HaLoBuilding: Optical Benchmark for Building Extraction

• HaLoBuilding is an optical benchmark that pairs degraded remote sensing images with clear-weather references to enable precise annotation transfer.
• The benchmark categorizes two degradation regimes—HaLo-L for low-light and HaLo-H for haze—to tackle challenges like low SNR, texture loss, and blurred boundaries.
• HaLoBuild-Net utilizes joint spatial-frequency modeling and dedicated semantic modules to enhance building extraction and ensure robust generalization across datasets.

HaLoBuilding is an optical benchmark for building extraction from remote sensing imagery under hazy and low-light conditions, introduced together with HaLoBuild-Net, an end-to-end baseline framework for adverse remote-sensing scenarios (Sang et al., 16 Apr 2026). It is defined by a same-scene multitemporal pairing strategy in which each degraded image is paired with a clear-weather reference from the same coordinates, enabling annotation transfer, manual refinement, and final masks with pixel-level alignment. Within the same study, HaLoBuild-Net combines spatial-frequency modeling and semantic calibration to mitigate meteorological interference, preserve building topology, and sharpen weather-induced blurred boundaries, while also reporting robust generalization on WHU, INRIA, and LoveDA (Sang et al., 16 Apr 2026).

1. Problem setting and benchmark scope

Building extraction from optical remote sensing imagery suffers from performance degradation under real-world hazy and low-light conditions, whereas existing optical methods and benchmarks focus primarily on ideal clear-weather conditions (Sang et al., 16 Apr 2026). The study situates this limitation against an alternative sensing modality: while SAR offers all-weather sensing, its side-looking geometry causes geometric distortions. HaLoBuilding is therefore framed as an optical benchmark rather than a SAR benchmark, with the explicit objective of preserving true building footprints under adverse visual conditions.

The benchmark is organized around two degradation regimes. The HaLo-L subset targets low-light imagery, where illumination spans dim twilight to near-dark and causes severe SNR drop, texture loss, and halo effects. The HaLo-H subset targets real haze densities ranging from thin mist to dense fog, where the dominant degradations are low contrast, color shift, and non-uniform scattering. This formulation defines adverse-weather building extraction as a joint problem of semantic segmentation, degradation robustness, and geometric fidelity rather than as a conventional clear-scene building-mapping task.

A common misconception is that building extraction in adverse weather can be reduced to a generic restoration pre-processing step followed by a standard segmentation model. The reported comparisons instead treat the adverse conditions as a distinct perception regime, motivating a dedicated benchmark and an end-to-end model optimized directly for the extraction objective.

2. Dataset construction, annotation workflow, and quality control

HaLoBuilding is built from GF2-PMS and GF7-DLC nadir-view optical satellites, with acquisitions from 2021–2023 across 10+ Chinese provinces, including Anhui, Gansu, Ningxia, Hebei, Hunan, and Yunnan, and covering urban, rural, and coastal scenes (Sang et al., 16 Apr 2026). The central design choice is the same-scene multitemporal pairing strategy.

The annotation workflow has two stages. In Stage 1, for each degraded image, either haze or low-light, a clear-weather reference captured within 1–3 months at the same coordinates is located. Rigid registration and cropping then yield aligned 1024×1024 RGB patches. In Stage 2, buildings are annotated on the clear reference, the masks are transferred to the paired degraded image, and the transferred labels are manually corrected where temporal changes or visibility differences introduce discrepancies.

This procedure is coupled to explicit quality control. Automatic transfer of clear-scene annotations is followed by meticulous manual correction on misaligned regions, including road changes, new structures, and building occlusions. The final masks guarantee pixel-level alignment, and no geometric correction is needed since nadir optical view preserves true footprints. This point is methodologically important: label alignment is not treated as a secondary post-processing issue but as a primary property of the benchmark design.

The released benchmark contains 4386 image pairs, each consisting of a 1024×1024 PNG or TIFF image and a binary mask. The split is 70% train, 10% val, and 20% test, applied independently for HaLo-L and HaLo-H, and the data are released in standard remote-sensing data format as GeoTIFF with accompanying metadata. The source code and datasets are publicly available at https://github.com/AeroVILab-AHU/HaLoBuilding.

3. Degradation regimes and data characteristics

HaLoBuilding separates adverse conditions into two subsets with distinct visual statistics and failure modes (Sang et al., 16 Apr 2026).

Subset	Size	Defining characteristics
HaLo-L	2514 images	Dim twilight to near-dark; severe SNR drop, texture loss, halo effects
HaLo-H	1872 images	Thin mist to dense fog; low contrast, color shift, non-uniform scattering

KDE analysis of pixel-value histograms provides a quantitative characterization of these regimes. Low-light images exhibit a shift toward the dark tail, described as left skew, together with intensity compression. Haze images exhibit a shift toward the bright tail, described as right skew, together with peak flattening. These distributional shifts indicate that the two degradations do not merely reduce visibility in a generic sense; they alter image statistics in opposite directions.

The low-light regime is associated with weak signal and suppressed texture cues, which directly affects boundary visibility and small-structure recoverability. The haze regime is associated with attenuated contrast and scattering-induced blur, which affects both local boundary sharpness and larger-scale structural continuity. This suggests that a single adverse-weather benchmark can still contain materially different feature-degradation mechanisms, and that architecture design must address both spatial discontinuity and unreliable high-frequency content.

4. HaLoBuild-Net architecture

HaLoBuild-Net is an end-to-end framework whose encoder uses an LWGANet-L2 backbone to produce four feature levels at $\tfrac14$, $\tfrac18$, $\tfrac1{16}$, and $\tfrac1{32}$ resolution, followed by a decoder with three cascaded stages (Sang et al., 16 Apr 2026). At each decoder stage, three components are applied: a Mutual-Guided Fusion Module (MGFM), a Global Multi-scale Guidance Module (GMGM), and an SFF Block consisting of a Spatial-Frequency Focus Module (SFFM) plus a lightweight MLP.

Module	Purpose	Key mechanism
MGFM	Suppress shallow noise and sharpen deep blurred boundaries	Bidirectional semantic–spatial calibration
GMGM	Anchor building topology	Global semantic constraints from shallow features
SFFM	Repair geometric discontinuities and suppress unreliable high-frequency noise	Large receptive field attention with frequency-aware channel reweighting

MGFM operates on encoder feature $F_e$ and decoder feature $F_d$. It first performs channel recalibration,

$$\tilde F_e = F_e \odot \sigma(Linear(GAP(F_e))), \qquad \tilde F_d = F_d \odot \sigma(Linear(GAP(F_d))),$$

then applies a Dual-stream Correction Module,

$$F'_e = V_e + (Q_d \odot K_e), \qquad F'_d = V_d + (Q_e \odot K_d),$$

followed by

$$F_\mathrm{DCM} = Conv_{1\times1}([F'_e;F'_d]), \qquad F_\mathrm{corr} = \tilde F_e \odot \tilde F_d,$$

and a residual fusion,

$$F_\mathrm{out} = FuseBlock([F_e; F_d; F_\mathrm{corr}; F_\mathrm{DCM}]).$$

Its stated role is to suppress shallow noise in low-light imagery and sharpen deep blurred boundaries in haze through bidirectional attention.

GMGM computes a global semantic map from shallow encoder features $\tfrac18$0 in order to anchor building topology. It aggregates pooled descriptors,

$\tfrac18$1

derives cross-scale attention weights,

$\tfrac18$2

forms

$\tfrac18$3

projects back to each scale,

$\tfrac18$4

and injects $\tfrac18$5 into the corresponding decoder stage by channel-wise concatenation. The module therefore imposes global topological priors rather than relying exclusively on local feature evidence.

SFFM is the architectural core for joint spatial-frequency modeling. In the spatial branch,

$\tfrac18$6

$\tfrac18$7

$\tfrac18$8

In the frequency branch,

$\tfrac18$9

$\tfrac1{16}$0

The final recalibration is

$\tfrac1{16}$1

followed by residual addition, MLP, and normalization. The module is explicitly designed to couple large receptive field attention with frequency-aware channel reweighting guided by stable low-frequency anchors.

5. Training protocol and empirical performance

Training uses random scaling and flipping, with crops to 512×512, and the optimizer is AdamW + Lookahead with batch size 16 and 100 epochs (Sang et al., 16 Apr 2026). The learning rate starts at $\tfrac1{16}$2, is cosine-annealed, and uses weight decay 0.01. The loss is

$\tfrac1{16}$3

Evaluation uses Intersection over Union, F1-score, Precision, and Recall for the building class, and multi-class metrics on LoveDA.

On the HaLoBuilding test sets, HaLoBuild-Net has 23.9 M parameters and achieves the strongest reported results among the listed methods.

Method	HaLo-L IoU / F1 (%)	HaLo-H IoU / F1 (%)
LWGANet-L2	66.81 / 80.11	68.84 / 81.54
DecoupleNet-D2	66.56 / 79.92	68.11 / 81.03
HaLoBuild-Net	68.90 / 81.59	70.88 / 82.96

The corresponding Precision and Recall for HaLoBuild-Net are 84.16 and 79.10 on HaLo-L, and 86.92 and 79.35 on HaLo-H. The paper’s key insight is that HaLoBuild-Net substantially closes the performance gap under extreme haze and low-light, boosting IoU by 2–3% over strong baselines.

The study also compares end-to-end learning with cascaded restoration-segmentation. The cascaded alternatives are GSAD low-light enhancement or DEA-Net dehazing followed by segmentation. HaLoBuild-Net improves over its cascaded counterpart by approximately 0.9% IoU on haze, while on low-light the cascaded strategy actually drops performance due to hallucinations and over-smoothness. This reported behavior is central to the interpretation of the framework: the model is not merely a segmentation head attached to restored images, but a direct extractor whose internal modules are optimized to avoid artifact propagation.

6. Generalization, interpretation, and limitations

HaLoBuild-Net is evaluated beyond HaLoBuilding on clear-weather and multi-class datasets (Sang et al., 16 Apr 2026). On WHU it reports 91.88 IoU and 95.77 F1, exceeding BOMSC-Net at 90.15/94.80 and LOGCAN++ at 91.51/95.57. On INRIA it reports 82.53 IoU and 90.43 F1, exceeding BOMSC-Net at 78.18/87.75 and LOGCAN++ at 82.13/90.19. On LoveDA it achieves the highest mIoU of 54.04%, building IoU of 60.25%, and best road/background separation of 59.81% and 47.72%.

These results are interpreted in the study as evidence that end-to-end joint spatial-frequency modeling through SFFM, together with semantic calibration through MGFM and GMGM, is superior to cascaded restoration followed by segmentation because it avoids artifact propagation. The reported generalization to WHU, INRIA, and LoveDA further indicates that the dual-domain priors do not degrade, but rather sharpen, boundary and structural fidelity in all scenarios. A plausible implication is that adverse-weather priors, when embedded directly into the segmentation pipeline, can regularize topology and boundaries even on clear-scene data.

The limitations are also explicit. The model size, approximately 24 M parameters, is non-trivial for real-time or edge deployment, and future work is directed toward compressing or distilling the framework. Additional directions include extension to multi-class segmentation in adverse weather, vector-level boundary recovery, and other extreme conditions such as rain, snow, and sandstorms. The longest-range objective stated in the work is the development of a unified all-weather foundation model for remote-sensing perception.