martibosch/lcd-bias-correction

Model description

Radiation-based temperature bias correction models for low-cost temperature sensors (LCDs), trained on parallel measurements against the MeteoSwiss Zollikofen (BER) reference station (summer 2025). One Explainable Boosting Machine (EBM) model per sensor type, wrapped in a scikit-learn Pipeline that selects the optimal radiation integration window. Models are serialized with skops.

Intended uses & limitations

[More Information Needed]

Training Procedure

Each model is fitted on hourly temperature bias (ΔT = T_LCD − T_ref) as a function of accumulated shortwave radiation from the nearest MeteoSwiss AWS. The optimal radiation integration window (1–5 h) is selected per sensor by maximizing Pearson correlation with ΔT before fitting the EBM.

Hyperparameters

Click to expand

Hyperparameter	Value
memory	None
steps	[('radiation_transformer', BestScaleRadiationTransformer(radiation_col='radiation_shortwave', time_col='time', window_minutes=[60, 120, 180, 240, 300])), ('model', ExplainableBoostingRegressor())]
transform_input	None
verbose	False
radiation_transformer	BestScaleRadiationTransformer(radiation_col='radiation_shortwave', time_col='time', window_minutes=[60, 120, 180, 240, 300])
model	ExplainableBoostingRegressor()
radiation_transformer__radiation_col	radiation_shortwave
radiation_transformer__time_col	time
radiation_transformer__window_minutes	[60, 120, 180, 240, 300]
model__callback	None
model__cat_smooth	10.0
model__cyclic_progress	False
model__early_stopping_rounds	100
model__early_stopping_tolerance	1e-05
model__exclude	None
model__feature_names	None
model__feature_types	None
model__gain_scale	5.0
model__greedy_ratio	10.0
model__inner_bags	0
model__interaction_smoothing_rounds	100
model__interactions	5x
model__learning_rate	0.04
model__max_bins	1024
model__max_delta_step	0.0
model__max_interaction_bins	64
model__max_leaves	2
model__max_rounds	50000
model__min_cat_samples	10
model__min_hessian	0.0
model__min_samples_leaf	4
model__missing	separate
model__monotone_constraints	None
model__n_jobs	-2
model__objective	rmse
model__outer_bags	14
model__random_state	42
model__reg_alpha	0.0
model__reg_lambda	0.0
model__smoothing_rounds	500
model__validation_size	0.15

Model Plot

Pipeline(steps=[('radiation_transformer',BestScaleRadiationTransformer(radiation_col='radiation_shortwave',time_col='time',window_minutes=[60, 120, 180,240, 300])),('model', ExplainableBoostingRegressor())])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Evaluation Results

Device	File	Training R²	Best radiation window
Abilium	abilium.skops	0.591	60 min
Barani	barani.skops	0.100	300 min
Decentlab	decentlab.skops	0.755	120 min
Koalasense	koalasense.skops	0.362	60 min
Onset_big	onset-big.skops	0.215	300 min

How to Get Started with the Model

from huggingface_hub import hf_hub_download
from skops import io as skops_io
from meteora.bias_correction import apply_bias_correction, parse_hf_path

model_str = "martibosch/lcd-bias-correction/decentlab.skops"  # replace with your sensor
repo_id, filename = parse_hf_path(model_str)
trusted = skops_io.get_untrusted_types(
    file=hf_hub_download(repo_id, filename)
)
cor_ts_df = apply_bias_correction(
    lcd_ts_df,
    ref_rad_ts,  # pd.Series/DataFrame/xr.Dataset with shortwave radiation
    model_str,
    trusted=trusted,
)

Model Card Authors

This model card is written by following authors:

[More Information Needed]

Model Card Contact

You can contact the model card authors through following channels: [More Information Needed]

Citation

@unpublished{bosch2026swiss,
  title  = {Revisiting urban heat indices in {Switzerland} using low-cost measurement networks},
  author = {{Climatology Group of the University of Bern}},
  year   = {2026},
  note   = {Manuscript under preparation}
}