Physics-informed methane plume screening from EnMAP and PRISMA SWIR hyperspectral imagery

Abstract

Spaceborne imaging spectrometers such as EnMAP and PRISMA provide contiguous shortwave-infrared (SWIR) spectra that resolve greenhouse-gas (GHG) absorption features and enable detection of localized emission plumes at high spatial resolution. For rapid screening of large scene archives, statistical retrievals (matched-filter and least-squares variants) are attractive due to their computational efficiency; however, their relationship to physically meaningful column enhancements is often scene- and regime-dependent.

We present an end-to-end workflow for EnMAP and PRISMA Level-1C data that combines physics-informed target generation with fast statistical retrievals, and benchmark it against a physics-based reference retrieval. Scene-dependent target spectra are derived from radiative-transfer motivated absorption templates and adjusted for geometry and atmospheric state (ERA5). In this contribution, methane is used as the primary case study; the same processing pattern is designed to be extensible to other SWIR-absorbing gases. Using these targets, we compute dense per-pixel enhancement proxies with classical statistical retrievals, including a classical matched filter (CMF), albedo-reweighted matched filtering, linear and generalized least-squares (LLS/GLS), and SVD-based variance suppression. As a reference product we run the Beer--Lambert InfraRed Retrieval Algorithm (BIRRA), which is physically interpretable but computationally expensive.

Across representative scenes, the statistical methods reproduce plume morphology and show per-pixel correlation with BIRRA, but their behavior is strongly scene- and regime-dependent. In particular, the mapping from statistical scores (e.g., matched-filter $\alpha$) to column enhancements can be non-linear and may saturate at high concentrations, limiting discrimination of very large enhancements. In multi-scene comparisons, GLS is generally the most robust default across diverse surface types and masking conditions, while albedo-reweighted variants can reduce false positives over heterogeneous backgrounds.

In addition, we evaluate a machine-learning acceleration pipeline that predicts GHG column proxies directly from per-pixel spectra and contextual information (demonstrated for methane). Training data are generated by sampling pixels within each scene and pairing the SWIR spectrum with per-pixel geometry, scene-level metadata, and ERA5 features, using collocated BIRRA retrievals as targets. A multilayer perceptron regressor is trained with scene-wise data splits to prevent information leakage and is tailored to heavy-tailed enhancement distributions by emphasizing plume ``tail'' samples in the loss.

Finally, retrieval outputs from all methods are compared using complementary metrics that target both quantitative agreement and plume detectability: pixelwise correlation on a common valid domain, plume-mask overlap (IoU/Dice) based on statistically derived masks, and multi-scale spatial consistency via the Fractional Skill Score (FSS). We find that neighborhood-scale agreement (FSS) typically improves with increasing window size, indicating higher consistency in coarse plume location than in pixel-level plume boundaries. An EnMAP--PRISMA collocation analysis further highlights that validity masking, surface heterogeneity, and sensor-specific artifacts (e.g., striping) can dominate cross-sensor differences.