PILA: Physics-Informed Low Rank Augmentation for Interpretable Earth Observation

Physically meaningful representations are essential for Earth Observation (EO), yet existing physical models are often simplified and incomplete. This leads to discrepancies between simulation and observations that hinder reliable forward model inversion. Common approaches to EO inversion either ignored this incompleteness or relied on case-specific preprocessing. More recent methods use physics-informed autoencoders but depend on auxiliary variables that are difficult to interpret and multiple regularizers that are difficult to balance. We propose Physics-Informed Low-Rank Augmentation (PILA), a framework that augments incomplete physical models using a learnable low-rank residual to improve flexibility, while remaining close to the governing physics. We evaluate PILA on two EO inverse problems involving diverse physical processes: forest radiative transfer inversion from optical remote sensing; and volcanic deformation inversion from Global Navigation Satellite Systems (GNSS) displacement data. Across different domains, PILA yields more accurate and interpretable physical variables. For forest spectral inversion, it improves the separation of tree species and, compared to ground measurements, reduces prediction errors by 40-71% relative to the state-of-the-art. For volcanic deformation, PILA’s recovery of variables captures a major inflation event at the Akutan volcano in 2008, and estimates source depth, volume change, and displacement patterns that are consistent with prior studies that however required substantial additional preprocessing. Finally, we analyse the effects of model rank, observability, and physical priors, and suggest that PILA may offer an effective general pathway for inverting incomplete physical models even beyond the domain of Earth Observation. The code is available at https://github.com/yihshe/PILA.git.

💡 Research Summary

The paper tackles a fundamental challenge in Earth Observation (EO): physical forward models are often simplified for computational efficiency, leaving them incomplete and causing a mismatch between simulated and real observations. This mismatch hampers the inversion of observations to retrieve physically meaningful variables (Z_phy). Existing solutions either ignore model incompleteness, rely on case‑specific preprocessing, or employ physics‑informed autoencoders that introduce hard‑to‑interpret auxiliary variables and require balancing multiple regularizers.

Inspired by Low‑Rank Adaptation (LoRA) originally developed for fine‑tuning large language models, the authors propose Physics‑Informed Low‑Rank Augmentation (PILA). The core idea is to augment an incomplete physical model F with a learnable low‑rank residual Δ (e.g., rank r ≪ d, where d is the observation dimension). The observation model becomes X ≈ F(Z_phy) + Δ, where Δ is generated by a mapping C that takes auxiliary latent variables Z_aux (and optionally the physical reconstruction X_F) as input. Crucially, a stop‑gradient is applied to X_F during back‑propagation, ensuring that the physical pathway (the parameters Z_phy and the forward model F) is not altered by the data‑driven residual. This separation preserves the interpretability of Z_phy while allowing Δ to capture unmodeled processes.

Training minimizes a simple reconstruction loss L_rec = ‖X − (X_F + Δ)‖². The low‑rank constraint on Δ acts as a “knob” that controls the trade‑off between flexibility (reducing reconstruction error) and physical fidelity (preventing Z_phy from drifting). The rank r is a hyper‑parameter; experiments show that modest ranks (5–10) provide sufficient expressivity without sacrificing interpretability.

The framework is evaluated on two disparate EO inverse problems:

1. Forest Radiative Transfer Inversion – Using Sentinel‑2 multispectral data over mixed Austrian forests, the authors augment a 1‑D radiative transfer model (RTM) based on Atzberger (2000). Compared with a state‑of‑the‑art physics‑informed VAE (Hybrid VAE), PILA achieves markedly better species separation and reduces errors in leaf‑area index, chlorophyll content, and soil moisture by 40–71 % relative to ground measurements. The low‑rank residual effectively learns to correct spectral discrepancies caused by complex canopy structure that the simplified RTM cannot capture.

2. Volcanic Deformation Inversion – GNSS time‑series from the Akutan volcano are fed directly into the model, which augments a classic Mogi point‑source deformation model. PILA automatically isolates a major inflation event in early 2008, estimating source depth, volume change, and surface displacement patterns that align with prior studies—yet without the extensive preprocessing (trend removal, ICA, etc.) those studies required. The residual captures tectonic and hydrological signals that the simple Mogi model omits.

Additional analyses explore how the rank r, the observability of each physical variable, and the choice of physical priors affect performance. Variables that are strongly observable (e.g., LAI) are robust to rank changes, whereas weakly observable variables (e.g., subsurface pressure) benefit from higher rank and well‑chosen priors. Over‑parameterizing Δ leads to loss of interpretability, while too low a rank leaves systematic errors uncorrected.

Overall, PILA demonstrates that a low‑rank data‑driven augmentation can bridge the gap between computationally tractable physical models and the richness of real EO data, delivering more accurate and physically interpretable inversions. The authors release their code, facilitating reproducibility and encouraging extensions to other domains where incomplete physics hampers inverse modeling, such as climate, hydrology, and atmospheric dynamics. Limitations include reliance on differentiable forward models and a linear‑plus‑nonlinear residual structure; future work may explore more expressive residuals and applications to non‑differentiable simulators.