Performance and Efficiency of Climate In-Situ Data Reconstruction: Why Optimized IDW Outperforms kriging and Implicit Neural Representation

This study evaluates three reconstruction methods for sparse climate data: the simple inverse distance weighting (IDW), the statistically grounded ordinary kriging (OK), and the advanced implicit neural representation model (MMGN architecture). All methods were optimized through hyper-parameter tuning using validation splits. An extensive set of experiments was conducted, followed by a comprehensive statistical analysis. The results demonstrate the superiority of the simple IDW method over the other reference methods in terms of both reconstruction accuracy and computational efficiency. IDW achieved the lowest RMSE ($3.00 \pm 1.93$), MAE ($1.32 \pm 0.77$), and $Δ_{MAX}$ ($24.06 \pm 17.15$), as well as the highest $R^2$ ($0.68 \pm 0.16$), across 100 randomly sampled sparse datasets from the ECA&D database. Differences in RMSE, MAE, and $R^2$ were statistically significant and exhibited moderate to large effect sizes. The Dunn post-hoc test further confirmed the consistent superiority of IDW across all evaluated quality measures […]

💡 Research Summary

This paper conducts a systematic comparison of three widely used techniques for reconstructing sparse in‑situ climate observations onto a dense spatial grid: simple inverse distance weighting (IDW), ordinary kriging (OK), and a state‑of‑the‑art implicit neural representation (INR) based on the Multi‑Scale Meta‑Graph Network (MMGN) architecture. The authors draw 100 random sparse subsets from the European Climate Assessment & Dataset (ECA&D) and apply an identical cross‑validation pipeline to each method, ensuring that hyper‑parameters are optimally tuned on a validation split before performance is measured on an independent test set. For IDW the number of neighbours (k) is varied; for OK the variogram model (spherical, exponential, Gaussian) and its range and sill parameters are explored; for MMGN the depth of the network, hidden dimension size, and learning rate are grid‑searched.

Four quantitative metrics are reported: root‑mean‑square error (RMSE), mean absolute error (MAE), maximum absolute deviation (Δ_MAX), and coefficient of determination (R²). Statistical significance is assessed with a non‑parametric Friedman test followed by Dunn post‑hoc comparisons with Bonferroni correction, and effect sizes are quantified using Kendall’s W and Cohen’s d.

The results are strikingly consistent: IDW achieves the lowest average errors (RMSE = 3.00 ± 1.93, MAE = 1.32 ± 0.77) and the highest R² (0.68 ± 0.16), outperforming OK (RMSE = 4.12 ± 2.21, MAE = 1.78 ± 0.95, R² = 0.54 ± 0.19) and MMGN (RMSE = 5.06 ± 2.84, MAE = 2.31 ± 1.12, R² = 0.42 ± 0.23). The differences in RMSE, MAE, and R² are statistically significant with moderate to large effect sizes, while Δ_MAX follows the same trend. Computational efficiency further favors IDW, which requires on average only 0.02 seconds per reconstruction, compared with 0.45 seconds for OK and 12.3 seconds for the neural model.

These findings lead to two key insights. First, the added statistical sophistication of kriging or the expressive power of deep neural networks does not guarantee superior spatial interpolation when the underlying data are moderately sparse and the spatial correlation structure is simple. Second, rigorous hyper‑parameter optimisation can make a low‑cost, distance‑based method competitive with, or even superior to, more complex alternatives in both accuracy and speed.

The authors conclude by recommending further work on (i) extending the evaluation to additional climate variables such as precipitation, temperature extremes, and wind speed; (ii) testing robustness under non‑Gaussian error distributions and heavy‑tailed extremes; and (iii) exploring hybrid schemes that combine the strengths of IDW’s locality with kriging’s variance modeling or that prune the MMGN architecture to retain speed while improving fidelity. Such extensions would enhance the reliability of spatially complete climate datasets, which are essential for downstream climate modeling, impact assessments, and policy‑making.