AeroDS-Syn: Synthetic ASI Shadow Dataset

• AeroDS-Syn is a synthetic paired shadow dataset that accurately simulates the soft penumbrae and gradual spatial attenuation typical of aerospace imagery.
• It is generated through PDSS-Net, which combines a linear illumination-degradation model with guided filtering and CycleGAN-based refinement to mimic real shadow formation.
• The dataset powers the Penumbra-aware Cascaded DeShadowing Network (PCDS-Net), yielding improved restoration metrics and better artifact handling compared to traditional methods.

AeroDS-Syn is a synthetic paired shadow dataset introduced within the AeroDeshadow framework for aerospace imagery (ASI). Its defining purpose is to provide supervised training data for shadow removal in a setting where strictly paired real shadowed and shadow-free ASI are severely lacking and where shadow formation is dominated by broad, soft penumbrae rather than hard binary boundaries. AeroDS-Syn is generated by the Physics-aware Degradation Shadow Synthesis Network (PDSS-Net), which combines a physics-guided illumination-degradation model with refinement for realistic spatial attenuation, and it serves as the sole supervised training source for the Penumbra-aware Cascaded DeShadowing Network (PCDS-Net) (Lu et al., 17 Apr 2026).

1. Motivation and problem setting

AeroDS-Syn was proposed to address two difficulties that make standard shadow-removal datasets and methods inadequate for ASI. First, paired ASI shadow data are extremely hard to obtain. Satellites revisit scenes only periodically, aerial flight paths vary, illumination and atmospheric conditions change, and the exact same scene without shadows usually cannot be captured under identical conditions. As a result, the large-scale paired supervision commonly used in deep shadow-removal pipelines is generally unavailable in aerospace imaging (Lu et al., 17 Apr 2026).

Second, aerospace shadows are not well described by hard-mask assumptions. In ASI, shadows often contain a dark umbra core, a wide penumbra transition zone, and gradual spatial attenuation due to scene geometry, long-range light propagation, and atmospheric scattering. Methods trained on homogeneous shadow assumptions tend to produce over-correction in the umbra, under-correction in the penumbra, visible boundary artifacts, and spectral inconsistency. AeroDS-Syn was constructed specifically to encode realistic soft-boundary shadows and physically plausible illumination decay.

A recurrent misconception is that a synthetic shadow dataset for ASI can be assembled by applying simple intensity attenuation under binary masks. AeroDS-Syn was designed against that assumption. Its role is not merely to increase sample count, but to approximate the physics of aerospace shadow formation closely enough that a network trained only on synthetic supervision can generalize to real-world ASI.

2. Physics-guided synthesis pipeline

AeroDS-Syn is produced by PDSS-Net, whose synthesis process begins from a linear illumination-degradation model relating a real shadow pixel $I_s$ and its adjacent lit pixel $I_l$: $I_s = w \cdot I_l + b$ where $w \in \mathbb{R}^3$ is a channel-wise illumination scaling factor and $b \in \mathbb{R}^3$ is a channel-wise environment light shift bias (Lu et al., 17 Apr 2026).

To estimate physically grounded degradation parameters, the framework uses real shadowed images and their masks. Morphological erosion isolates a shadow core, morphological dilation defines nearby lit regions, and statistical moments provide

$$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}$$

where $\sigma_{sc}, \mu_{sc}$ are the standard deviation and mean of the shadow core, and $\sigma_{lit}, \mu_{lit}$ are the standard deviation and mean of the adjacent lit region. The resulting parameter library is sampled during synthesis so that synthetic shadows inherit attenuation patterns measured from real ASI.

Hard pseudo-shadow masks are then converted into soft masks through guided filtering: $$I_{psm}^{soft} = \text{GuidedFilter}(I_{psm}^{hard}, I_{free})$$ This edge-preserving transformation is central to the penumbra-aware character of AeroDS-Syn, because it inserts a gradual transition zone before image synthesis.

Given a shadow-free image $I_{free}$, a soft pseudo-shadow mask $I_l$0, and sampled parameters $I_l$1, the initial synthetic shadow is generated by

$I_l$2

Outside the mask, the original image is preserved; inside the mask, a physically grounded de-exposure transform is applied.

This coarse shadow is then refined by a CycleGAN-based generator with a U-Net-style architecture. A key component is the Spatial-Decay Coordinate Attention (SDCA) module, inserted at skip connections to model the transition from umbra to penumbra. Its feature reweighting is written as

$I_l$3

with coordinate pooling

$I_l$4

concatenation

$I_l$5

and separate local-perception and decay-simulation branches. The stated function of SDCA is to balance local sharpness for umbra edges with distance-dependent decay for penumbra diffusion.

3. Dataset composition and statistics

AeroDS-Syn is one of two sub-datasets in the broader AeroDS benchmark. AeroDS contains AeroDS-Syn, a synthetic paired dataset for supervised training and validation, and AeroDS-Real, a real-world unpaired test set for generalization evaluation (Lu et al., 17 Apr 2026).

The paper reports an initial collection of 2,260 real shadowed images, 2,260 unpaired shadow-free images, and 2,260 manually annotated pseudo-shadow masks. After PDSS-Net-based synthesis, AeroDS-Syn contains 2,260 triplets composed of a shadow-free image, a synthesized shadow image, and a corresponding shadow mask. The split is 2,000 for training and 260 for testing. AeroDS-Real contains 260 real shadowed ASI samples with manually annotated shadow masks but no paired shadow-free ground truth.

Component	Size / split	Notes
AeroDS-Syn	2,260 triplets	shadow-free image, synthesized shadow image, shadow mask
AeroDS-Syn train	2,000	supervised training
AeroDS-Syn test	260	paired evaluation
AeroDS-Real	260 samples	real shadowed ASI, unpaired

The imagery is reported at 0.3 m spatial resolution. Samples are initially cropped to 256 × 256 and then upsampled and normalized to 512 × 512 using Real-ESRGAN. The stated rationale is resolution normalization, easier mask annotation, and minimal disruption of radiometric properties.

The dataset is distinguished not only by scale but by its reported shadow statistics. On the AeroDS-Syn test set, the paper gives the following reference values: Real reference: MeanSLR $I_l$6, SLR range $I_l$7, $I_l$8; PDSS-Net: MeanSLR $I_l$9, SLR range $I_s = w \cdot I_l + b$0, $I_s = w \cdot I_l + b$1. The intended interpretation is that the synthetic data approximate real shadow-intensity behavior while retaining smooth spatial attenuation.

4. Function within AeroDeshadow

AeroDS-Syn is the sole supervised training source for PCDS-Net, the restoration stage of AeroDeshadow. This design is consequential because the method does not rely on paired real annotations. Instead, the framework uses physics-guided synthetic supervision to learn a deshadowing model that is subsequently evaluated on real ASI (Lu et al., 17 Apr 2026).

PCDS-Net is structured around the specific properties encoded in AeroDS-Syn. Its Penumbra-aware Two-stream Encoder separates the input into an umbra mask $I_s = w \cdot I_l + b$2 and a penumbra mask $I_s = w \cdot I_l + b$3; the umbra stream captures high-frequency texture, while the penumbra stream uses dilated convolutions to model boundary context. Attention Feature Fusion (AFF) then blends the two streams with

$I_s = w \cdot I_l + b$4

and

$I_s = w \cdot I_l + b$5

A Cascaded Refinement Decoder progressively reconstructs the output from deep semantics to fine detail: $I_s = w \cdot I_l + b$6

The training objective includes adversarial, reconstruction, color-consistency, and boundary-aware physical smoothness terms: $I_s = w \cdot I_l + b$7 with $I_s = w \cdot I_l + b$8 and $I_s = w \cdot I_l + b$9. The physical smoothness term is

$w \in \mathbb{R}^3$0

Because AeroDS-Syn explicitly contains soft masks, realistic penumbra, physically consistent degradation, and paired ground truth, it is suited to training a network that does not treat the shadow as a single homogeneous region.

5. Empirical properties and reported performance

The empirical value of AeroDS-Syn is assessed at two levels: the realism of the synthetic shadows themselves and the downstream performance of PCDS-Net trained only on this synthetic corpus (Lu et al., 17 Apr 2026).

On synthesis evaluation, PDSS-Net-generated shadows are reported to match real shadow statistics more closely than competing approaches, particularly in MeanSLR, SLR range, and boundary profile behavior. The boundary profiles are described as exhibiting a smooth, non-linear attenuation curve rather than step-like transitions. This is consistent with the dataset’s stated objective of reproducing wide penumbrae and gradual spatial decay.

On the AeroDS-Syn test set, PCDS-Net achieved PSNR-S = 21.91, SSIM-S = 0.79, and RMSE-S = 11.35. On the real AeroDS-Real test set, despite being trained solely on synthetic data, it achieved Entropy = 7.40 and BRISQUE = 12.70. The paper interprets these outcomes as evidence of stronger restored detail, fewer artifacts, and better perceptual quality than competing methods, which are said to exhibit blue-green color distortions, boundary fractures, halo artifacts, and blurred texture recovery.

The reported generalization extends beyond the AeroDS benchmark to AISD and SRGTA. Ablation results further support the dataset design: removing the physically guided synthesis components degrades realism, with the paper stating that without DEP the shadow-intensity statistics deviate more from real shadows, and without SDCA the boundary transitions become sharp and unnatural.

A plausible implication is that AeroDS-Syn functions not merely as synthetic pretraining data, but as a domain-specific approximation to the supervisory signal that would otherwise require infeasible paired ASI acquisition.

6. Scope, terminology, and relation to other synthetic-data systems

Within the cited literature, AeroDS-Syn denotes the synthetic paired shadow dataset introduced by AeroDeshadow, rather than a generic aerial synthetic-data resource. The term can be confused with lexically similar systems in adjacent areas, but those systems address different tasks. AeroDGS is a physics-guided 4D Gaussian splatting framework for monocular UAV videos and dynamic aerial reconstruction, not a shadow dataset (Liu et al., 25 Feb 2026). AeroGen is a layout-controllable diffusion model for remote sensing image object detection and synthetic detection augmentation, not a deshadowing benchmark (Tang et al., 2024).

This distinction matters because AeroDS-Syn is defined by paired restoration supervision and penumbra-aware shadow physics. Its target variable is the transformation between shadowed and shadow-free ASI under realistic illumination decay, whereas AeroGen targets object-detection augmentation through horizontal and rotated bounding box control, and AeroDGS targets temporally coherent 4D scene reconstruction from a single monocular aerial sequence.

The broader context nevertheless suggests a common trend: aerospace and remote-sensing research increasingly uses domain-specific synthetic supervision when real annotations are scarce or structurally difficult to obtain. AeroDS-Syn occupies the shadow-removal end of that spectrum. Its differentiating claim is that the synthetic process is constrained by illumination decay, spatial attenuation, soft pseudo masks, and penumbra-aware refinement, rather than by purely visual plausibility alone.

In that sense, AeroDS-Syn is best understood as a physically grounded paired dataset for aerospace shadow restoration, built to encode the umbra-penumbra structure of ASI shadows and to support real-data generalization in the absence of paired real labels.