Fixing persistent state-space bias in land surface models using SSU-LSF

When deploying Mamba-family Structured State Space Models for operational land surface forecasting, dealing with stubborn non-stationary bias can become an architectural bottleneck. If you have ever watched your NDVI, LST, and crop phenology predictions remain skewed long after an unrecorded irrigation boom, a sudden dam-operation shift, or a sensor recalibration has ended, you are likely hitting the core limitation of state-transition matrix absorption. These abrupt environmental changes get baked directly into the model's internal memory states, creating persistent forecast errors that standard fine-tuning fails to excise cleanly.
This post breaks down the mechanics of the SSU-LSF (State-Space Unlearning for Land Surface Forecasting) framework proposed by Anidipta Pal under arXiv:2610.02248, detailing how it targets these specific state matrix artifacts without requiring a full retraining cycle from scratch.

Diagnosing State-Transition Matrix Contamination

The root cause of the failure mode lies in how Mamba-based architectures handle continuous temporal dependencies. Unlike traditional recurrent neural networks or transformer architectures that might dilute past anomalies over a fixed attention window, Structured State Space Models store sequential histories inside continuous-time transition parameters. When a non-stationary confounding event occurs—such as an undocumented agricultural irrigation surge or a sudden hardware recalibration on a satellite sensor—the shift is absorbed into the state matrices.
The primary symptom in production is a sustained residual bias in downstream metrics (like Normalized Difference Vegetation Index and Land Surface Temperature) even when physical field conditions have returned to baseline. Checking for this issue involves tracking divergence between expected baseline climatology and model predictions across localized temporal windows. If the error vector fails to decay according to the system's expected eigenvalue decay rate, the transition matrices are likely retaining a confounding footprint.

How SSU-LSF Computes State-Space Unlearning

To isolate and remove these temporal artifacts without destabilizing the broader model parameters, the SSU-LSF framework introduces a specialized unlearning pipeline tailored for geoscientific SSMs. The process relies on several mathematical components working in sequence:

1. EKFac Influence Functions: The framework adapts Empirical Kronecker-factored Approximate Curvature (EKFac) influence functions specifically for Mamba state matrices. By utilizing a closed-form matrix-exponential gradient, it maps the exact influence of the confounding temporal window onto the internal transition parameters.
2. Spectral-Radius-Weighted Elbow Thresholding: Localization of the temporal confounding footprint, denoted as $\Phi$, is achieved through spectral-radius-weighted elbow thresholding. This step prevents the unlearning algorithm from stripping out valid seasonal signals by mathematically determining the boundary where the confounding event stops influencing the state dynamics.
3. Hessian-Free Projected Gradient Ascent: Once the footprint $\Phi$ is mapped, the framework executes a Hessian-free projected gradient ascent. To ensure the unlearning procedure does not catastrophically degrade performance on clean, unaffected spatial domains, this optimization is bounded within a Kullback-Leibler (KL-divergence) trust region.
4. Spatial Total-Variation Regularization: The gradient step is further augmented by spatial total-variation (TV) regularization, which preserves spatial smoothness across neighboring land surface grids and prevents checkerboard artifacts in the updated state matrices.

Evaluating Residual Confounding Bounds

A key theoretical contribution of the paper is Proposition 1, which establishes that residual confounding is strictly bounded by $\mathcal{O}\big((1-\rho(\bar{A})^{T_c})/((1-\rho(\bar{A}))\mu)\big)$.
Analyzing this bound reveals why window length matters so much in operational setups. As the confounding window length $T_c$ increases, the upper bound on residual confounding grows, making prompt unlearning critical. Waiting too long to purge an anomaly enlarges the footprint $\Phi$, making it harder for the spectral-radius-weighted thresholding to cleanly separate the artifact from valid long-term environmental trends.

Implementation Performance and Benchmarks

When integrating this unlearning workflow into an operational pipeline, computational efficiency dictates feasibility. Across three heterogeneous benchmarks and eleven baseline methods, the framework achieves notable confounding reduction rates:

• CropHarvest benchmark: $0.773$ confounding reduction rate
• NDVI-LST benchmark: $0.821$ confounding reduction rate
• ERA5 benchmark: $0.859$ confounding reduction rate

In terms of stability, the worst-case clean-domain Root Mean Square Error (RMSE) degradation on the ERA5 benchmark sits at $4.2\%$, showing that the trust region and TV regularization successfully protect unaffected regions. From a compute perspective, the unlearning requests converge rapidly in $3$ to $5$ epochs, achieving an $8.4×$ lower GPU-cost per unlearning request compared to executing a full model retraining run from scratch.
For developers managing large-scale environmental Mamba deployments, adopting this closed-form matrix-exponential gradient approach offers a viable path to correct lingering non-stationary anomalies without discarding historical training investments.