Uncertainty-Aware Spatial Mixture Classification of Imbalanced CO₂ Plume States: Application to the Sleipner 2019 Benchmark

Technical Summary: Uncertainty-Aware Spatial Mixture Classification of Imbalanced CO₂ Plume States

Problem Statement
Monitoring geological carbon dioxide (CO₂) storage requires reliable classification of plume states (e.g., outside the plume, plume boundary, plume interior) across spatially heterogeneous reservoirs. Benchmark datasets, such as the Sleipner 2019 model, present significant statistical challenges:

1. Severe Class Imbalance: The vast majority of grid bins (approx. 87% in the study) represent the "outside plume" background class, while scientifically critical boundary and interior states are rare minorities.
2. Spatial and Stratigraphic Heterogeneity: The relationship between reservoir predictors (depth, velocity, petrophysical properties) and plume state varies across geological layers and spatial regimes. A single global classifier fails to capture these local variations.
3. Collinearity and Instability: Reservoir predictors are often highly correlated due to shared geological processes. In mixture models, where data is partitioned into latent components, the effective sample size for each component can be small, leading to ill-conditioned local estimates and unstable multinomial logit coefficients.
4. Lack of Uncertainty Quantification: Standard predictive machine learning approaches often provide point estimates without posterior uncertainty or interpretable latent regime identification.

Methodology
The authors propose a Spatially Informed Cluster-Weighted Multinomial Logit Mixture model. This framework integrates three distinct methodological elements:

1. Cluster-Weighted Modeling (CWM): Unlike standard finite mixtures that model only the marginal distribution of covariates, the CWM factorizes the joint distribution of covariates ($X$) and the response ($Y$). The model assumes $G$ latent statistical regimes, where each regime $g$ has:

   • A component-specific covariate density $f_g(s_i, x_i, c_i)$, which includes spatial coordinates (easting, northing, depth) to ensure regimes occupy distinct spatial regions.
   • A component-specific conditional response model $q_g(y_i | x_i, c_i)$, implemented as a multinomial logit regression.
   • Constant mixing weights $\pi_g$ across the domain (spatial heterogeneity is driven by the covariate densities, not the weights).

2. Class-Weighted Estimation: To address severe class imbalance, the conditional response part of the likelihood is fitted using class weights derived from an inverse-power transform of class frequencies. This shifts the optimization target from overall accuracy to balanced recovery of minority classes.

3. Two-Parameter Shrinkage Update: To stabilize the Expectation-Maximization (EM) algorithm's M-step under collinearity and small effective sample sizes, the authors replace the standard Maximum Likelihood Estimation (MLE) update for multinomial logit coefficients with a component-specific two-parameter shrinkage update:
   $$\hat{\beta}_{g, TPE} = (A_g + I_r)^{-1}(A_g + d_g I_r)(A_g + k_g I_r)^{-1} b_g$$
   where $A_g$ and $b_g$ are the working information matrix and vector, and $k_g, d_g$ are shrinkage parameters selected via local pseudo-BIC or cross-validation. This update generalizes Ridge ($d_g=1$) and Liu-type estimators, allowing the degree of regularization to adapt to the conditioning of each latent regime.

Key Contributions
The paper positions itself as an applied methodological contribution that integrates established techniques rather than proposing entirely new statistical theories. The specific contributions are:

• Integration for Reservoir Monitoring: Combining spatially informed CWM, multinomial logit responses, and two-parameter shrinkage specifically for imbalanced, binned reservoir data.
• Component-Specific Stabilization: Embedding a two-parameter shrinkage update within the EM algorithm to handle ill-conditioned local fits caused by collinearity and rare classes within specific latent regimes.
• Rigorous Validation Protocol: Distinguishing between apparent in-sample performance and honest predictive performance using spatial-block cross-validation (to handle spatial autocorrelation) and leave-one-layer-out transfer tests. The authors explicitly report "pseudo-BIC" and "pseudo-ICL" when class weights are used, acknowledging the deviation from standard likelihood theory.
• Ablation Analysis: Systematically isolating the contributions of class weighting, spatial covariates, latent mixtures, and shrinkage to determine the drivers of performance.

Empirical Results (Sleipner 2019 Benchmark)
The model was applied to a 200-metre binned table containing 3,409 bins across 9 layers.

• Latent Regimes: The model identified three interpretable, depth-organized latent regimes. One regime dominated the deeper reservoir stack, while the other two partitioned the shallower layers, consistent with known Utsira stratigraphy.
• Conditioning: All three regimes exhibited severe ill-conditioning (condition numbers $\sim 10^9$). The shrinkage parameters adapted accordingly, applying strong shrinkage to the most ill-conditioned regime and near-MLE updates to better-conditioned ones.
• Performance Metrics:
  • Baseline: A majority-class baseline achieved a balanced accuracy of 0.333. An unweighted global multinomial logit achieved 0.343.
  • Proposed Model: The proposed spatial CWM with two-parameter shrinkage achieved a cross-validated balanced accuracy of 0.561 (SD 0.093).
  • Ablation Findings: The improvement over the baseline was driven primarily by class weighting (raising global MLE from 0.343 to ~0.548) and the inclusion of spatial covariates (non-spatial mixture dropped to 0.422).
  • Marginal Gains: The addition of the latent mixture structure and the two-parameter shrinkage provided small, statistically indistinguishable gains over a class-weighted global multinomial logit (0.548–0.550) or a ridge-regularized mixture. In leave-one-layer-out tests, the unregularized mixture even slightly outperformed the shrinkage variant.

Significance and Claims
The authors are notably modest in their claims, explicitly stating that the proposed framework does not significantly outperform simpler class-weighted alternatives in terms of raw predictive accuracy on this specific benchmark.

• Primary Value: The significance of the framework lies in interpretability and uncertainty quantification. The model provides:
  • Identification of depth-organized latent statistical regimes.
  • Posterior uncertainty maps for every bin, highlighting transition zones between regimes.
  • Adaptive, regime-specific stabilization of local coefficients, which is crucial for robust inference in collinear settings even if it does not boost aggregate accuracy scores.
• Methodological Lessons: The paper emphasizes that for imbalanced spatial data, "honest" evaluation (spatial-block CV) is critical, as in-sample performance (0.732 balanced accuracy) is overly optimistic. It also highlights that correcting the objective function (via class weighting) is often more impactful than increasing model complexity.
• Conclusion: The work offers an "interpretable and uncertainty-aware approach to minority plume-state recovery under severe imbalance," rather than a claim that two-parameter regularization fundamentally improves classification accuracy over simpler baselines. The stabilization mechanism is presented as a tool for reliable local inference in ill-conditioned regimes.