Hyperspectral Variational Autoencoders for Joint Data Compression and Component Extraction

TL;DR

This work tackles the challenge of terabytes-scale TEMPO hyperspectral data by introducing a variational autoencoder (VAE) that jointly compresses spatial and spectral information to a compact latent representation, achieving a compression of $\times 514$ with spectral reconstruction errors in the $1$–$2$ order-of-magnitude range below the signal. Beyond data reduction, the study probes whether atmospheric information is retained in the latent space by training linear and nonlinear probes to recover Level-2 products such as NO$_2$, O$_3$, HCHO, and cloud fraction, finding strong retrievals for cloud fraction ($R^2$ about $0.92$) and total ozone ($R^2$ about $0.81$) but more modest performance for NO$_2$ and HCHO. The results reveal a semi-linear encoding of atmospheric information in the latent space, with nonlinear probes substantially outperforming linear ones, and show that explicit latent supervision during training yields minimal improvement, highlighting fundamental encoding challenges for certain products. Overall, the approach demonstrates that neural compression can dramatically reduce hyperspectral data volumes while preserving essential atmospheric signals, offering significant benefits for data archival, sharing, and analysis in next-generation Earth observation systems.

Abstract

Geostationary hyperspectral satellites generate terabytes of data daily, creating critical challenges for storage, transmission, and distribution to the scientific community. We present a variational autoencoder (VAE) approach that achieves x514 compression of NASA's TEMPO satellite hyperspectral observations (1028 channels, 290-490nm) with reconstruction errors 1-2 orders of magnitude below the signal across all wavelengths. This dramatic data volume reduction enables efficient archival and sharing of satellite observations while preserving spectral fidelity. Beyond compression, we investigate to what extent atmospheric information is retained in the compressed latent space by training linear and nonlinear probes to extract Level-2 products (NO2, O3, HCHO, cloud fraction). Cloud fraction and total ozone achieve strong extraction performance (R^2 = 0.93 and 0.81 respectively), though these represent relatively straightforward retrievals given their distinct spectral signatures. In contrast, tropospheric trace gases pose genuine challenges for extraction (NO2 R^2 = 0.20, HCHO R^2 = 0.51) reflecting their weaker signals and complex atmospheric interactions. Critically, we find the VAE encodes atmospheric information in a semi-linear manner - nonlinear probes substantially outperform linear ones - and that explicit latent supervision during training provides minimal improvement, revealing fundamental encoding challenges for certain products. This work demonstrates that neural compression can dramatically reduce hyperspectral data volumes while preserving key atmospheric signals, addressing a critical bottleneck for next-generation Earth observation systems. Code - https://github.com/cfpark00/Hyperspectral-VAE

Figures (13)

• Figure 1: TEMPO satellite data products. Representative $131 \times 131$ pixel region showing (a) Level-1 radiance data as three-channel composite (channels 100, 500, 900 out of 1028), and (b-e) Level-2 atmospheric products: tropospheric NO$_2$, total ozone, HCHO, and cloud fraction. All L2 products are normalized (see App. \ref{['app:data_norm']}). The distinct spatial patterns—from localized NO$_2$ pollution to broad stratospheric ozone—illustrate the challenge of preserving multiple atmospheric signals through compression.
• Figure 2: VAE reconstruction quality. Two representative samples showing VAE reconstruction performance at 200,000 training steps. For each sample, we display: (left) original TEMPO radiance as three-channel composite (channels 100, 500, 900), (center-left) VAE reconstruction, (center) per-pixel MSE on log scale, and (right panels) normalized spectra at two random spatial locations (marked with red + and o). The VAE achieves $\times$514 compression ($1028 \times 64 \times 64 \rightarrow 32 \times 16 \times 16$) while accurately preserving both spatial structures and spectral features across the 290-490nm range. Note: The large features visible in low-index channels (left end, near 290 nm) are primarily artifacts from measurement uncertainties and stray light rather than genuine atmospheric signals.
• Figure 3: Channel-wise reconstruction errors across validation set. Root mean squared error (RMSE) computed per spectral channel over 61,440 validation spectra sampled from 960 tiles. (a) RMSE for normalized spectra showing wavelength-dependent reconstruction quality with errors ranging from $10^{-2}$ to $10^{-1}$ normalized units. (b) RMSE for physical radiance (red line) overlaid with mean radiance spectrum (black dashed line), demonstrating that reconstruction errors are 1-2 orders of magnitude smaller than the signal across the 290-490nm range, with slightly elevated errors only at the shortest wavelengths (290-310nm). Shaded regions indicate ±1 standard deviation.
• Figure 4: Unsupervised VAE probing results. Predicted vs. ground truth scatter plots for all four atmospheric products extracted from the base VAE latent space using (left) linear probes and (right) MLP probes. Each 2×2 grid shows: (a) NO$_2$ tropospheric vertical column, (b) total O$_3$ column, (c) HCHO vertical column, (d) cloud fraction. Red dashed lines indicate perfect prediction. MLP probes substantially outperform linear probes, particularly for cloud fraction (R$^2$: 0.785$\rightarrow$0.930) and total ozone (R$^2$: 0.545$\rightarrow$0.811), demonstrating that atmospheric information is encoded nonlinearly in the unsupervised latent representation. NO$_2$ remains challenging (R$^2$=0.203) despite the nonlinear extraction, indicating fundamental limitations of purely unsupervised compression for this trace gas.
• Figure 5: Joint optimization of reconstruction and L2 prediction. Combined view of L2 product prediction losses (left y-axis) and reconstruction error (right y-axis, red dashed line) during latent supervised VAE training over 220,000 steps. L2 product losses (NO$_2$ in purple, O$_3$ in blue, HCHO in teal, cloud in orange) improve significantly once reconstruction error plateaus around step 40,000. Cloud, O$_3$, and HCHO show overfitting in later stages, while NO$_2$ barely improves throughout training, indicating fundamental encoding difficulties for this trace gas even under explicit supervision.