PBLH Estimation from Satellite Radiances via a Dual-Encoder Transformer

AI summary

Overview

Research area: Satellite remote sensing and computer vision, specifically estimating the Planetary Boundary Layer Height (PBLH) from passive infrared and microwave sounder radiances using deep learning.

Technical level: Intermediate. The paper assumes familiarity with regression metrics (MAE, RMSE, Pearson r) and Transformer/Vision Transformer concepts, but the problem and evaluation are explained clearly enough for a reader new to atmospheric remote sensing.

Scope: This paper benchmarks eight machine learning approaches for estimating PBLH from complete MetOp satellite overpasses, analyzes which input channels drive accuracy via Shapley attribution, and proposes a dual-encoder Transformer that handles missing satellite data directly.

What This Paper Is About

The Planetary Boundary Layer is the turbulent lowest layer of the troposphere, and its height determines how much vertical volume is available to dilute surface emissions and exchange energy with the free atmosphere. Measuring it currently depends on sparse ground instruments (radiosondes, ceilometers, lidars) or on slow numerical weather models, while satellite sounders could in principle provide continuous global coverage at every overpass. The core difficulty is that the relationship between top-of-atmosphere radiances and near-surface structure is indirect and non-linear, and no architecture had previously been built to handle the multimodal, spatially incomplete nature of a full satellite passage. This paper asks whether PBLH can be estimated from MetOp radiances alone, using a dual-encoder Transformer that processes an entire orbital passage at once.

Key Contributions

1. A comparative benchmark of eight models spanning three spatial granularities, pixel-wise regression (ridge, Random Forest, XGBoost, MLP), swath-wise sequence models (UNet-1D, Transformer-1D), and passage-wise convolutional and Transformer models (ResNet-2D and the proposed Transformer-2D), reported with parameter counts and per-pixel inference cost so accuracy can be weighed against computational footprint. The authors state this is the first comparison of its kind for satellite-based PBLH estimation.

2. A grouped Shapley attribution analysis over the 13 input blocks, computed exhaustively rather than sampled, and scored directly in metres of RMSE reduction with 95% bootstrap confidence intervals.

3. A dual-encoder Transformer architecture that routes infrared (IASI) and microwave (AMSU-A, MHS) radiances through two parallel Vision Transformer encoders, fuses them by element-wise addition, and passes the result through a Transformer decoder to produce dense per-pixel PBLH maps. A dynamic masking schedule used as a training-time augmentation lets the model consume passages containing gaps directly, whereas the convolutional and sequence baselines require missing values to be imputed first.

4. Out-of-distribution evaluation on 30 MetOp granules from two dates within the TEAMx observational campaign period, plus release of code and dataset at github.com/links-ads/pblh-transformer.

Main Findings

• Best accuracy on the held-out global test set: Transformer-2D achieves MAE = 155.8 m, RMSE = 215.3 m, and Pearson r = 0.850, outperforming every baseline on all three metrics across all four evaluation subsets (full, Europe-only, clear-sky, cloudy-sky).

• Spatial context helps, but the single-footprint signal already carries most of the skill: the MLP, using a single footprint with no spatial context and only 0.56 M parameters, reaches 177.6 m MAE on the full test set, beating both swath-wise models (UNet-1D at 191.9 m, Transformer-1D at 187.5 m) and approaching ResNet-2D (166.8 m). ResNet-2D, the other passage-wise model, is beaten by the MLP on full-test RMSE and by UNet-1D over Europe on both error metrics.

• The advantage is widest in cloudy conditions: Transformer-2D reaches 159.3 m MAE on the cloudy subset versus 183.2 m for the MLP, consistent with the microwave encoder continuing to provide a complete representation when cloud cover renders infrared channels unavailable.

• Out-of-distribution performance degrades only moderately: on the 30 TEAMx-period granules (14 February 2025 and 25 June 2025, 15 granules each), Transformer-2D achieves MAE = 165.3 m, RMSE = 226.6 m, and r = 0.852, versus the pixel-wise MLP baseline from the authors' previous work at MAE = 197 m and RMSE = 254 m.

• Ablations isolate the importance of the dual-encoder design: on global all-sky test RMSE, replacing the dual encoder with a single encoder costs +4.1 m (219.4 m), removing masking costs +2.5 m relative to the full model's 215.3 m (217.0 m with no masking), capping the masking ratio at 50% costs +2.5 m (217.8 m), and concatenation fusion instead of element-wise addition costs +1.1 m (216.4 m).

• Microwave channels dominate the attribution: microwave brightness temperatures contribute 101.1 m (32% of the total Shapley value) and infrared radiances 56.4 m (18%). Latitude (51.0 m) and the land/sea flag (50.2 m) are individually comparable to the entire 168-channel infrared block. Satellite zenith (19.1 m) and solar zenith (12.6 m) contribute more modestly, while day of year is statistically indistinguishable from zero.

• Errors are higher over Europe, consistent with target variability: the Europe-only subset shows higher errors for every model, with ERA5 PBLH standard deviation of 438 m over Europe versus 410 m globally.

• Fast inference: on a single consumer-grade NVIDIA RTX 2080 Ti, Transformer-2D produces a full-passage PBLH map of 765 × 30 footprints in approximately 0.46 s (50k pixels/s), allowing a full day of global MetOp coverage to be processed in a couple of minutes on one such card.

• Reported errors are of the same order as the label's own error: the authors note that Guo et al. found ERA5 underestimates daytime PBLH by around 130 m against near-global radiosondes, while their model reaches MAE = 155.8 m against ERA5, so small differences between the better-performing models should be interpreted with caution.

Methodology in Plain English

The researchers started from an existing large-scale dataset that pairs raw top-of-atmosphere radiances from MetOp satellites with ERA5 reanalysis PBLH values. The radiances come from IASI (infrared) and AMSU-A/MHS (microwave), distributed openly by EUMETSAT. The dataset covers the year 2022 sampled at weekly intervals (one date per week from January 3 to December 26) with four synoptic overpass times per date (00, 06, 12, and 18 UTC), yielding 194 orbital passages and over 4.3 million samples.

Each satellite footprint is turned into a 207-channel input vector: 20 microwave brightness temperatures, 168 infrared radiances, and 19 auxiliary features (sine/cosine–encoded latitude, longitude, time of day, day of year, and satellite and solar angles, plus surface elevation, cloud flag, and land flag). Channels are standardized using training-set statistics, and the PBLH target is normalized by a fixed 4000 m scale factor. Splits are made month by month, with the first sampled date of each month going to test, the second to validation, and the rest to training, so every split spans the full annual cycle.

They then built a ladder of models at three spatial granularities to see how much context actually matters. Pixel-wise models see one footprint at a time. Swath-wise models see one 30-position scan line at a time. Passage-wise models see the whole 2D overpass. The proposed model pads a passage to 800 × 30 pixels, treats each pixel as a token with patch size 1 × 1, and splits the input into two streams: a microwave stream of 39 channels (20 brightness temperatures plus 19 auxiliary features) and an infrared stream of 187 channels (168 radiances plus the same auxiliary features). Each stream gets its own ViT-style encoder (dimension 384, depth 2, 16 attention heads with head dimension 24), decoupled in weight space so each can learn modality-specific representations. The two outputs are added element-wise, learnable missing-token embeddings are inserted at masked positions, and a shared Transformer decoder of depth 4 produces a scalar PBLH per token.

The key trick is that during training a dynamic masking schedule randomly selects a masking ratio between 0% and 100% of pixels in each window, replacing masked tokens with a learnable missing-token embedding. Microwave and infrared missing tokens are learned separately. This mimics the gaps real passages contain (cloud contamination of infrared channels, fill-flagged or invalid values) and trains the model to recover through the downstream regression loss rather than through a separate reconstruction objective. At inference, the full passage is decomposed into overlapping 16 × 16 tiles with stride 8, and overlapping predictions are averaged. Baselines that cannot handle sparse input (UNet-1D, Transformer-1D, ResNet-2D) receive interpolated values instead.

Finally, the authors ran two analyses: an ablation study changing one architectural choice at a time, and a grouped Shapley decomposition over 13 input blocks, computed on 1024 fully observed 16 × 16 evaluation crops drawn from test-set passages, with each block's contribution scored in metres of RMSE reduction.

Why This Matters

Impact on research. The paper shows that passive sounder radiances, instruments designed for other purposes, can support PBLH estimation at accuracy competitive with lidar- and reanalysis-based approaches, and it provides the first systematic comparison of model families for this task. The Shapley analysis is a notable methodological contribution: it quantifies that geolocation and surface-type context carry nearly as much information as the spectral radiances themselves, which is a useful signal for anyone designing future retrieval models. The paper is also candid that its labels are ERA5 reanalysis rather than direct observations, and that its reported error is of the same order as ERA5's own error against radiosondes, a framing that is unusually honest for a benchmark paper.

Real-world applications.

• Air quality forecasting and regulation, since PBLH sets the vertical volume available to dilute surface emissions.
• Wind energy siting and operation, where boundary layer structure governs turbine-relevant flow.
• Numerical weather prediction, where dense satellite-derived PBLH could supplement sparse radiosonde networks.
• Filling observational gaps over oceans, polar regions, and developing-world areas that ground-based networks such as the pan-European E-PROFILE network leave unobserved.

Industry relevance. The model runs at 50k pixels/s on a single consumer-grade GPU, meaning a full day of global MetOp coverage can be processed in a couple of minutes on one card. That cost profile makes operational deployment plausible for weather services and commercial forecasting providers, and the masking-based design means no separate imputation pipeline is needed for cloudy scenes, which is where infrared-only retrievals typically fail.

Future Directions

• Replace ERA5 with direct observations as supervision. The authors identify this as the most consequential direction, proposing ground-based networks such as the pan-European E-PROFILE microwave radiometer and ceilometer network, or the TEAMx Observational Campaign once released, to evaluate against real observations and independently check the reanalysis.

• Extend the dataset to multi-year records to improve representation of seasonal variability and rare meteorological regimes.

• Incorporate temporally adjacent overpasses as additional context. The current model processes each orbital passage independently, capturing spatial but not temporal continuity, which could help resolve rapidly evolving boundary layer conditions. The authors describe this as a deliberate scoping choice rather than an architectural constraint.

• Add uncertainty quantification. The model currently outputs a single point estimate per footprint with no confidence measure; quantile regression, ensembling, or conformal prediction would let predictions be flagged by reliability, which matters most where inputs are heavily masked or reconstructed.

Target Audience

This paper suits machine learning researchers working on geospatial and remote sensing applications, especially those interested in multimodal architectures, masked-input handling, and model attribution methods. It is also relevant to atmospheric scientists and meteorologists evaluating satellite-based PBLH retrieval, to practitioners at weather services and commercial forecasting providers considering operational deployment, and to reviewers or policymakers tracking the readiness of PBL as an Earth observation target. Readers should be comfortable with regression metrics and basic Transformer terminology, but no deep background in radiative transfer is required.

Authors’ abstract

Estimating the Planetary Boundary Layer Height (PBLH) from satellite observations is a challenging regression problem due to the indirect relationship between top-of-atmosphere radiances and near-surface atmospheric structure. Progress has been limited both by the lack of architectures capable of handling the multimodal, spatially incomplete nature of satellite overpasses, and by the scarcity of suitable datasets. In this paper, we build upon the large-scale dataset pairing MetOp radiances with ERA5 PBLH labels that we introduced in our previous work, making three contributions. First, we establish a benchmark across eight approaches spanning pixel-wise regression, swath-wise sequence models, and convolutional and Transformer models operating on the full orbital passage. Second, we quantify what the resulting model actually relies on, using grouped Shapley decomposition over the input blocks. Third, we present the best-performing architecture found: a dual-encoder Transformer whose masked-input handling lets it operate in all weather conditions. The proposed model achieves MAE = 155.8 m on the held-out global test set, outperforming all baselines on every evaluation subset. On 30 out-of-distribution granules acquired on two days overlapping the TEAMx observational campaign, it achieves MAE = 165.3 m, outperforming a pixel-wise baseline trained on the same data (MAE = 197 m).