Source-linked AI summary
PBLH Estimation from Satellite Radiances via a Dual-Encoder Transformer
Lorenzo Innocenti, Luca Catalano, Edoardo Arnaudo, Claudio Rossi, Salvatore Larosa, Domenico Cimini, Paolo Garza
TL;DR
Satellite PBLH retrieval is challenging because radiances are indirectly related to near-surface structure, and suitable datasets and architectures for incomplete multimodal overpasses are limited. This paper benchmarks approaches and proposes a dual-encoder Transformer; it achieves the best reported performance on the global test set and TEAMx-period granules.
Problem
Estimating PBLH from satellite radiances is challenging, and progress is limited by architectures for heterogeneous, spatially incomplete overpasses and scarce large-scale radiance–PBLH datasets.
Source: §1 · p. 2Method
The paper benchmarks eight model approaches and proposes a dual-encoder Transformer that processes infrared and microwave radiances with dynamic masking for missing observations.
Source: §1 · p. 2Results
Transformer-2D outperforms every baseline on all evaluation subsets, with MAE of 155.8 m and Pearson r = 0.850 on the global test set; on 30 TEAMx-period granules, its MAE is 165.3 m versus 197 m for the pixel-wise MLP baseline.
Source: §5.2 · p. 8, §6 · p. 10Takeaways & Limitations
The results support estimating PBLH from MetOp radiances across clear-sky, cloudy, and TEAMx-period out-of-distribution cases.
Source: §6 · p. 10Evaluation measures agreement with ERA5 reanalysis rather than ground measurements, so reported errors do not represent distance from true PBLH.
Source: §6.1 · pp. 10–11
Abstract
from arXiv · showhide
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).
1 Introduction
The paper addresses the challenge of estimating PBLH from satellite observations amid gaps in suitable datasets and architectures for heterogeneous, incomplete overpasses. It benchmarks eight approaches, analyzes input-group contributions, and proposes a dual-encoder Transformer for complete MetOp passages.
1 Introduction
- Satellite remote sensing can enable continuous, globally dense PBLH mapping at each orbital overpass, but progress is limited by architecture and dataset gaps.The challenge includes jointly handling heterogeneous, spatially incomplete multi-instrument overpasses and a scarcity of large-scale datasets pairing radiances with reliable PBLH labels.Source: §1 · p. 2
- The paper benchmarks eight models spanning pixel-wise regression, swath-wise sequences, and convolutional and Transformer models over complete passages.The benchmark reports parameter counts and per-pixel inference cost alongside accuracy.Source: §1 · p. 2
- Grouped Shapley decomposition quantifies model reliance across 13 input blocks, exhaustively and directly in RMSE.The blocks comprise microwave and infrared channels plus auxiliary inputs.Source: §1 · p. 2
- The dual-encoder Transformer handles gaps in satellite passages directly and outperforms the pixel-wise baseline of on the same data.Infrared and microwave radiances are processed in parallel, and dynamic masking mimics real acquisition gaps rather than training a separate reconstruction objective.Source: §1 · p. 2
2 Related Work
Prior PBLH estimation combines sparse ground measurements with satellite approaches based largely on active sensing or reanalysis. Passive sounder radiances offer broad coverage, but direct retrieval from them remains largely unexplored.
2 Related Work
- Satellite-derived PBLH estimates, including this work, use instruments designed for other purposes rather than a dedicated PBLH mission.The PBL was identified as an Incubation Targeted Observable in the 2017 US Decadal Survey.Source: §2 · p. 3
- Ground networks provide continuous records at instrumented sites, while satellite observations offer global coverage at each overpass.Existing satellite approaches include CALIPSO lidar and GNSS radio occultation, whose sampling is sparse and irregular and whose footprint is too coarse for sharp gradients.Source: §2 · p. 3
- Machine-learning retrievals have largely used active-sensing backscatter or reanalysis predictors rather than passive radiances.Examples include CALIOP and CATS backscatter models and an XGBoost model trained on ERA5 meteorological and geographical predictors.Source: §2 · p. 3
- Passive infrared and microwave sounders offer wide-swath, high-revisit sampling, and microwave sensitivity through cloud, but PBLH retrieval from sounder radiances alone remains largely unexplored.The closest work estimates PBLH from combined infrared and microwave radiances footprint by footprint.Source: §2 · pp. 3–4
- Earlier approaches treat retrieval locations independently, whereas this work uses raw infrared and microwave radiances with spatial context across the full overpass.The authors describe this as the first such dual-encoder approach and an encouraging first step, not a final answer on how broadly MetOp-class radiances can support PBLH retrieval.Source: §2 · p. 4
3 Dataset
The study uses a year of MetOp radiances paired with ERA5 PBLH labels, partitions dates to limit leakage while representing every season, and evaluates separately on 2025 TEAMx-period granules.
3 Dataset
- The dataset contains 194 orbital passages and over 4.3 million samples from 2022, pairing MetOp IASI and AMSU-A/MHS radiances with ERA5 PBLH targets.Weekly dates across the year are sampled at four synoptic overpass times; ERA5 targets average the four nearest grid points.Source: §3 · p. 4
- The out-of-distribution evaluation uses 30 MetOp granules from 14 February and 25 June 2025, outside the 2022 dataset and within the TEAMx campaign period.Fifteen granules were selected from each date.Source: §3 · p. 4
- The train, validation, and test split assigns one sampled day per month to each held-out split, preserving seasonal representation and preventing data leakage.Channels are standardized using training-set statistics, and PBLH is normalized by a fixed 4000 m scale factor.Source: §3 · p. 4
- Figure 1 compares ERA5 reference PBLH, Transformer-2D predictions, and prediction-minus-reference along MetOp ground tracks on 14 February 2025.The reference and prediction panels share a 0–3000 m colour scale; the difference panel spans ±1000 m.Source: Figure 1 · p. 5
- Each observation has 207 channels: 20 microwave temperatures, 168 infrared radiances, and 19 auxiliary features.Auxiliary inputs include cloud and land flags, surface elevation, and encoded periodic quantities.Source: §3 · p. 5
4 Methodology
The methodology compares models that use no spatial context, a single scan line, or a complete orbital passage. The proposed passage-wise Transformer processes infrared and microwave streams with separate encoders and handles missing observations through masking.
4 Methodology
- The benchmark compares seven baselines with Transformer-2D across three spatial granularities, each using a different level of satellite-data context.The proposed model operates on complete two-dimensional orbital passages.Source: §4 · p. 5, §4.3 · p. 6
4.1 Pixel-wise Models
- Pixel-wise models map one IASI footprint's 207-feature vector to a scalar PBLH estimate without spatial context.The baselines include ridge-regularized linear regression, Random Forest, XGBoost, and a residual-block MLP.Source: §4.1 · p. 5
4.2 Swath-wise Models
- Swath-wise models process each scan line as a 30-position sequence, filling missing values by linear interpolation before inference.UNet-1D uses an encoder-decoder with skip connections; Transformer-1D applies self-attention along the scan line with positional encodings.Source: §4.2 · p. 5
4.3 Passage-wise Models
- ResNet-2D requires dense input, so missing pixels are interpolated before its convolutional processing.Its training uses flips, rotations, affine transforms, and random crops; inference uses mirror padding.Source: §4.3 · p. 6
- Transformer-2D splits microwave and infrared inputs into separate streams, each with its own encoder, then fuses their representations by element-wise addition.The microwave stream has 20 radiance and 19 auxiliary channels; the infrared stream has 168 radiance and the same auxiliary features.Source: Proposed approach · p. 6, Figure 2 · p. 6
- The passage-wise Transformer encodes visible tokens, inserts learned missing-token embeddings at masked positions, and decodes a complete sequence into per-pixel PBLH estimates.Each pixel is a token; inference tiles the passage into overlapping 16 × 16 windows with stride 8 and averages overlapping predictions.Source: Proposed approach · pp. 6–7
5 Experiments
The experiments compare models across four evaluation subsets, test architectural choices, and quantify the contribution of input blocks. Transformer-2D leads the benchmark, while attribution identifies microwave radiances as the largest single input-block contributor.
5.1 Configurations
- All models share a 207-channel input and use grid-searched hyperparameters; Transformer-2D uses batch size 16, depth-2 ViT encoders, and a depth-4 decoder.The two ViT encoders each use dimension 384 and 16 attention heads.Source: §5.1 · p. 7
5.2 Quantitative Results
- Transformer-2D outperforms every baseline on all three metrics across four subsets; global-test MAE is 155.8 m, and 30 campaign-period granules yield MAE 165.3 m versus the MLP baseline's 197 m.The campaign-period evaluation uses granules from two dates outside the training and validation splits.Source: §5.2 · p. 8, Table 1 · p. 9
5.3 Ablation Study
- Every tested change raised global all-sky RMSE: a single encoder cost +4.1 m, concat fusion was smallest at +1.1 m, and the 50% masking cap added +2.5 m.The variants change one architectural choice at a time relative to the full model.Source: §5.3 · p. 9, Table 2 · p. 10
5.4 Channel Attribution
- Microwave channels contribute the largest Shapley share (101.1 m, 32%), ahead of infrared (56.4 m, 18%); latitude (51.0 m) and land/sea (50.2 m) are also substantial.Shapley values are expressed as test-set RMSE reduction, with 95% bootstrap intervals shown in Fig. 3.Source: Figure 3 · p. 10, §5.4 · p. 10
6 Conclusion
The dual-encoder Transformer performed best across evaluation subsets, including on out-of-distribution TEAMx-period granules. The conclusion is bounded by ERA5-based labels and identifies direct observations, temporal context, broader records, and uncertainty estimates as future directions.
6 Conclusion
- MAE of 155.8 m and Pearson r = 0.850 on the global held-out test set; the dual-encoder Transformer performed best on all evaluation subsets and degraded moderately on TEAMx-period granules.The result includes clear-sky and cloudy subsets.Source: §6 · p. 10
6.1 Limitations
- Reported errors measure agreement with ERA5 rather than ground measurements, and ERA5 underestimates daytime PBLH by around 130 m, warranting caution about small model differences.Some residual error may reflect the label’s noise floor rather than model deficiency.Source: §6.1 · pp. 10–11
- Each orbital passage is processed independently, so the model does not use temporal continuity that could help resolve rapidly evolving boundary-layer conditions.The paper describes this as a deliberate scoping choice, not an architectural constraint.Source: §6.1 · p. 11
6.2 Future Work
- Replacing ERA5 supervision with direct ground-based measurements could enable evaluation against observations and an independent check on the reanalysis.Suggested sources include E-PROFILE instruments and the TEAMx campaign once released.Source: §6.2 · p. 11
- Multi-year data could broaden seasonal and rare-regime representation, while temporally adjacent overpasses could add context for rapidly changing conditions.The paper identifies both as future directions.Source: §6.2 · p. 11
- Adding uncertainty estimates could flag prediction reliability, particularly when inputs are heavily masked or reconstructed, where errors are expected to be higher.Proposed approaches include quantile regression, ensembling, and conformal prediction.Source: §6.2 · p. 11