Rule-based explanations of extreme-event AI models matched the models’ internal feature activations at about 93% accuracy on wildfire data and 85% on data used to detect tropical cyclones and atmospheric rivers, according to an arXiv preprint. That score measures agreement with the model’s own features, not whether the rules identify the real-world causes of those events.
GeoTopK feeds geographic coordinates into a location encoder before a k-sparse autoencoder, a model that keeps only a limited number of features active at once. SAE-Xplainers then use rule-based surrogate models—if-then descriptions trained at several complexity levels—to predict those activations.
A geographic signal is built into the model
The tests covered fire prediction, plus detection of tropical cyclones and atmospheric rivers. SeasFire supplied 44k samples across 17.7M patches, 98% of which were no-fire patches; ClimateNet supplied 459 samples across 25.4M patches, 94% of which were no-event patches.
In the reconstruction comparisons, location-aware GeoTopK generally scored higher than TopK for the ClimaX model: 0.937 versus 0.927 on SeasFire and 0.953 versus 0.943 on ClimateNet. Dead-feature rates also fell, from 0.159 to 0.002 and from 0.513 to 0.391, respectively.
The pattern was not universal. For the ViT model on SeasFire, the event-conditioned reconstruction score was lower with GeoTopK, 0.633 versus 0.651 for TopK.
The rules tracked some familiar fire patterns
At low rule complexity, integrating spatial context was associated with a 13% SAE-Xplainer accuracy gain on ClimateNet and a 3% gain on SeasFire. Apart from a low-complexity edge case on ClimateNet, absorption score increased with rule accuracy on both datasets; co-occurrence increased with rule complexity on SeasFire but decreased on ClimateNet.
On SeasFire, the reported gains were 22% over neuron baselines and 9% over prototype baselines; on ClimateNet, they were 9% and 5%, respectively.
In the fire-specific analysis, 92% of the top 50 GeoTopK explanations contained conditions linked to at least one of seven literature-supported fire-driver categories. The remaining 8% were labeled Unknown.
GeoTopK’s fire rules contained literature-consistent conditions in 46% of cases, while inverse conditions appeared in 2%. The corresponding inverse-condition rates were 16% for ClimaX neurons and 20% for prototypes.
Where the evidence stops
But the fire percentages are not an independent validation of the explanations. They measure agreement with literature-supported categories; the study did not test the rules against independent ground-truth semantic labels, human evaluation or operational deployment.
In a separate three-run sampling analysis, adversarial negative sampling produced more stable rules, less geographic-condition weighting and slightly higher fire-condition consistency than random sampling. The paper reports these results as means with standard deviations across the three runs.
The wider evidence remains narrow: only two datasets and three event types were evaluated, and rule-based feature interpretation focused on ClimaX. The main comparisons did not report confidence intervals or inferential tests, so the results support a computational interpretation finding in these setups rather than generalization to other regions, forecast horizons, models or operational decisions.
The document is an arXiv preprint, version 1 dated 20 Aug 2026; journal publication or peer-review status is not reported. The paper says its code is available in the SAE-Xplainers GitHub repository.
Paper data and sources
Original title: SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events
Authors: Hugo Porta, Emanuele Dalsasso, Chang Xu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text