Pith. sign in

REVIEW 3 major objections 4 minor 61 references

Data fusion of complementary data sources using Machine Learning enables higher accuracy Solar Resource Maps

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fusing three solar-irradiance products with simple machine learning yields hourly Norway radiation estimates more accurate than any single product, and the paper releases a 30-year map built this way.

desk verdict A practical solar resource map for Norway with real data release; the fusion result is solid via linear regression, but the FCNN edge is likely selection-inflated and needs resampling validation. read the letter →

arxiv 2501.04381 v2 pith:X3TUMAQZ submitted 2025-01-08 physics.ao-ph

J Rabault , ML S{ae}tra , A Dobler , S Eastwood , E Berge This is my paper · ORCID
classification physics.ao-ph
keywords solarresourcemappingglobalhorizontalirradiancedatafusionmachinelearninglinearregressionneuralnetworkreanalysissatellite
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that a straightforward supervised machine-learning fusion of three publicly available solar and atmospheric products—two numerical reanalyses and a satellite-derived irradiance record—can estimate hourly Global Horizontal Irradiance (GHI) over Norway more accurately than any of the three products alone. On eight held-out validation stations, a fully connected neural network lowers mean absolute error from 59 W/m2 for the best single product to 46 W/m2, with a linear regression close behind. The trained models are then applied to generate a 30-year (1991–2020), hourly, roughly 3.3 km resolution GHI map for continental Norway, released as an open data product. This matters because solar-resource planning needs locally accurate irradiance estimates, especially at high latitudes where global satellite-based atlases stop.

What carries the argument

The carrying object is the trained fusion model: a linear regression (ordinary least squares on all predictors) and a fully connected neural network with 4 hidden layers of 60 neurons, ReLU activations, mean-absolute-error loss, and about 13,000 trainable parameters, fed by roughly 30 variables from ERA5 and NORA3 plus the three closest half-hourly SARAH-3 values. The linear model performs the BLUE combination of inputs with distinct error sources; the network adds a modest nonlinear correction, which the paper interprets as compensating residual nonlinear biases such as snow, ice, and fog effects in the satellite product. A fallback hierarchy chooses the best available model at each grid point and hour, using the FCNN only when all inputs are within training ranges and GHI is at least 50 W/m2.

What would settle it

Compare the published map's hourly GHI estimates against independent quality-controlled pyranometer observations from before 2016 (or from stations not used in training); if mean absolute error clearly exceeds the 46 W/m2 validation level, or the bias drifts with year, the stationarity assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that fusing complementary input products with simple supervised models yields a GHI estimate that beats every input individually, and that the gain follows from Best Linear Unbiased Estimator logic: products whose error sources are not fully correlated can be combined to reduce error. Using pyranometer measurements from 47 quality-controlled stations (2016–2020) as ground truth and holding out 8 stations for validation, the paper trains a linear regression and a small fully connected neural network on ERA5 and NORA3 model fields plus SARAH-3 satellite irradiance. On validation stations the FCNN reaches MAE 46 W/m2 versus 59 W/m2 for SARAH-3, 72 W/m2 for ERA5, and 78 W/m2 for NORA3, and its intrinsic variability stays close to observed variability. The same models, with a fallback hierarchy when satellite data are unavailable north of 65°N or inputs fall outside training ranges, are used to produce the 30-year hourly map.

Load-bearing premise

The mapping from the input products to true irradiance, learned on 2016–2020 pyranometer data, stays valid for the entire 1991–2020 map, with no independent historical ground-truth check.

Editorial extensions

If this is right

  • If the claim holds, users get 30-year, hourly, roughly 3.3 km GHI fields over continental Norway that are more accurate than ERA5, NORA3, or SARAH-3 alone, including north of 65°N where satellite data are absent.
  • The fused estimates retain realistic variability, so downstream solar-yield estimates will not be artificially smoothed the way raw ERA5 irradiance is.
  • The same training recipe should transfer to other regions wherever quality-controlled pyranometer networks and at least two complementary gridded products exist, with region-specific retraining.
  • Because the fallback hierarchy uses simpler models when inputs are missing or out of range, the map remains complete in time and space even when one product fails.
  • The open data release enables independent checks of solar-resource planning and climate studies in Norway.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's validation is confined to 2016–2020; the strongest implicit assumption is that the predictor-to-truth mapping is stationary, so a direct test against independent ground truth from 1991–2015 would sharpen confidence.
  • The FCNN's nonlinear gain over linear regression is small (46 versus 54 W/m2), suggesting the dominant benefit is linear blending; a testable corollary is that richer architectures or new predictors yield further gains only if they add non-redundant error structure.
  • The method naturally extends to other renewable-resource variables, such as direct or diffuse irradiance or wind, whenever multiple reanalyses or observational products with partially independent error structures cover the same period.
  • The deliberate exclusion of latitude, longitude, and altitude from predictors is an anti-overfitting choice; in regions with denser station networks, including such coordinates may become safe and could further reduce local error.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper describes a machine-learning data-fusion methodology for estimating global horizontal irradiance (GHI) over Norway. Hourly predictors from ERA5, NORA3, and the SARAH-3 satellite product are combined with pyranometer ground truth from 47 quality-controlled stations (8 held out for validation). Both a linear regression (LR) and a fully connected neural network (FCNN) are trained; on the validation stations the fused models report lower MAE than any individual product (e.g., 54 and 46 W/m2 for LR and FCNN versus 59 W/m2 for SARAH-3). The trained models are then used to produce a 30-year (1991–2020), hourly, 0.03-degree GHI map over Norway, which is released openly.

Significance. If the validation results are unbiased, the paper delivers a practical, transferable fusion method and a valuable open solar-resource dataset for a region poorly covered by global satellite products. The strengths are the reproducibility (open code and data), the quality-controlled pyranometer dataset, and the use of a simple BLUE-style linear model, which is robust because OLS has no tuned hyperparameters. The LR-vs-individual-products improvement is credible. The main weakness is that the FCNN's reported advantage is potentially inflated by hyperparameter selection on the same validation set and by the small number of validation stations.

major comments (3)
  1. [Section 3.2 and Table 3] The FCNN architecture, optimizer, loss, and metaparameters were chosen as those giving the best validation loss (Section 3.2), and the same validation stations are then used to report the headline 46 W/m2 MAE in Table 3. This is a post-selection estimate: it reflects the best of the configurations tried, not an unbiased estimate of generalization performance. The difference between the FCNN (46) and LR (54) MAE is the most vulnerable comparison, since LR has no hyperparameter search. Please report performance from a nested cross-validation or a separate test set that is not used for any model or hyperparameter choice, and give confidence intervals for the metrics.
  2. [Section 2.3 / Table 3] The validation set has only 8 stations, of which 5 are used for the models that include SARAH-3, and hourly GHI is strongly autocorrelated, so the 78k-sample metric is not equivalent to 78k independent samples. No per-station MAE/RMSE, error bars, or significance tests are given. Station-level bootstrap or paired tests are needed to establish that the FCNN-LR difference and, more importantly, the fused-vs-single-product differences are not within station-to-station noise. Also, the 8 validation stations are selected subjectively; leave-one-station-out or repeated random splits would assess sensitivity to that choice.
  3. [Section 4.2 / requirements R1-R2] The 30-year map is produced by applying models trained on 2016-2020 to the full 1991-2020 period without any historical ground-truth validation. The paper acknowledges the stationarity assumption (R2) but does not test it. If any early-period or pre-2016 pyranometer data exist, they should be used as a temporal holdout; otherwise the map's accuracy claims for 1991-2020 are unsupported. At minimum, discuss the risk quantitatively (e.g., how stable the predictor-target relationship is expected to be given the reanalysis and SARAH-3 processing changes).
minor comments (4)
  1. [Section 3.3] In the sentence 'since we set a lowest value cutoff of 50 W/m2 to filter the training data to avoid class imbalance due to the night, as highlighter higher up', 'highlighter' should be 'highlighted'.
  2. [Abstract and Section 4.1] The word 'significantly' is used without a statistical test. Please either add significance tests or rephrase to avoid implying formal significance.
  3. [Figure 3] The Taylor plot in the left panel shows training-station performance; consider adding a note about the risk of interpreting training-set clustering as evidence of generalization, even though the validation panel is reassuring.
  4. [Table 2] The model_index ordering (6 best, 0 worst) is counterintuitive; a sentence in the caption clarifying that higher index means better would help readers.

Circularity Check

2 steps flagged · score 6.0 of 10

FCNN headline accuracy is a post-selection validation minimum; the data-fusion claim retains independent LR support, so circularity is partial.

  1. fitted input called prediction [Section 3.2 (Neural Network); Table 3]
    "We have tested several different FCNN architectures, optimizers, loss functions, and metaparameters in the early phases of the training process. We found that a simple FCNN with 4 layers of 60 neurons each, using the rectified linear unit (ReLu) activation function, the mean absolute error loss, and the Adagrad optimizer, gives the best validation loss over the present task."

    The FCNN configuration is selected as the one minimizing the validation loss on the 8 validation stations, and the chosen loss is the mean absolute error. Table 3 then reports the FCNN MAE of 46 W/m2 on those same validation stations, and the abstract and conclusion present this as the model's accuracy. The reported 46 W/m2 is therefore the minimum over the tried configurations on the evaluation set, i.e., a post-selection statistic offered as an independent accuracy estimate. The headline 46 vs 59 W/m2 gap over SARAH-3 is partly an artifact of having selected the model on the very same validation comparison, rather than an unbiased out-of-sample prediction.

  2. fitted input called prediction [Section 3.3 (GHI map production); Table 2]
    "we select the best (i.e., the one with the highest model_index following the validation presented below) that can be successfully produced as the corresponding best estimate for the GHI."

    The model_index hierarchy used to assemble the final 30-year map is justified by 'the validation presented below', which is the same 8-station validation set used to compute Table 3. The final product's composition (FCNN where available, otherwise LR, then single products) is therefore selected on the same validation-station metrics that are later used to display the product's performance in Fig. 7 and Table 3. This makes the 'best estimate' label of the map a validation-selected outcome rather than an independently confirmed prediction. The map itself remains a deployment of fitted models, so this is partial rather than total circularity.

full rationale

The paper's main data-fusion result is not circular in the ordinary sense: the LR and FCNN models are trained on pyranometer ground truth at stations disjoint from the 8 validation stations, and comparing LR/FCNN against ERA5, NORA3, and SARAH-3 on those held-out stations is a genuine out-of-sample evaluation. The 30-year map is a straightforward deployment of the fitted models, and the stationarity assumption (R1/R2 in Section 2) is an external-validity concern, not a circularity. The self-citations to the authors' own quality-control report and dataset [30,31] are not load-bearing circularity because the dataset and QC scripts are openly released and can be checked independently. However, one significant step does reduce to a fit: the FCNN architecture, optimizer, loss function, and metaparameters were chosen as the configuration with the lowest validation loss on the same 8 validation stations that are then used to report the headline MAE (46 W/m2 vs 59 W/m2 for SARAH-3). That value is the minimum of the searched configurations on the evaluation set, so the FCNN-specific superiority claim is a post-selection estimate. The same validation set is used to rank the model_index hierarchy that assembles the final map, and the product is then re-compared at those same stations. The LR-based improvement (54 vs 59 on the same samples) is less affected because ordinary least squares has no meaningful hyperparameter search, so the central claim that data fusion beats single products retains independent content; only the FCNN 'best model' and map-composition claims are partially circular. Additional concerns, such as the small number of validation stations, strong temporal autocorrelation of hourly GHI, and the absence of confidence intervals, are statistical robustness issues rather than circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No invented physical entities are introduced. The load-bearing choices are the hand-picked validation split, the night-filter threshold, and the FCNN hyperparameters, plus the stationarity assumptions that allow a 30-year map to be generated from a 5-year training window.

free parameters (4)
  • Pmin night-filter threshold = 50 W/m2
    Chosen to exclude night and very low light samples to avoid class imbalance in FCNN training; affects which samples are used.
  • FCNN architecture and hyperparameters = 4x60 neurons, ReLU, MAE loss, Adagrad, initial LR 2e-4, ~2000 epochs
    Selected by trial and error based on validation loss; this model selection on the validation set may bias reported performance.
  • Train/validation station split = 39 training / 8 validation stations (subjective)
    Split chosen to represent Norway's range of conditions; not random, and no independent test set is held out.
  • Predictor variable set = Listed in Table 1
    Chosen based on expert discussions rather than automated feature selection; the set affects model capacity.
assumptions (5)
  • domain assumption Pyranometer measurements are accurate ground truth for GHI after quality control.
    The training and validation target is the QC'd pyranometer data; any systematic error in these measurements propagates to the models.
  • domain assumption Input product biases are stationary over 1991-2020.
    Stated in Section 2.1 as requirement R2; the ML models are trained on 2016-2020 data and applied to the full 30-year period.
  • domain assumption The predictor-target relationship learned on 2016-2020 generalizes to 1991-2020.
    No historical ground truth is used to validate the map before 2016; the paper relies on temporal stationarity.
  • domain assumption Nearest-neighbor interpolation to station locations and grid points introduces negligible error.
    Used to build the training set and to generate the final grid; the paper does not quantify interpolation errors.
  • domain assumption The 8 validation stations are representative of all of Norway, including the data gap between 64N and 67N.
    The paper acknowledges the gap and assumes the models generalize across it because similar conditions are seen elsewhere.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data fusion of complementary data sources using Machine Learning enables higher accuracy Solar Resource Maps." pith.science (2026). https://pith.science/paper/X3TUMAQZ

@misc{pith2026250104381,
  author       = {Pith},
  title        = {Pith review of: Data fusion of complementary data sources using Machine Learning enables higher accuracy Solar Resource Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X3TUMAQZ}},
  note         = {Machine review of arXiv:2501.04381}
}
read the original abstract

In the present work, we collect solar irradiance and atmospheric condition data from several products, obtained from both numerical models (ERA5 and NORA3) and satellite observations (CMSAF-SARAH3). We then train simple supervised Machine Learning (ML) data fusion models, using these products as predictors and direct in-situ Global Horizontal Irradiance (GHI) measurements over Norway as ground-truth. We show that combining these products by applying our trained ML models provides a GHI estimate that is significantly more accurate than that obtained from any product taken individually. Using the trained models, we generate a 30-year ML-corrected map of GHI over Norway, which we release as a new open data product. Our ML-based data fusion methodology could be applied, after suitable training and input data selection, to any geographic area on Earth.

Figures

Figures reproduced from arXiv: 2501.04381 by the authors.

Figure 1
Figure 1. Exported Global Solar Atlas map (https://globalsolaratlas.info, recovered 2025-01-09). This provides a global overview of the average daily and yearly-accumulated Global Horizontal Irradiance (GHI), in kWh/m2 . Note that data are typically available only within the latitude range [65S; 65N], owing to the use of geostationary satellite data to generate the product. 2/21 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Map of the in-situ pyranometer stations selected to provide the ground truth data (red dots) used to perform model training and stations for validation (black stars). 6/21 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Taylor plots summarizing the performance of the different predictors and models. The acronyms used correspond to [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Estimate of the linearized importance of the predictors used by the LR and FCNN models. This shows (in a linearized sense for the FCNN), how much weight is given to the different input predictors. Top: individual overview for all predictors. This can be misleading, wit…
Figure 5
Figure 5. Figure 5: Illustration of the final GHI product. Snapshot for one point in time (2019-07-05 13:00 UTC) for the instantaneous GHI (W/m2 ) prediction over Norway (left panel) and the associated model_index values (right panel). 5 Conclusion and future work In the present work, we …
Figure 6
Figure 6. Figure 6: Average of the yearly accumulated GHI over Norway (kWh/m2 ) for the period 1991-2020. and the maintenance and calibration of the pyranometers operating in Norway has received increased attention following the growing interest in evaluating possible solar resources. As …
Figure 7
Figure 7. Figure 7: Comparison at the validation stations between the GHI estimates from our final ML product, other products (ERA5, NORA3, and SARAH-3 geosatellite data when these are available), and in-situ pyranometer observations. Our GHI product compares generally well with in-situ o…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 57 canonical work pages

  1. [1]

    Gueymard, C. A. Clear-sky irradiance predictions for solar resource mapping and large-scale applications: Improved validation methodology and detailed performance analysis of 18 broadband radiative models. Sol. Energy 86, 2145–2169 (2012)

    work page 2012

  2. [2]

    Polo, J. et al. Solar resources and power potential mapping in vietnam using satellite-derived and gis-based information. Energy conversion management 98, 348–358 (2015)

    work page 2015

  3. [3]

    & Alvarado, R

    Wegertseder, P., Lund, P., Mikkola, J. & Alvarado, R. G. Combining solar resource mapping and energy system integration methods for realistic valuation of urban solar energy potential. Sol. Energy 135, 325–336 (2016)

    work page 2016

  4. [4]

    & Bandoc, G

    Pr˘av˘alie, R., Patriche, C. & Bandoc, G. Spatial assessment of solar energy potential at global scale. a geographical approach. J. Clean. Prod. 209, 692–721 (2019)

    work page 2019

  5. [5]

    C., Hinkley, J., Burmester, D

    Brent, A. C., Hinkley, J., Burmester, D. & Rayudu, R. Solar atlas of new zealand from satellite imagery. J. Royal Soc. New Zealand 50, 572–583 (2020)

    work page 2020

  6. [6]

    Global solar atlas 2.0 technical report

    Suri, M. Global solar atlas 2.0 technical report. Washington, DC: World Bank (2019)

    work page 2019

  7. [7]

    P., e Silva, F

    Castillo, C. P., e Silva, F. B. & Lavalle, C. An assessment of the regional potential for solar power generation in eu-28. Energy policy 88, 86–99 (2016)

    work page 2016

  8. [8]

    & Šúri, M

    Perez, R., Cebecauer, T. & Šúri, M. Semi-empirical satellite models. Sol. energy forecasting resource assessment 21–48 (2013)

    work page 2013

Show all 61 references
  1. [9]

    & Suri, M

    Cebecauer, T. & Suri, M. Site-adaptation of satellite-based dni and ghi time series: Overview and solargis approach. AIP Conf. Proc. 1734 (2016)

    2016

  2. [10]

    M., Polo, J., Martín, L

    Fernández-Peruchena, C. M., Polo, J., Martín, L. & Mazorra, L. Site-adaptation of modeled solar radiation data: The siteadapt procedure. Remote. Sens. 12, 2127 (2020)

    2020

  3. [11]

    & Gueymard, C

    Yang, D. & Gueymard, C. A. Probabilistic post-processing of gridded atmospheric variables and its application to site adaptation of shortwave solar radiation. Sol. Energy 225, 427–443 (2021)

    2021

  4. [12]

    F., Bressan, M

    Narvaez, G., Giraldo, L. F., Bressan, M. & Pantoja, A. Machine learning for site-adaptation and solar radiation forecasting. Renew. Energy 167, 333–342 (2021)

    2021

  5. [13]

    Ruiz-Munoz, J. F. & Hoyos-Gómez, L. S. Accurate solar radiation site adaptation: Harnessing satellite data and in situ measurements. J. Renew. Sustain. Energy 16 (2024)

    2024

  6. [14]

    & Campana, P

    Zainali, S., Yang, D., Landelius, T. & Campana, P. E. Site adaptation with machine learning for a northern europe gridded global solar irradiance product. Energy AI 15, 100331 (2024)

    2024

  7. [15]

    Müller, M. et al. Arome-metcoop: A nordic convective-scale operational weather prediction model. Weather. F orecast.32, 609–627 (2017)

    2017

  8. [16]

    Hersbach, H. et al. The ERA5 global reanalysis. Q. J. Royal Meteorol. Soc. 146, 1999–2049 (2020)

    2020

  9. [17]

    Haakenstad, H. et al. NORA3: A nonhydrostatic high-resolution hindcast of the North Sea, the Norwegian Sea, and the Barents Sea. J. Appl. Meteorol. Climatol. 60, 1443–1464 (2021)

    2021

  10. [18]

    Eaton, B. et al. Netcdf climate and forecast (cf) metadata conventions. URL: http://cfconventions. org/Data/cf- conventions/cf-conventions-1.8/cf-conventions. pdf (2003). 17/21

    2003

  11. [19]

    & Milagro-Pérez, M

    Aschbacher, J. & Milagro-Pérez, M. P. The european earth monitoring (gmes) programme: Status and perspectives.Remote. Sens. Environ. 120, 3–8 (2012)

    2012

  12. [20]

    & Festa, R

    Noia, M., Ratto, C. & Festa, R. Solar irradiance estimation from geostationary satellite data: I. statistical models. Sol. Energy 51, 449–456 (1993)

    1993

  13. [21]

    Huang, G. et al. Estimating surface solar irradiance from satellites: Past, present, and future perspectives. Remote. Sens. Environ. 233, 111371 (2019)

    2019

  14. [22]

    Pfeifroth, U. et al. Surface Radiation Data Set - Heliosat (SARAH) - Edition 3, 10.5676/EUM_SAF_CM/SARAH/V003 (2023)

    2023 doi

  15. [23]

    Trentmann, J. Algorithm theoretical baseline document, meteosat solar surface radiation and effective cloud albedo climate data records - heliosat, sarah-3, the magicsol method applied for the generation of sarah-3. https://www.cmsaf.eu/SharedD ocs/Literatur/document/2023/saf_...

    2023

  16. [24]

    & Trentmann, J

    Posselt, R., Mueller, R., Stöckli, R. & Trentmann, J. Remote sensing of solar surface radiation for climate monitoring—the cm-saf retrieval in international comparison. Remote. Sens. Environ. 118, 186–198 (2012)

    2012

  17. [25]

    Skartveit, A., Olseth, J. A. & Tuft, M. E. An hourly diffuse fraction model with correction for variability and surface albedo. Sol. Energy 63, 173–183 (1998)

    1998

  18. [26]

    M., Stackhouse Jr, P

    Hinkelman, L. M., Stackhouse Jr, P. W., Wielicki, B. A., Zhang, T. & Wilson, S. R. Surface insolation trends from satellite and ground measurements: Comparisons and challenges. J. Geophys. Res. Atmospheres 114 (2009)

    2009

  19. [27]

    & Philipona, R

    Dürr, B., Zelenka, A., Müller, R. & Philipona, R. Verification of cm-saf and meteoswiss satellite based retrievals of surface shortwave irradiance over the alpine region. Int. J. Remote. Sens. 31, 4179–4198 (2010)

    2010

  20. [28]

    Castelli, M. et al. The heliomont method for assessing solar irradiance over complex terrain: Validation and improvements. Remote. sensing environment 152, 603–613 (2014)

    2014

  21. [29]

    & Pfeifroth, U

    Müller, R. & Pfeifroth, U. Remote sensing of solar surface radiation–a reflection of concepts, applications and input data based on experience with the effective cloud albedo. Atmospheric Meas. Tech. 15, 1537–1561 (2022)

    2022

  22. [30]

    O., Grini, S

    Dobler, A., Berge, E., Hygen, H. O., Grini, S. & Nygard Riise, H. Quality control of radiation data for solar resource mapping. MET Norway Report (2024)

    2024

  23. [31]

    & Berge, E

    Dobler, A. & Berge, E. Quality controlled observations of hourly incoming shortwave radiation data at the surface for solar resource mapping in norway (2016-2020), 10.5281/ZENODO.11029185 (2024)

    2024 doi

  24. [32]

    & Tsai, C.-L

    Su, X., Yan, X. & Tsai, C.-L. Linear regression. Wiley Interdiscip. Rev. Comput. Stat. 4, 275–294 (2012)

    2012

  25. [33]

    & Taylor, J

    James, G., Witten, D., Hastie, T., Tibshirani, R. & Taylor, J. Linear regression. In An introduction to statistical learning: With applications in python, 69–134 (Springer, 2023)

    2023

  26. [34]

    Applied linear regression (2005)

    Weisberg, S. Applied linear regression (2005)

    2005

  27. [35]

    Moser, W. R. Linear models: A mean model approach (Elsevier, 1996)

    1996

  28. [36]

    Scikit-learn: Machine learning in python

    Pedregosa, F. Scikit-learn: Machine learning in python. J. machine learning research 12, 2825 (2011)

    2011

  29. [37]

    On ordinary ridge regression in generalized linear models

    Segerstedt, B. On ordinary ridge regression in generalized linear models. Commun. Stat. Methods 21, 2227–2246 (1992)

    1992

  30. [38]

    McDonald, G. C. Ridge regression. Wiley Interdiscip. Rev. Comput. Stat. 1, 93–100 (2009)

    2009

  31. [39]

    & White, H

    Hornik, K., Stinchcombe, M. & White, H. Multilayer feedforward networks are universal approximators. Neural networks 2, 359–366 (1989)

    1989

  32. [40]

    Approximation by superpositions of a sigmoidal function

    Cybenko, G. Approximation by superpositions of a sigmoidal function. Math. control, signals systems 2, 303–314 (1989)

    1989

  33. [41]

    & Hinton, G

    LeCun, Y ., Bengio, Y . & Hinton, G. Deep learning.nature 521, 436–444 (2015)

    2015

  34. [42]

    Residual attention network for image classification

    Wang, F.et al. Residual attention network for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3156–3164 (2017)

    2017

  35. [43]

    Hurtik, P. et al. Poly-yolo: higher speed, more precise detection and instance segmentation for yolov3. Neural Comput. Appl. 34, 8275–8290 (2022)

    2022

  36. [44]

    Zhao, W. X. et al. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)

    2023 arXiv

  37. [45]

    M., Weinhardt, C., van der Aalst, W

    Teubner, T., Flath, C. M., Weinhardt, C., van der Aalst, W. & Hinz, O. Welcome to the era of chatgpt et al. the prospects of large language models. Bus. & Inf. Syst. Eng. 65, 95–101 (2023). 18/21

    2023

  38. [46]

    & Ganguli, S

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N. & Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning , 2256–2265 (PMLR, 2015)

    2015

  39. [47]

    Yang, L. et al. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv. 56, 1–39 (2023)

    2023

  40. [48]

    Silver, D. et al. Mastering the game of go with deep neural networks and tree search. nature 529, 484–489 (2016)

    2016

  41. [49]

    Bi, K. et al. Accurate medium-range global weather forecasting with 3d neural networks. Nature 619, 533–538 (2023)

    2023

  42. [50]

    Lam, R. et al. Graphcast: Learning skillful medium-range global weather forecasting. arXiv preprint arXiv:2212.12794 (2022)

    2022 arXiv

  43. [51]

    Poulton, M. M. Computational neural networks for geophysical data processing (Elsevier, 2001)

    2001

  44. [52]

    Krasnopolsky, V . M. & Schiller, H. Some neural network applications in environmental sciences. part i: forward and inverse problems in geophysical remote measurements. Neural Networks 16, 321–334 (2003)

    2003

  45. [53]

    Bias correction of operational storm surge forecasts using neural networks

    Tedesco, P.et al. Bias correction of operational storm surge forecasts using neural networks. Ocean. Model. 188, 102334 (2024)

    2024

  46. [54]

    Guo, W. et al. Information fusion for gnss-r wind speed retrieval using statistically modified convolutional neural network. Remote. Sens. Environ. 272, 112934 (2022)

    2022

  47. [55]

    Abiodun, O. I. et al. State-of-the-art in artificial neural network applications: A survey. Heliyon 4 (2018)

    2018

  48. [56]

    & Liu, X.-Y

    Zhou, Z.-H. & Liu, X.-Y . Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on knowledge data engineering 18, 63–77 (2005)

    2005

  49. [57]

    Abadi, M. et al. TensorFlow: Large-scale machine learning on heterogeneous systems (2015). Software available from tensorflow.org

    2015

  50. [58]

    Chollet, F. et al. Keras. https://keras.io (2015)

    2015

  51. [59]

    & Berthouze, N

    Alqaraawi, A., Schuessler, M., Weiß, P., Costanza, E. & Berthouze, N. Evaluating saliency map explanations for convolutional neural networks: a user study. In Proceedings of the 25th international conference on intelligent user interfaces, 275–285 (2020)

    2020

  52. [60]

    Selvaraju, R. R. et al. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision , 618–626 (2017)

    2017

  53. [61]

    Hsu, C.-Y . & Li, W. Explainable geoai: can saliency maps help interpret artificial intelligence’s learning process? an empirical study on natural feature detection. Int. J. Geogr. Inf. Sci. 37, 963–987 (2023). 19/21 (the figure continues on next page) 20/21 Figure 7. Comparis...

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.