Scattering Mechanisms as a Shared Target for Federated PolSAR Learning

Radar archives around the world cannot be pooled. They sit with different agencies and companies, come from different satellites and aircraft, use different radar wavelengths, and are labeled with incompatible land-cover taxonomies: crops in one archive, wetland plants in another, urban density in a third. Federated learning is the obvious tool here. It trains one model by sending it to each archive and averaging the returned weights. But every archive predicts different classes, and some archives record all four transmit/receive polarization combinations while others record only two, so there is little for the averaged model to agree on. We propose to give the federation a common task. Every client trains its shared encoder to predict the physical scattering makeup of each pixel: how much of the radar echo came from a single bounce off a surface, how much from a double bounce off a wall-ground corner, and how much from diffuse scattering inside vegetation. Clients compute these quantities from their own data with standard formulas, so the targets cost no human labels.

The claim is that polarimetric scattering decomposition, the classical technique that splits a radar return into those canonical mechanisms, acts as a physically defined supervision signal that all clients share. That shared signal should let federated training work across clients whose label sets do not overlap and whose polarization modes differ. It should also teach clients that have only two polarizations to infer mechanisms they cannot compute directly. If this holds, each client's local classifier should gain accuracy, and a new sensor joining the federation should need fewer labels, compared with federating on raw channels.

Background and the gap

Three things make polarimetric radar (PolSAR) different from ordinary images. A fully polarimetric, or quad-pol, sensor records a 2×2 complex scattering matrix per pixel. After local averaging this becomes a 3×3 coherency matrix $\mathbf{T}$. Decompositions turn $\mathbf{T}$ into interpretable quantities. Model-based methods give surface, double-bounce and volume powers ($P_s, P_d, P_v$). The eigenvalue method of Cloude and Pottier gives entropy $H$, which measures how mixed the mechanisms are, and the angle $\alpha$, which indicates the dominant mechanism.

Two centralized studies motivate using these quantities rather than raw channels. Research on Input Schemes for Polarimetric SAR Classification Using Deep Learning fed AlexNet and VGG16 combinations of coherency elements and reflection-symmetry decomposition parameters on Gaofen-3 Yellow River Delta scenes. The 21-parameter scheme reached 98.1% overall accuracy, 14.51 points above a 6-parameter scheme, and tidal flats fell below 50% with the smaller input. Polarimetric Convolutional Network for PolSAR Image Classification preserved complex-matrix structure through a sparse real-valued coding and reported 98.24% on RADARSAT-2 and 97.73% on AIRSAR San Francisco. Both assume pooled data from a single sensor or scene family.

Two further lines constrain the design. The first concerns what two-channel sensors can recover. Scattering Mechanism Extraction by a Modified Cloude-Pottier Decomposition for Dual Polarization SAR showed that HH-VV dual-pol retains about 68% of scattering mechanisms, while HH-HV and HV-VV retain under 30%. A shared target must therefore distinguish quantities a dual-pol client can compute from those it can only estimate. The second is Quad-Pol SAR Data Reconstruction from Dual-Pol SAR Mode Based on a Multiscale Feature Aggregation Network (D2QNet), which learned, centrally, to reconstruct the full covariance from dual-pol input and reached 97.08% classification on UAVSAR Manitoba. That shows the missing information is partly predictable. It is also the closest competitor to our dual-pol pathway.

On the federated side, Learning Across Decentralized Multi-Modal Remote Sensing Archives with Federated Learning handles clients holding Sentinel-1 or Sentinel-2 data. It uses modality-specific backbones, zero-padding for the missing modality, batch whitening and contrastive alignment. It treats radar as two intensity images and assumes one shared label space. FedRS-Bench documents the non-IID gap on optical data: roughly 68–71% federated, about 20% local-only, and 75–79% centralized. It also reports that SCAFFOLD and FedDyn can destabilize, which warns us against relying on drift-correcting optimizers. Noise-Aware and Dynamically Adaptive Federated Defense Framework for SAR Image Target Recognition is the only federated SAR work in the file. It addresses backdoors hidden in speckle on single- or dual-pol chips, not heterogeneity. None of these federates polarimetric decomposition, and none addresses disjoint label taxonomies.

Core Claim

In a federation of PolSAR archives that differ in sensor, wavelength, polarization mode and label taxonomy, the scattering decomposition is the one per-pixel prediction every client can supervise. Training the shared encoder on it produces a representation that averages well across clients, because the physics is common to all of them even when the labels are not. We do not claim decomposition quantities are invariant across wavelengths. L-band penetrates canopies more deeply than C-band, so the same forest has a different volume fraction at each. The target is therefore conditioned on a wavelength token rather than assumed identical.

Research Program

Each client holds a regional archive and its own class list.

Architecture. The model has three parts:

  • a shared convolutional encoder, averaged across clients;
  • a shared decomposition head, also averaged;
  • a private classification head per client, never averaged.

Input. The encoder takes the coherency matrix in a fixed nine-channel real layout. Channels a client does not have are masked, and a binary indicator of which channels are present is appended. We apply orientation-angle compensation before computing anything. This removes the spurious cross-polarized power that tilted terrain and rotated buildings otherwise add to the volume term.

Decomposition targets. Targets come in two tiers:

  • Dual-pol tier. Quantities computable from both quad-pol and HH-VV data, namely the dual-pol $H/\alpha$ of Ji and Wu. Every client supervises this tier.
  • Quad-only tier. $P_s, P_d, P_v$ from a nonnegative model-based decomposition, plus the full-pol $H/\alpha$. Only quad-pol clients supervise this tier.

Channel dropout. During local training, quad-pol clients randomly mask their data down to HH-VV or VV-VH. They still supervise the quad-only targets. Dual-pol clients thereby receive, through averaging, an encoder that has learned to infer quad-only mechanisms from two channels.

Local objective. Client $k$ minimizes

$$\mathcal{L}_k = \mathcal{L}^{\text{cls}}_k + \lambda \sum_{j \in \mathcal{A}_k} \| \hat{d}_j - d_j \|_1,$$

where:

  • $\mathcal{L}^{\text{cls}}_k$ is cross-entropy on the client's own labeled pixels;
  • $\mathcal{A}_k$ is the set of decomposition quantities the client can compute from its native data;
  • $d_j$ is the computed value and $\hat{d}_j$ the prediction;
  • $\lambda$ trades off the two terms.

The decomposition term is applied to all pixels, labeled or not. For clients with few labels this is the main training signal.

Aggregation. The server averages encoder and decomposition-head weights with plain FedAvg, weighted by the number of pixels contributing to the decomposition loss. Classifier heads stay local.

Baselines

The naive reference point is local-only training: each client trains the same encoder and head on its own data with no communication. FedRS-Bench suggests this is weak under skew, but PolSAR scenes are internally large, so its strength here is an empirical question.

The main federated reference is shared encoder plus local heads, raw channels, no decomposition loss. It uses the identical architecture, masking, number of rounds, local epochs and tuning budget, so it isolates the effect of the decomposition target.

The strongest federated competitor from prior work is Büyüktas et al.'s recipe adapted to polarization mismatch: zero-padded missing channels, batch whitening in place of batch norm, and their contrastive alignment loss on co-located patches where sensors overlap. They report 76.7% F1 on BigEarthNet-MM under IID, 9.5 points above multi-sensor FedAvg, but no reported result found on Lacuna for any federated PolSAR benchmark. We will therefore run it ourselves under the same budget.

A second strong competitor is federated D2QNet. The shared model regresses the nine covariance parameters from masked input, using the same channel-dropout protocol, and classification heads sit on top. It tests whether reconstructing the whole matrix beats predicting a few physical quantities.

The centralized upper bound pools all clients, using PCN-style input and a multi-head classifier. The published anchors are 98.24% on RADARSAT-2 San Francisco and 97.73% on AIRSAR San Francisco from PCN, 98.1% on Yellow River Delta from the 21-parameter scheme, and 97.08% on UAVSAR Manitoba from D2QNet-v2. These are single-scene centralized numbers with their own splits. We report our pooled re-implementation alongside them, not as directly comparable targets.

Two ablations isolate the key ingredient:

  • Shuffled targets. Decomposition targets are permuted spatially within each client. If the gain is generic regularization from any dense auxiliary loss, this variant will match ours; if it depends on correct physics, it will not.
  • No orientation compensation. This tests the rotation-ambiguity issue directly.

Criteria for success. Beating the baselines means higher mean per-client overall accuracy than both Büyüktas-adapted and federated-D2QNet across three seeds. The gains must concentrate on dual-pol and few-label clients. The shuffled-target ablation must fall back toward the raw-channel baseline.

Why not just…?

Why not just feed decomposition parameters as input features and run FedAvg? Only quad-pol clients can compute the quad-only parameters. Zero-filling them on dual-pol clients gives the averaged encoder inputs whose meaning differs by client, which is the drift we want to avoid. Input stacking also does nothing for label-taxonomy mismatch. We include this variant as an extra arm. If it matches our method on quad-pol clients but trails on dual-pol clients, that confirms the target, not the features, is what transfers.

Why not federate D2QNet-style quad-pol reconstruction and decompose afterwards? This is the serious threat. D2QNet shows centralized reconstruction is accurate, and a reconstructed matrix contains everything a decomposition does. Our argument is that reconstructing nine correlated, speckle-dominated parameters spends model capacity on noise. The few decomposition powers are bounded, physically constrained and smoother. That argument is untested. The federated-D2QNet baseline settles it on two measures: classification accuracy, and mechanism error on dual-pol clients against withheld quad-pol truth. If reconstruction wins, we adopt it and the contribution narrows to the federated channel-dropout protocol.

Why not federated self-supervised pretraining, such as contrastive learning with physically realistic speckle views? PCM-SAR shows physics-aware contrastive views help SAR classification centrally, with an 11.52-point gain over SimCLR on FUSAR-Ship. However, contrastive objectives learn invariances chosen by the view generator, not the mechanism content that distinguishes flooded from dry vegetation or urban from forest. We add a federated contrastive arm with speckle augmentation, pretrained then fine-tuned, as a baseline under the same compute.

Why not rely on batch whitening alone? Whitening aligns feature statistics but cannot supply information a dual-pol client never measured, such as the co-polar phase relations that separate double bounce from volume. This is covered by the Büyüktas-adapted baseline and by per-mechanism error on dual-pol clients.

Proposed Experiments

Clients. The federation uses real archives from the file:

  • one client per PolSF sensor (PolSF): AIRSAR L-band, RADARSAT-2, ALOS-2, RISAT-1 and GF-3;
  • AIRSAR Flevoland crops;
  • Gaofen-3 Yellow River Delta wetlands;
  • UAVSAR Southern Manitoba crops.

Three clients, chosen to span both bands, are downgraded to dual-pol. Two keep HH-VV and one keeps VV-VH, which mirrors the retention gap Ji and Wu measured.

Splits. Train and test regions are spatially disjoint blocks within each scene, because neighboring pixels are correlated. PolSF clients image the same city, so their test blocks are aligned geographically to prevent leakage through co-location.

Metrics. We report:

  • per-client overall and average accuracy;
  • mean accuracy across clients;
  • mechanism mean absolute error and $H/\alpha$ zone agreement on dual-pol clients, measured against withheld quad-pol truth;
  • rounds to reach 95% of final accuracy.

Decisive checks. Three results would decide the question:

  1. Leave-one-sensor-out. The held-out client joins after training with 1%, 5% and 20% of its labels. The proposal predicts the largest margin over baselines at 1%.
  2. Shuffled-target ablation. This must lose most of the gain.
  3. Dual-pol mechanism error versus local-only. A dual-pol client's mechanism error should fall below what it achieves training alone on quad-pol simulation of its own scene. That would show the federation transferred physics it could not learn by itself.

A negative result on the first check, together with parity on the third, would indicate the decomposition target helps only as a regularizer. We would report it that way.