations for the last 18 months. We considered here 16 layers up to 1m depth. The chosen cost function is based on the Kling-Gupta efficiency. The model was optimized for 29 parameters, which control soil, surface and vegetation properties. The chosen optimization algorithm was CMA-ES. In this case 256 CPUs were used.

The reports show the decrease of the cost function as function of the generations and the comparison of the simulated time series with the observations after the calibration, at different time scales, aggregated hourly, daily and monthly.

# Defaults
model_path = '../data/testbed/inputs'
timeout = 120
observations_path = '../data/testbed/observations/obs.csv'
parameters_path = '../data/parameters/testbed.csv'
default_parameters = {}
algorithm = 'Random'
popsize = 2
num_generations = 2
scheduler_file = None
num_cpus = 2
num_workers = 2
performance_report_filename='report.html'

# Parameters
model_path = "/scratch/scampane/MHPC-project/data/DOMES/inputs"
timeout = 150
observations_path = "/scratch/scampane/MHPC-project/data/DOMES/observations/obs.csv"
parameters_path = "/scratch/scampane/MHPC-project/data/parameters/all.csv"
algorithm = "NGO"
popsize = 128
num_generations = 32
scheduler_file = "/scratch/scampane/MHPC-project/scheduler_files/scheduler-Q61.json"
num_cpus = 256
num_workers = 32
performance_report_filename = "/scratch/scampane/MHPC-project/runs/short/DOMES-NGO-4096-256-DeW-performance-report.html"

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import scrapbook as sb
import dask.config
from dask.distributed import Client, performance_report

from mhpc_project.utils import date_parser, kge_cmp, calibrate, delta_mim
from mhpc_project.parameters import UniformSoilParameters as Parameters
from mhpc_project.models import UniformSoilModel as Model
import mhpc_project.plots as plots

# Store dask config
sb.glue('dask_config', dask.config.config)

parameters = Parameters(parameters_path, default_parameters)
model = Model(model_path, store=False, timeout=timeout)
observations = pd.read_csv(observations_path,
                           parse_dates=[0],
                           date_parser=date_parser,
                           index_col=0)

if scheduler_file:
    client = Client(scheduler_file=scheduler_file)
else:
    client = Client()
client.wait_for_workers(n_workers=num_workers, timeout=240)

with performance_report(filename=performance_report_filename):
    recommendation, predictions, log = calibrate(model,
                                                 parameters,
                                                 observations,
                                                 algorithm,
                                                 popsize,
                                                 num_generations,
                                                 client,
                                                 num_cpus)

sb.glue('loss', kge_cmp(predictions, observations))
sb.glue('log', [((x.generation, x.args[0]), l, t) for (x, l, t) in log])

(64_w,128)-aCMA-ES (mu_w=34.2,w_1=6%) in dimension 29 (seed=nan, Thu Feb 18 09:36:33 2021)

plot = plots.convergence([(x.generation, l) for x, l, _ in log if np.isfinite(l)])
plt.close(plot)
sb.glue('convergence_plot', plot, 'display')

for name, plot in plots.comparisons(predictions, observations).items():
    plt.close(plot)
    sb.glue(name + '_plot', plot, 'display')

report = delta_mim(parameters, [(x.args[0], l) for x, l, t in log if np.isfinite(l)])
report['best'] = parameters.from_instrumentation(recommendation)
sb.glue('report', report, 'pandas')

report.sort_values('delta', key=np.abs, ascending=False)

	delta	delta_conf	S1	S1_conf	best
FieldCapacity	0.196860	0.016790	0.083046	0.231401	3.870706e-02
ThetaRes	0.180005	0.014483	0.104353	0.282825	2.985520e-03
ThetaSat	0.168758	0.015870	0.097108	0.266497	5.557729e-01
DecayCoeffCanopy	0.167036	0.012630	0.087294	0.237451	1.165855e+00
NormalHydrConductivity	0.163122	0.018452	0.043401	0.147470	1.975318e-04
CanopyFraction	0.162160	0.017316	0.054883	0.169293	8.927573e-01
MinStomatalRes	0.152129	0.013413	0.043169	0.110797	2.081448e+02
AlphaVanGenuchten	0.137440	0.014014	0.056531	0.136567	2.450772e-03
NVanGenuchten	0.136678	0.017649	0.039043	0.125220	1.827232e+00
LSAI	0.127544	0.017510	0.039181	0.119751	2.178355e+00
SoilEmissiv	0.125561	0.017766	0.041374	0.119356	4.216430e-01
VegSnowBurying	0.119362	0.021109	0.026987	0.086476	3.022595e+00
ThermalCapacitySoilSolids	0.117335	0.018992	0.025219	0.095252	3.709680e+06
SoilAlbNIRWet	0.116171	0.016200	0.025704	0.095070	6.082924e-01
SoilAlbNIRDry	0.112665	0.014766	0.040809	0.094023	3.240326e-01
SoilAlbVisDry	0.111565	0.015490	0.023665	0.097089	3.549442e-01
ThermalConductivitySoilSolids	0.109694	0.018087	0.030302	0.079683	2.883452e+00
SoilAlbVisWet	0.106393	0.018053	0.026151	0.059908	2.886453e-01
VMualem	0.104845	0.019058	0.021472	0.065113	2.914712e-01
VegReflectVis	0.104551	0.013805	0.015882	0.067012	5.010492e-01
SpecificStorativity	0.102199	0.018316	0.021701	0.078875	1.643646e-07
RootDepth	0.100619	0.018432	0.023663	0.080024	5.512460e+02
CanDensSurface	0.099391	0.017148	0.027862	0.076559	4.069320e+00
WiltingPoint	0.097252	0.013059	0.032982	0.074388	7.603242e-01
SoilRoughness	0.096809	0.015322	0.030851	0.062250	1.077662e+00
VegTransVis	0.088482	0.016014	0.014867	0.052614	2.124748e-01
VegTransNIR	0.077449	0.018123	0.012616	0.032653	4.469694e-01
VegReflNIR	0.068249	0.018172	0.009401	0.027690	1.406107e-01
VegHeight	0.037858	0.014848	0.003969	0.005968	9.226614e+02

Calibration of the GEOtop model using evolutionary algorithms on supercomputers

Example GEOtop Calibration Report¶