Paper summary: Spectral-Spatial Interpretation

Problem

Post-hoc explanations usually give either a spatial heatmap (where the model looks) or global channel scores (which channels matter). For multispectral data, such as satellite bands, channel relevance is often local and physically meaningful: a band may matter only over burned land.

Method

Spectral-Spatial Interpretation (SSI) attributes over position–channel pairs. For each spatial position, it fixes the observed surrounding context, defines a conditional game over that position’s channels, and applies the Shapley value. The result is a full spectral-spatial attribution matrix. It satisfies Shapley fairness within each position, and summing a position’s channel attributions exactly recovers that position’s contribution relative to the baseline.

Key results

From the abstract:

• on HLS Burn Scar, SSI beats an MC-Owen estimator on retained attribution metrics, using about 6% of its evaluation budget
• across CUB-200, EuroSAT and HLS Burn Scar with two baselines, it has the best joint deletion and insertion AUC over dense, flat and rank-1 baselines, plus the strongest spectral hit rates

Limitations

There’s no dedicated limitations section in the extracted text. Shapley-based methods still need many model evaluations, which is why the evaluation budget is reported.