AeroDeshadow: Physics-Guided ASI Shadow Removal

• The paper introduces AeroDeshadow, a two-stage framework that combines physics-guided shadow synthesis with penumbra-aware restoration for aerospace imagery.
• It leverages synthetic paired data via the AeroDS benchmark and separates umbra from penumbra to address spatially-varying shadow effects.
• The method demonstrates competitive performance across multiple benchmarks and improves over traditional homogeneous shadow removal approaches.

Searching arXiv for the focal paper and closely related remote-sensing deshadowing work. AeroDeshadow is a two-stage shadow-removal framework for high-resolution aerospace imagery (ASI) that combines physics-guided shadow synthesis with penumbra-aware restoration. It is formulated around two ASI-specific constraints: strictly paired real shadow/shadow-free data are severely lacking, and homogeneous shadow assumptions fail in the broad penumbra transition zones produced by complex scene geometry, long-distance light propagation, and atmospheric scattering. The framework therefore couples the Physics-aware Degradation Shadow Synthesis Network (PDSS-Net), the AeroDS benchmark, and the Penumbra-aware Cascaded DeShadowing Network (PCDS-Net) in order to restore illumination consistency while reducing spectral distortion, texture suppression, and boundary ambiguity in satellite and aerial imagery (Lu et al., 17 Apr 2026).

1. Domain formulation and shadow model

AeroDeshadow treats shadows in ASI as a radiometric and structural degradation rather than a merely aesthetic artifact. The paper identifies illumination inconsistency, spectral distortion, texture suppression, and information loss as direct consequences of shadows, with downstream impact on urban analysis, disaster monitoring, object detection, segmentation, and remote-sensing interpretation. In this formulation, ASI shadow removal is materially different from natural-image shadow removal because the same scene is rarely available under matched acquisition conditions with and without shadows, and because ASI shadows often contain both a darker umbra core and a broad penumbra where illumination changes continuously rather than abruptly (Lu et al., 17 Apr 2026).

A central premise is that binary or homogeneous shadow models are too crude for aerospace scenes. Uniform correction can over-brighten the umbra, under-restore the penumbra, and introduce spectral inconsistency or visible artifacts at soft boundaries. A common misconception in adjacent literature is that natural-image shadow removal pipelines can be transferred directly to aerospace imagery; AeroDeshadow explicitly rejects that assumption by making penumbra modeling and synthetic paired-data construction first-class components of the method.

The physical shadow-formation prior is written as a linear illumination degradation between adjacent lit pixels $I_l$ and shadow pixels $I_s$: $I_s = w \cdot I_l + b,$ where $w \in \mathbb{R}^3$ is a channel-wise illumination scaling factor and $b \in \mathbb{R}^3$ is a channel-wise environmental light shift bias. The illumination decay parameters are estimated from real ASI by isolating a shadow core region and an adjacent lit region: $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$ These $(w,b)$ values form the physical prior library used by the synthesis stage.

2. Two-stage architecture and data flow

The framework is organized as an overview stage followed by a restoration stage. In Stage 1, PDSS-Net takes a shadow-free ASI image and a pseudo-shadow mask and produces a realistic synthetic shadow image. In Stage 2, PCDS-Net takes a shadowed image and its mask and predicts a restored shadow-free image. The paper describes this as a bridge from clean real aerospace images to synthetic shadowed paired samples and then to real-world deshadowing (Lu et al., 17 Apr 2026).

The stage-level data flow is: $$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$ followed by

$$(\text{shadowed ASI}, \text{mask}) \xrightarrow{\text{PCDS-Net}} \text{restored shadow-free ASI}.$$

This division of labor is fundamental. PDSS-Net addresses the absence of paired real data by constructing synthetic supervision with soft boundary transitions, while PCDS-Net addresses the failure of homogeneous restoration by explicitly separating umbra and penumbra. The two-stage design is therefore not a generic cascade; it encodes the claim that data realism and restoration bias must be solved jointly, but not necessarily by a single end-to-end model.

3. Physics-guided synthesis and the AeroDS benchmark

PDSS-Net begins by transforming a hard pseudo-shadow mask into a soft mask with a guided filter: $$I_{psm}^{soft} = \text{GuidedFilter}(I_{psm}^{hard}, I_{free}).$$ A sampled $I_s$0 pair is then applied to produce a physics-guided initial shadow image: $I_s$1 This initial synthesis is refined by a CycleGAN-style generator built on a U-Net-style encoder-decoder, yielding

$I_s$2

The main spatial prior inside PDSS-Net is the Spatial-Decay Coordinate Attention (SDCA) module. For a feature map $I_s$3, SDCA reweights features by horizontal and vertical attention: $I_s$4 Coordinate pooling is defined by

$I_s$5

followed by

$I_s$6

The local perception branch is

$I_s$7

and the decay simulation branch is

$I_s$8

After fusion,

$I_s$9

the attention maps are

$I_s = w \cdot I_l + b,$0

The paper interprets the local branch as preserving umbra edge sharpness and the decay branch as simulating non-linear spatial attenuation in the penumbra.

PDSS-Net is trained with

$I_s = w \cdot I_l + b,$1

with $I_s = w \cdot I_l + b,$2, $I_s = w \cdot I_l + b,$3, and $I_s = w \cdot I_l + b,$4. The adversarial term is

$I_s = w \cdot I_l + b,$5

Cycle consistency is

$I_s = w \cdot I_l + b,$6

Background consistency is

$I_s = w \cdot I_l + b,$7

and identity preservation is

$I_s = w \cdot I_l + b,$8

The resulting benchmark, AeroDS, contains two subsets. All images come from diverse regions worldwide, have original spatial resolution of $I_s = w \cdot I_l + b,$9 m, are initially cropped to $w \in \mathbb{R}^3$0, and are then normalized to $w \in \mathbb{R}^3$1 using Real-ESRGAN.

Subset	Content	Split
AeroDS-Syn	2,260 triplets from real shadow-free images and pseudo-shadow masks	2,000 train / 260 test
AeroDS-Real	260 real shadowed ASI samples with manually annotated masks	test only

AeroDS as a whole contains 2,260 real shadowed images, 2,260 unpaired shadow-free images, and 2,260 manually annotated pseudo-shadow masks (Lu et al., 17 Apr 2026).

4. Penumbra-aware restoration in PCDS-Net

PCDS-Net begins by decomposing the input mask into an umbra mask $w \in \mathbb{R}^3$2 and a penumbra mask $w \in \mathbb{R}^3$3 through dynamic morphological erosion and dilation. The architectural claim is that umbra and penumbra obey different restoration dynamics: the umbra requires stronger illumination compensation and texture recovery, whereas the penumbra requires smooth boundary handling and contextual compensation (Lu et al., 17 Apr 2026).

The Umbra Feature Encoder (UFE) extracts high-frequency core-shadow features: $w \in \mathbb{R}^3$4 while the Penumbra Feature Encoder (PFE) uses dilated convolutions: $w \in \mathbb{R}^3$5

At each scale, the two streams are fused by Attention Feature Fusion (AFF). Starting from

$w \in \mathbb{R}^3$6

the module computes

$w \in \mathbb{R}^3$7

and fuses the branches as

$w \in \mathbb{R}^3$8

This gives the model a way to prefer umbra features where texture recovery is critical and penumbra features where smooth transition modeling is dominant.

The decoder first forms a semantic feature

$w \in \mathbb{R}^3$9

and then restores progressively: $b \in \mathbb{R}^3$0 The final prediction is denoted

$b \in \mathbb{R}^3$1

PCDS-Net is trained with

$b \in \mathbb{R}^3$2

where $b \in \mathbb{R}^3$3 and $b \in \mathbb{R}^3$4. The adversarial term is

$b \in \mathbb{R}^3$5

The reconstruction term is

$b \in \mathbb{R}^3$6

with $b \in \mathbb{R}^3$7 and $b \in \mathbb{R}^3$8. To preserve spectral fidelity, the color consistency loss is

$b \in \mathbb{R}^3$9

The boundary-aware physical smoothness term uses

$$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$0

and is written as

$$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$1

This enforces smooth illumination outside penumbra while preserving texture and transition behavior inside penumbra.

5. Benchmarking, ablations, and reported performance

The evaluation covers AeroDS-Syn, AeroDS-Real, AISD, and SRGTA. For synthesis, the paper reports Shadow-to-Lit Intensity Ratio (SLR) and chromaticity deviation in Lab $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$2-channel $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$3. For paired shadow removal it uses PSNR, SSIM, and RMSE, and for real ASI without ground truth it uses Entropy and BRISQUE (Lu et al., 17 Apr 2026).

The headline removal results are organized below.

Dataset	Metrics	AeroDeshadow
AeroDS-Syn	PSNR-S / SSIM-S / RMSE-S	21.91 / 0.79 / 11.35
AeroDS-Real	Entropy / BRISQUE	7.40 / 12.70
AISD	Entropy / BRISQUE	7.17 / 11.56
SRGTA	PSNR-S / SSIM-S / RMSE-S	22.37 / 0.78 / 9.41

On AeroDS-Syn, the strongest baseline RS-GSSR is listed at $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$4. On SRGTA, RS-GSSR is $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$5. On AeroDS-Real, RS-GSSR has a slightly better BRISQUE of $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$6, while PCDS-Net remains highly competitive at $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$7. This makes the performance profile mixed rather than uniformly dominant: AeroDeshadow leads on the paired synthetic and transferred synthetic benchmarks and on AISD, but does not produce the lowest BRISQUE on AeroDS-Real.

For synthesis realism, the real reference shadow distribution has MeanSLR $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$8, SLR Range $$w = \frac{\sigma_{sc}}{\sigma_{lit}}, \quad b = \mu_{sc} - w \cdot \mu_{lit}.$$9, and $(w,b)$0. PDSS-Net reports MeanSLR $(w,b)$1, SLR Range $(w,b)$2, and $(w,b)$3. RS-GSSR is numerically close on MeanSLR and range, with MeanSLR $(w,b)$4 and SLR Range $(w,b)$5. The paper nevertheless argues that PDSS-Net better captures local physical realism and non-linear penumbra transitions, and that global chromatic metrics can miss local transition fidelity.

The ablation study separates the roles of DEP, SDCA, UFE, PFE, and AFF. Without DEP, PDSS-Net drops to MeanSLR $(w,b)$6, SLR Range $(w,b)$7, and $(w,b)$8. For restoration, UFE only gives $(w,b)$9 on AeroDS-Syn and BRISQUE $$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$0 on AeroDS-Real; PFE only gives $$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$1 and BRISQUE $$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$2; UFE + PFE without AFF gives $$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$3 and BRISQUE $$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$4. These results support the paper’s main architectural claim: umbra-specific recovery, penumbra-specific context, and adaptive fusion are all necessary, and AFF is especially important for real-image quality.

The qualitative analysis reported in the paper attributes the final gains to smoother non-linear boundary attenuation, fewer step-like transitions, reduced halo artifacts, stronger texture preservation inside core shadows, and fewer blue-greenish distortions than competing baselines. The method is also evaluated in a downstream vehicle-detection setting with SCAF-Net, where the paper argues that the restoration is preferable to GAN-based removal because it preserves small-object structure more faithfully.

6. Position within remote-sensing deshadowing research

AeroDeshadow belongs to the line of remote-sensing deshadowing methods that respond to the lack of paired real supervision by constructing synthetic or weakly paired training data, but its distinctive claim is that the synthetic data must encode penumbra physics rather than only hard shadow masks. This places it near, but not identical to, other recent RSI frameworks. SARU unifies detection and removal through DBCSF-Net and the training-free N$$\text{shadow-free ASI} + \text{pseudo-shadow mask} \xrightarrow{\text{PDSS-Net}} \text{synthetic shadowed ASI},$$5SGSR algorithm, emphasizing single-image restoration without paired training data (Bo et al., 28 Apr 2026). deSEO addresses the satellite case by turning S-EO into a geometry-aware weakly paired benchmark and pairing it with a DSM-aware deshadowing model (Beltrame et al., 5 May 2026). ShadowGS moves in a different direction, using 3D Gaussian Splatting to disentangle geometry, albedo, and shadow-consistent illumination in multi-temporal satellite reconstruction (Luo et al., 4 Jan 2026).

This suggests that AeroDeshadow is best understood as an optical ASI framework centered on synthetic paired supervision and explicit penumbra modeling, rather than as a training-free single-image method or a 3D inverse-rendering system. Its nearest conceptual neighbors outside remote sensing are shadow-removal models that challenge hard-mask assumptions, such as DeS3’s adaptive attention for soft and self shadows (Jin et al., 2022), but AeroDeshadow relocates that concern to the much larger spatial scales and radiometric constraints of aerospace imagery.

The paper is relatively light on an explicit limitations section, but several limitations are implicit. The framework depends on shadow masks and morphological decomposition; its physical shadow model is still simplified, primarily a linear channel-wise attenuation plus bias; it is targeted at cast shadows in optical ASI and is less directly framed for cloud shadows or stronger atmospheric interactions; and the two-stage design increases system complexity. The authors accordingly identify two future directions: integrating deshadowing into downstream ASI tasks such as object detection and segmentation, and extending the physical degradation model to more complex atmospheric occlusions, especially cloud shadows (Lu et al., 17 Apr 2026).