cs.CVOct 16, 2026

ShadowCast: On-the-Fly Context-Aware Shadow Augmentation for Remote Sensing Image Segmentation

Authors: Dongho Yoon, Yechan Kim, Jongmin Joo, JongHyun Park, Moongu Jeon

Organizations: Gwangju Institute of Science and Technology (GIST), 123 Cheomdangwagi-ro, Buk-gu, Jeonnam-Gwangju, 61005, Korea · AI Laboratory, LIG Defense&Aerospace, 333 Pangyo-ro, Bundang-gu, Seongnam-si, Gyeonggi-do, 13488, Korea

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Shadows cast by elevated objects alter the spectral appearance and obscure the boundaries of objects in remote sensing images, thereby degrading semantic segmentation performance. Most existing approaches handle shadows as degradations to be detected and removed, requiring additional costs from dedicated networks and shadow-specific annotations. We propose ShadowCast, which takes the opposite approach: it synthesizes shadows and injects them during training for remote sensing segmentation. It extracts masks of elevated objects, such as buildings and trees, directly from the segmentation labels and estimates an image-specific cast-shadow direction cue by probing the brightness around them. It then projects synthetic shadows along the inferred cast-shadow direction, requiring no learned network and no acquisition metadata. Experiments on three widely used remote sensing benchmarks, Vaihingen, Potsdam, and LoveDA, show that ShadowCast consistently outperforms the baseline across diverse state-of-the-art segmentation architectures. Moreover, it brings additional performance gains when combined with standard augmentation and regularization techniques, indicating that stochastic shadow injection serves as a simple and complementary augmentation strategy for remote sensing segmentation.

Figures & tables

Fig. 1: Pipeline of our ShadowCast for remote sensing image segmentation.
Fig. 2: Behaviors of ShadowCast under varying values of θ⋆\theta^{\star} , dd , and δ\delta .
Cast-shadow directionVaihingenPotsdamLoveDA
No shadow augmentation77.53 ±\pm 0.3082.95 ±\pm 0.1043.29 ±\pm 0.77
Random direction78.14 ±\pm 0.1083.13 ±\pm 0.1443.61 ±\pm 0.50
Estimated direction78.20 ±\pm 0.0783.22 ±\pm 0.1344.45 ±\pm 0.30
+ Jittering ( Ours )78.46 ±\pm 0.1983.32 ±\pm 0.0944.89 ±\pm 0.09
TABLE I: Effect of cast-shadow direction selection in ShadowCast across three benchmarks using DeepLabV3+ ( p=0.5p=0.5 ; dmin⁡=5d_{\min}=5 ; dmax⁡=30d_{\max}=30 ; δmin⁡=0.5\delta_{\min}=0.5 ; δmax⁡=0.8\delta_{\max}=0.8 ; η=15\eta=15 ; τ=50\tau=50 ; r=8r=8 ).
ModelVaihingenPotsdamLoveDA
BaselineShadowCastBaselineShadowCastBaselineShadowCast
DeepLabV3+ [ 20 ]77.53 ±\pm 0.30 (+0.00)78.46 ±\pm 0.19 ( +0.93 )82.95 ±\pm 0.10 (+0.00)83.32 ±\pm 0.09 ( +0.37 )43.29 ±\pm 0.77 (+0.00)44.89 ±\pm 0.09 ( +1.60 )
DANet [ 21 ]79.59 ±\pm 0.61 (+0.00)80.20 ±\pm 0.06 ( +0.61 )84.47 ±\pm 0.24 (+0.00)84.88 ±\pm 0.04 ( +0.41 )43.35 ±\pm 0.78 (+0.00)44.43 ±\pm 0.85 ( +1.08 )
APCNet [ 22 ]79.15 ±\pm 0.85 (+0.00)80.53 ±\pm 0.18 ( +1.38 )84.73 ±\pm 0.13 (+0.00)85.05 ±\pm 0.01 ( +0.32 )46.46 ±\pm 0.60 (+0.00)47.94 ±\pm 0.17 ( +1.48 )
OCRNet [ 23 ]75.39 ±\pm 0.79 (+0.00)77.88 ±\pm 0.43 ( +2.49 )82.62 ±\pm 0.11 (+0.00)83.03 ±\pm 0.12 ( +0.41 )44.26 ±\pm 0.66 (+0.00)44.84 ±\pm 0.32 ( +0.58 )
InternImage [ 24 ]81.95 ±\pm 0.11 (+0.00)82.35 ±\pm 0.05 ( +0.40 )86.22 ±\pm 0.08 (+0.00)86.35 ±\pm 0.00 ( +0.13 )51.87 ±\pm 0.18 (+0.00)53.11 ±\pm 0.20 ( +1.24 )
TABLE II: Impact of ShadowCast on training state-of-the-art segmentation models on Vaihingen, Potsdam, and LoveDA.
AugmentationVaihingenPotsdamLoveDA
No shadow augmentation77.53 ±\pm 0.3082.95 ±\pm 0.1043.29 ±\pm 0.77
RandomShadow [ 13 ]78.36 ±\pm 0.0483.15 ±\pm 0.1243.67 ±\pm 0.40
PlasmaShadow [ 13 ]76.98 ±\pm 0.1982.37 ±\pm 0.1144.01 ±\pm 0.17
SynShadow [ 30 ]78.18 ±\pm 0.0583.11 ±\pm 0.1443.58 ±\pm 0.33
SDAA [ 31 ]77.87 ±\pm 0.0883.21 ±\pm 0.1243.12 ±\pm 0.60
Zhong et al. [ 32 ]78.07 ±\pm 0.1983.24 ±\pm 0.1243.46 ±\pm 0.56
TABLE III: Comparison of ShadowCast with additional shadow augmentation baselines using DeepLabV3+ on Vaihingen, Potsdam, and LoveDA.
Existing methodVaihingenPotsdamLoveDA
BaselineShadowCastBaselineShadowCastBaselineShadowCast
Horizontal Flip [00]78.62 ±\pm 0.19 (+0.00)79.08 ±\pm 0.06 ( +0.46 )84.00 ±\pm 0.12 (+0.00)84.32 ±\pm 0.08 ( +0.32 )43.81 ±\pm 0.60 (+0.00)45.88 ±\pm 0.52 ( +2.07 )
Random Resize [00]80.04 ±\pm 0.38 (+0.00)80.36 ±\pm 0.02 ( +0.32 )84.37 ±\pm 0.13 (+0.00)84.50 ±\pm 0.03 ( +0.13 )45.11 ±\pm 0.31 (+0.00)46.70 ±\pm 0.97 ( +1.59 )
Photometric Distortion [00]77.33 ±\pm 0.35 (+0.00)77.86 ±\pm 0.12 ( +0.53 )82.76 ±\pm 0.13 (+0.00)83.13 ±\pm 0.05 ( +0.37 )47.19 ±\pm 0.64 (+0.00)48.45 ±\pm 0.21 ( +1.26 )
CLAHE [ 36 ] [00]78.33 ±\pm 0.12 (+0.00)78.46 ±\pm 0.04 ( +0.13 )83.13 ±\pm 0.25 (+0.00)83.30 ±\pm 0.27 ( +0.17 )43.92 ±\pm 0.78 (+0.00)44.59 ±\pm 0.49 ( +0.67 )
CutOut [ 14 ] [00]78.29 ±\pm 0.69 (+0.00)79.18 ±\pm 0.30 ( +0.89 )83.27 ±\pm 0.03 (+0.00)83.56 ±\pm 0.07 ( +0.29 )44.54 ±\pm 0.46 (+0.00)46.27 ±\pm 0.75 ( +1.73 )
TABLE IV: Impact of combining ShadowCast with existing augmentation/regularization methods on DeepLabV3+ for Vaihingen, Potsdam, and LoveDA.
DatasetMethodShadowedNon-shadowed
VaihingenBaseline57.88 ±\pm 0.9582.77 ±\pm 0.37
 ShadowCast58.18 ±\pm 0.48 ( +0.30 )83.26 ±\pm 0.33 ( +0.49 )
PotsdamBaseline81.85 ±\pm 0.9381.50 ±\pm 3.00
 ShadowCast82.27 ±\pm 0.13 ( +0.42 )82.85 ±\pm 2.07 ( +1.35 )
LoveDABaseline26.83 ±\pm 1.3438.86 ±\pm 1.48
 ShadowCast27.14 ±\pm 0.71 ( +0.31 )39.41 ±\pm 1.13 ( +0.55 )
TABLE V: Evaluation of ShadowCast using DeepLab V3+ across shadowed and non-shadowed regions of Vaihingen, Potsdam, and LoveDA (Ten shadow-annotated test images per dataset).
(dmin⁡,dmax⁡)(d_{\min},d_{\max})(δmin⁡,δmax⁡)(\delta_{\min},\delta_{\max})η\etaτ\taurrmIoU
(5, 30)(0.5, 0.8)1550878.46 ±\pm 0.19
(5, 20)(0.5, 0.8)1550878.25 ±\pm 0.33
(5, 40)(0.5, 0.8)1550878.22 ±\pm 0.18
(5, 30)(0.4, 0.7)1550878.35 ±\pm 0.12
(5, 30)(0.6, 0.9)1550878.10 ±\pm 0.36
(5, 30)(0.5, 0.8)050878.20 ±\pm 0.07
TABLE S1: Parameter sensitivity of ShadowCast on the ISPRS Vaihingen benchmark using DeepLabV3+ ( p=0.5p=0.5 ).
(dmin⁡,dmax⁡)(d_{\min},d_{\max})(δmin⁡,δmax⁡)(\delta_{\min},\delta_{\max})η\etaτ\taurrmIoU
(5, 30)(0.5, 0.8)1550883.32 ±\pm 0.09
(5, 20)(0.5, 0.8)1550883.24 ±\pm 0.13
(5, 40)(0.5, 0.8)1550883.27 ±\pm 0.11
(5, 30)(0.4, 0.7)1550883.26 ±\pm 0.12
(5, 30)(0.6, 0.9)1550883.18 ±\pm 0.11
(5, 30)(0.5, 0.8)050883.22 ±\pm 0.13
TABLE S2: Parameter sensitivity of ShadowCast on the ISPRS Potsdam benchmark using DeepLabV3+ ( p=0.5p=0.5 ).
(dmin⁡,dmax⁡)(d_{\min},d_{\max})(δmin⁡,δmax⁡)(\delta_{\min},\delta_{\max})η\etaτ\taurrmIoU
(5, 30)(0.5, 0.8)1550844.89 ±\pm 0.09
(5, 20)(0.5, 0.8)1550844.15 ±\pm 0.23
(5, 40)(0.5, 0.8)1550844.20 ±\pm 0.21
(5, 30)(0.4, 0.7)1550844.29 ±\pm 0.47
(5, 30)(0.6, 0.9)1550844.21 ±\pm 0.23
(5, 30)(0.5, 0.8)050844.45 ±\pm 0.30
TABLE S3: Parameter sensitivity of ShadowCast on the LoveDA benchmark using DeepLabV3+ ( p=0.5p=0.5 ).
Fig. S1: Examples of the manually annotated real shadow regions ( magenta -colored) on (a) Vaihingen, (b) Potsdam, and (c) LoveDA.
Fig. S2: Shadows synthesized by different augmentation methods on the same Vaihingen training image. CLAHE [ 36 ] , which enhances the image rather than synthesizing shadows, is omitted.
ModelBackboneOptimizerInitial LRWeight decayMinimum LR
DeepLabV3+ResNet-18SGD0.015×10−45\times 10^{-4}10−410^{-4}
DANetResNet-50SGD0.015×10−45\times 10^{-4}10−410^{-4}
APCNetResNet-50SGD0.015×10−45\times 10^{-4}10−410^{-4}
OCRNetHRNet-W18-SmallSGD0.015×10−45\times 10^{-4}10−410^{-4}
InternImageInternImage-TSGD0.015×10−45\times 10^{-4}10−410^{-4}
DOCNetHRNet-W32AdamW10−410^{-4}10−410^{-4}0
TABLE S4: Training configurations of the segmentation models
MethodImplementationParameters
Horizontal FlipMMSegp=0.5p=0.5
Random ResizeMMSegScale ratio ∼U(0.5,2.0)\sim\mathcal{U}(0.5,2.0) relative to 512×512512\times 512 ( 1024×10241024\times 1024 for LoveDA); aspect ratio preserved
Photometric DistortionMMSegBrightness shift ∼U(−32,32)\sim\mathcal{U}(-32,32) ; contrast and saturation factors ∼U(0.5,1.5)\sim\mathcal{U}(0.5,1.5) ; integer hue shift ∈{−18,…,17}\in\{-18,\ldots,17\} (OpenCV hue scale); each applied with probability 0.5
CLAHE [ 36 ]Albumentationsp=0.5p=0.5 ; clip limit ∼U(1,4)\sim\mathcal{U}(1,4) ; 8×88\times 8 tiles; applied to the CIELAB lightness channel
CutOut [ 14 ]MMSegp=0.5p=0.5 ; 5-10 holes of 16×1616\times 16 or 32×3232\times 32 pixels; image filled with 0 and labels with the ignore index 255
CutMix [ 15 ]Customp=0.5p=0.5 ; patch area ratio ∼U(0.2,0.5)\sim\mathcal{U}(0.2,0.5) ; image and label patches from another training sample
TABLE S5: Settings of the augmentation and regularization methods
MethodImplementationParameters
ShadowCast (Ours)-p=0.5p=0.5 ; (dmin⁡,dmax⁡)=(5,30)(d_{\min},d_{\max})=(5,30) pixels; (δmin⁡,δmax⁡)=(0.5,0.8)(\delta_{\min},\delta_{\max})=(0.5,0.8) ; η=15∘\eta=15^{\circ} ; τ=50\tau=50 pixels; r=8r=8 pixels; E={Building,Tree}\mathcal{E}=\{\text{Building},\text{Tree}\} for Vaihingen and Potsdam and {Building,Forest}\{\text{Building},\text{Forest}\} for LoveDA
RandomShadow [ 13 ]Albumentationsp=0.5p=0.5 ; 1 or 2 triangles with vertices sampled over the whole image; shadow intensity U(0.5,0.8)\mathcal{U}(0.5,0.8)
PlasmaShadow [ 13 ]Albumentationsp=0.5p=0.5 ; roughness 3.0; plasma size 256; shadow intensity U(0.5,0.8)\mathcal{U}(0.5,0.8)
SynShadow [ 30 ]Portedp=0.5p=0.5 ; 10 000 rendered mattes of [ 30 ] ; l1∼U(0,0.25)l_{1}\sim\mathcal{U}(0,0.25) ; s1∼U(0.1,0.9)s_{1}\sim\mathcal{U}(0.1,0.9) ; channel offsets ∼N(0.05,0.0252)\sim\mathcal{N}(0.05,0.025^{2}) ; random horizontal flip; minimum matte coverage 0.005; up to 5 trials
SDAA [ 31 ]Re-implementedp=0.5p=0.5 ; mask M\mathbf{M} from E\mathcal{E} as in ShadowCast ; direction ∼U(0∘,360∘)\sim\mathcal{U}(0^{\circ},360^{\circ}) for each image; shift ∼U(5,30)\sim\mathcal{U}(5,30) pixels; gamma γ∈{0.800,0.825,…,0.975}\gamma\in\{0.800,0.825,\ldots,0.975\}
Zhong et al. [ 32 ]Portedp=0.5p=0.5 ; one triangle with vertices ∼U(−0.5,1.5)\sim\mathcal{U}(-0.5,1.5) in units of the image height and width; lightness scaled by k∼U(0.2,0.7)k\sim\mathcal{U}(0.2,0.7) ; 3-pixel edge blur; up to 10 trials
TABLE S6: Settings of the shadow augmentation methods
ClassVaihingenPotsdamClassLoveDA
BaselineShadowCastBaselineShadowCastBaselineShadowCast
Non-elevated classes 
Impervious surface83.08 ±\pm 0.1783.93 ±\pm 0.06 (+0.85)84.49 ±\pm 0.2584.87 ±\pm 0.14 (+0.38)Background49.30 ±\pm 0.7350.19 ±\pm 0.12 (+0.89)
Low vegetation68.44 ±\pm 0.5569.36 ±\pm 0.60 (+0.92)73.50 ±\pm 0.2573.82 ±\pm 0.05 (+0.32)Road48.40 ±\pm 0.9649.24 ±\pm 0.71 (+0.84)
Car69.48 ±\pm 0.6171.05 ±\pm 0.19 (+1.57)89.94 ±\pm 0.2490.35 ±\pm 0.26 (+0.41)Water47.89 ±\pm 1.7154.39 ±\pm 0.97 (+6.50)
     Barren19.48 ±\pm 0.9620.38 ±\pm 1.45 (+0.90)
TABLE S7: Per-group/Per-class mIoU of ShadowCast (DeepLabV3+).
 VaihingenPotsdamLoveDA
Baseline77.53 ±\pm 0.3082.95 ±\pm 0.1043.29 ±\pm 0.77
ShadowCast (Default: Hard-shadow) ( b=0b{=}0 )78.46 ±\pm 0.1983.32 ±\pm 0.0944.89 ±\pm 0.09
ShadowCast (Soft-shadow) ( b=1b{=}1 )77.79 ±\pm 0.1282.98 ±\pm 0.1343.36 ±\pm 0.24
TABLE S8: mIoU (%) of DeepLabV3+ with Hard ( b=0b{=}0 ) and Soft ( b=1b{=}1 ) Synthetic Shadows in ShadowCast .
Fig. S3: Hard ( b=0b=0 ) and soft ( b=1b=1 ) synthetic shadow boundaries on a Vaihingen patch. The bottom row enlarges the region marked by the yellow box.
MethodVaihingenPotsdamLoveDA
Baseline77.53 ±\pm 0.3082.95 ±\pm 0.1043.29 ±\pm 0.77
ShadowCast (Default)78.46 ±\pm 0.19 ( ++ 0.93)83.32 ±\pm 0.09 ( ++ 0.37)44.89 ±\pm 0.09 ( ++ 1.60)
ShadowCast (+ Confidence checks)78.18 ±\pm 0.57 ( ++ 0.65)83.22 ±\pm 0.04 ( ++ 0.27)43.90 ±\pm 0.30 ( ++ 0.61)
TABLE S9: Effect of the confidence-gated direction estimation ( τgap=0.03\tau_{\mathrm{gap}}=0.03 , τsym=0.10\tau_{\mathrm{sym}}=0.10 ) on DeepLabV3+ for the Vaihingen, Potsdam, and LoveDA benchmarks (mIoU, %)
Fig. S4: Representative failure cases of ShadowCast on training images from (a) Vaihingen, (b) Potsdam, and (c) LoveDA; see Gold bounding boxes. The examples show incorrect shadow length and shape, an unclear direction in an already shaded area, and misleading shadows from neighboring objects, respectively.
LimitationExampleObserved failureUnderlying cause
Shadow length and shapeFig. S4 (a)Shadows too short; shape does not follow the objectSame shift range for all object heights; single shift of the top-view mask
Unclear directionFig. S4 (b)Arbitrary direction in an already shaded areaNo brightness difference between sides
Neighboring shadowsFig. S4 (c)Wrong side appears darkestNo link between shadows and casters; averaging over all objects
TABLE S10: Categorization of the Failure Cases in Fig. S4
More in this topic

Remote Sensing Image Segmentation

Remote sensing image segmentation labels land cover and objects in satellite, aerial, radar, and multispectral imagery.

See all 220 papers

Explore similar work

CardsList

  1. SARU: A Shadow-Aware and Removal Unified Framework for Remote Sensing Images with New Benchmarks

    Apr 28, 2026Zi-Yang Bo, Wei Lu, Hongruixuan Chen +2Remote Sensing Image UnderstandingShadow Removal

    MOE Key Lab of ICSP, Anhui Provincial Key Lab of Multimodal Cognitive Computation, IMIS Lab of Anhui Province, School of Computer Science and Technology, Anhui University, Hefei, China · Graduate School of Frontier Sciences, The University of Tokyo, Chiba, 277-8561, Japan

  2. BRIC-Net: Boundary-Reliable Illumination-Color Interaction for Remote Sensing Image Deshadowing

    Aug 1, 2026Wei Lu, Yi Liu, Si-BaoShadow RemovalImage Enhancement

    MOE Key Lab of ICSP, IMIS Lab of Anhui, Anhui Provincial Key Lab of Multimodal Cognitive Computation, Zenmorn-AHU AI Joint Lab, School of Computer Science and Technology, Anhui University, Hefei 230601, China

  3. PDA++: Field-Aligned Planning and Scene-Adaptive Insertion in Remote Sensing

    Sep 16, 2026Xianchi Dong, Yingyan Hou, Chao Ren +6Remote Sensing Image UnderstandingSynthetic Data Augmentation