HaLoBuild-Net: Hazy/Low-Light Building Extraction

• HaLoBuild-Net is a single-stage encoder–decoder framework that extracts building footprints from degraded optical remote sensing images under hazy and low-light conditions.
• It integrates spatial-frequency focus, global multi-scale guidance, and mutual-guided fusion modules to jointly optimize restoration and segmentation without cascading separate enhancement steps.
• Empirical evaluations on the HaLoBuilding benchmark demonstrate significant improvements in IoU and F1 scores over baselines, confirming robust performance in adverse conditions.

HaLoBuild-Net is a single-stage encoder–decoder framework for building extraction from optical remote sensing imagery under hazy and low-light conditions, introduced together with HaLoBuilding, the first optical benchmark specifically designed for this setting (Sang et al., 16 Apr 2026). The framework addresses a regime in which conventional clear-weather optical benchmarks and models exhibit severe degradation, while SAR-based alternatives preserve all-weather sensing at the cost of layover and shadowing that compromise footprint fidelity. In this formulation, the central problem is not geometric correction but signal and semantic recovery from degraded nadir-view optical observations.

1. Problem setting and motivation

Building extraction from optical remote sensing imagery degrades substantially under haze and low light because the underlying image statistics shift in ways that directly damage boundary localization and small-target detectability (Sang et al., 16 Apr 2026). Under haze, atmospheric scattering elevates background luminance and suppresses contrast, washing out edges and textures. Under low light, signal-to-noise ratio is reduced, sensor noise is amplified, and dynamic range is compressed, causing boundary diffusion and small-target omission.

The HaLoBuilding study characterizes this degradation through Kernel Density Estimation analyses. The haze subset exhibits a rightward intensity shift with peak attenuation, corresponding to brighter backgrounds and contrast loss. The low-light subset exhibits severe leftward skew, corresponding to signal suppression. The benchmark therefore targets a severe domain shift without relying on synthetic degradation.

The paper positions this setting against two common alternatives. First, mainstream optical building datasets such as WHU, INRIA, Massachusetts, OpenEarthMap, and LoveDA are predominantly clear-weather, and models trained on them show catastrophic generalization failures under haze and nighttime conditions. Second, while SAR is all-weather, its side-looking geometry introduces layover and shadowing that displace rooftops from footprints, which is problematic for footprint extraction. Optical imagery, even when degraded, preserves nadir-view geometric correctness.

A further design decision follows from this diagnosis. Cascaded restoration–segmentation pipelines are described as imposing disjoint objectives, propagating enhancement artifacts, blurring edges via denoising, and incurring extra computation. In low-light conditions, enhancement can hallucinate structures and induce domain shift for the segmentor. HaLoBuild-Net therefore adopts an end-to-end, task-oriented formulation that learns directly from degraded observations.

2. HaLoBuilding benchmark

HaLoBuilding is built from real optical imagery from GF2-PMS and GF7-DLC acquired during 2021–2023, with geographic coverage spanning Anhui, Gansu, Ningxia, Shaanxi, Hebei, Shandong, Jiangsu, Hunan, Hubei, and Yunnan, and including diverse urban, rural, and coastal scenes (Sang et al., 16 Apr 2026). Each patch is a 1024×1024 RGB image, and the native ground sampling distance is about 0.8 m.

The benchmark contains two subsets totaling 4,386 images. HaLo-L contains 2,514 low-light images, and HaLo-H contains 1,872 hazy images. The split is 7:1:2 for train, validation, and test.

A distinctive feature is the same-scene multitemporal pairing strategy. For each degraded image, a clear-image reference of the same scene within a 1–3 month window is located using geographic coordinate metadata and precisely aligned and cropped to 1024×1024. Building masks are delineated on the clear reference and transferred to the degraded image. Annotators then manually rectify inconsistencies, including temporal changes and visibility-driven omissions, to guarantee pixel-accurate alignment with the actual visible buildings in the degraded image. Registration is georegistration-based followed by precise patch alignment, and extreme degradation is handled by conservative manual curation so that labels reflect true visible footprints.

The evaluation protocol is standard semantic segmentation on the building class. Reported metrics are Intersection over Union, F1/Dice score, Precision, and Recall, with results reported separately on HaLo-L and HaLo-H test splits.

3. Architectural design

HaLoBuild-Net is organized as a single-stage encoder–decoder with three core modules: the Spatial-Frequency Focus Module (SFFM), the Global Multi-scale Guidance Module (GMGM), and the Mutual-Guided Fusion Module (MGFM) (Sang et al., 16 Apr 2026). The encoder is LWGANet-L2, a lightweight group attention backbone tailored for remote sensing, producing features at $1/4$, $1/8$, $1/16$, and $1/32$ resolution. The backbone has 13.0M parameters.

The SFFM performs dual-domain collaborative optimization. Let $$F_{\mathrm{fused}} \in \mathbb{R}^{B \times C \times H \times W}$$ denote the decoder-stage input. Its spatial branch combines local and contextual receptive fields:

$$A_1 = \mathrm{Conv}_{1 \times 1}(\mathrm{DWConv}_{5 \times 5}(F_{\mathrm{fused}})), \qquad A_2 = \mathrm{Conv}_{1 \times 1}(\mathrm{DWConv}_{7 \times 7, d=3}(A_1)).$$

After concatenation, channel-wise average and max pooling, a $7 \times 7$ convolution, and sigmoid activation, a 2-channel spatial attention map is obtained:

$$M_{\mathrm{spatial}} = \sigma\!\left(\mathrm{Conv}_{7 \times 7}(\mathrm{Concat}_c(\mathrm{Avg}(A), \mathrm{Max}(A)))\right).$$

Splitting $M_{\mathrm{spatial}}$ into $S_1$ and $1/8$0, the recalibrated spatial response is

$1/8$1

This branch is intended to recover weather-blurred boundaries and large buildings through an enlarged effective receptive field.

The frequency branch computes a 2D FFT per channel, centers the spectrum, and crops a central low-frequency band:

$1/8$2

A channel descriptor is then formed from the magnitude spectrum:

$1/8$3

The output is

$1/8$4

The implementation uses a central box crop of size $1/8$5, reported as optimal in ablations.

GMGM provides global semantic constraints when local cues are unreliable. With shallow encoder features $1/8$6, it aggregates global context by

$1/8$7

learns scale weights $1/8$8, aligns features to the highest resolution by bilinear upsampling, and redistributes the resulting guidance to the original scales. The stated purpose is to anchor building topology and inject a scene-level prior into decoder stages.

MGFM performs bidirectional semantic–spatial calibration between encoder feature $1/8$9 and decoder feature $1/16$0 at the same resolution. After channel-attention pre-calibration,

$1/16$1

the module applies dual-stream correction:

$1/16$2

with $1/16$3, $1/16$4, and $1/16$5 realized by learned $1/16$6 projections. The corrected features are fused, and a correlation path

$1/16$7

is included before lightweight fusion. The stated role is to use deep semantics to suppress shallow noise and use shallow spatial anchors to sharpen blurred deep boundaries.

4. Optimization and training protocol

The segmentation objective is a weighted sum of cross-entropy and Dice loss (Sang et al., 16 Apr 2026):

$1/16$8

with $1/16$9 in practice. The soft cross-entropy term is

$1/32$0

and the Dice term is

$1/32$1

The paper explicitly states that it does not use explicit boundary or auxiliary losses; guidance is injected through GMGM and MGFM within the main head.

Training uses random scaling and flipping, followed by cropping from 1024×1024 to 512×512. The batch size is 16, the training schedule is 100 epochs, the optimizer is AdamW with Lookahead, the initial learning rate is $1/32$2, the schedule is cosine annealing, the weight decay is 0.01, and the fixed seed is 42. No explicit haze or low-light augmentations are used because the degradations are real and diverse.

At 512×512 input, HaLoBuild-Net totals 23.9M parameters and 90.3 GFLOPs. The paper characterizes this as a lightweight-moderate computational profile and identifies the chosen LWGANet-L2 encoder as the accuracy–efficiency balance. For restoration–segmentation comparisons, GSAD is used for low-light enhancement and DEA-Net for dehazing, with restored images fed into the same segmentors and results averaged over three runs.

5. Empirical performance

On HaLo-L, HaLoBuild-Net achieves IoU 68.90%, F1 81.59%, Precision 84.16%, and Recall 79.10% (Sang et al., 16 Apr 2026). The reported best baseline encoder, LWGANet-L2, achieves IoU 66.81%, F1 80.11%, Precision 82.68%, and Recall 77.68%, so the gains are +2.09 IoU, +1.48 F1, and +1.42 Recall. Other strong baselines include DecoupleNet-D2 with IoU 66.56 and F1 79.92, and UNetFormer with IoU 66.40 and F1 79.81.

On HaLo-H, HaLoBuild-Net achieves IoU 70.88%, F1 82.96%, Precision 86.92%, and Recall 79.35%. LWGANet-L2 achieves IoU 68.84%, F1 81.54%, Precision 87.93%, and Recall 76.02%, corresponding to gains of +2.04 IoU and +1.42 F1. DecoupleNet-D2 reaches IoU 68.11 and F1 81.03, while UNetFormer reaches IoU 67.92 and F1 80.90.

The qualitative analysis emphasizes condition-specific behavior. Under low light, the model maintains complete building interiors, separates buildings from texture-similar roads and grass, and detects small targets amidst weak cues. Under haze, it suppresses confusion with greenhouses and roads and recovers low-contrast buildings via global guidance.

A central empirical claim concerns the comparison between direct end-to-end learning and cascaded enhancement. On HaLo-L, direct HaLoBuild-Net yields IoU 68.90 and F1 81.59, while HaLoBuild-Net* with GSAD→segmentation yields IoU 67.53 and F1 80.62, a drop of $1/32$3 IoU. On HaLo-H, direct HaLoBuild-Net yields IoU 70.88 and F1 82.96, while HaLoBuild-Net* with DEA-Net→segmentation yields IoU 69.95 and F1 82.32. The paper attributes this gap to enhancement-induced hallucinations, over-smoothed edges, and shifted statistics.

The model also generalizes to standard clear-weather datasets. On WHU, it achieves IoU 91.88% and F1 95.77%, exceeding LOGCAN++ at IoU 91.51% and F1 95.57%, and BCTNet/FD-Net at IoU about 91.1% and F1 about 95.36–95.37%. On INRIA, it achieves IoU 82.53% and F1 90.43%, exceeding LOGCAN++ at IoU 82.13% and F1 90.19%, and HD-Net at IoU 82.10% and F1 90.17%. On LoveDA, the building IoU is 60.25% and mIoU is 54.04%, with per-class IoU values of background 47.72, building 60.25, road 59.81, water 81.15, barren 20.31, forest 45.75, and agricultural 63.29.

6. Ablation results, limitations, and outlook

The ablation study decomposes performance gains by module and configuration (Sang et al., 16 Apr 2026). Relative to baseline UNetFormer, SFFM contributes +0.88 IoU on HaLo-L and +0.94 on HaLo-H, GMGM contributes +1.25 and +1.04, and MGFM contributes +1.68 and +1.71, making MGFM the largest single-module gain. Combinations also matter: SFFM + MGFM gives +2.46 IoU on HaLo-H, SFFM + GMGM gives +1.37 IoU on HaLo-L, and the full model gives +2.50 on HaLo-L and +2.96 on HaLo-H.

SFFM path decoupling identifies complementary behavior. The spatial-only version $1/32$4 yields IoU 68.24 on HaLo-L and 68.59 on HaLo-H. The frequency-only version $1/32$5 yields 67.84 and 70.03. Full SFFM yields 68.90 and 70.88. The paper concludes that the spatial path favors low-light structure recovery, while the frequency path better combats haze scattering.

The low-frequency crop-size study reports $1/32$6 at 68.13/70.19, $1/32$7 at 66.15/70.43, $1/32$8 at 68.90/70.88, and $1/32$9 at 68.19/70.06 for HaLo-L/HaLo-H. The stated interpretation is that the full spectrum is contaminated by noisy high frequencies, $$F_{\mathrm{fused}} \in \mathbb{R}^{B \times C \times H \times W}$$0 is too broad and includes unstable bands, and $$F_{\mathrm{fused}} \in \mathbb{R}^{B \times C \times H \times W}$$1 is too small and loses topology.

Backbone analysis reports IoU values of 59.36 and 65.81 for ResNet-18, 67.90 and 65.65 for ResNet-34, 66.81 and 67.35 for MobileNetV2-1.4, and 68.90 and 70.88 for LWGANet-L2 on HaLo-L and HaLo-H, respectively. The paper attributes the best trade-off to group attention and multi-scale sensitivity.

The study also specifies several limitations. Extremely dense high-rise clusters with complex shadows may still induce local confusion without additional 3D priors. Glass or reflective roofs and illumination bloom can perturb frequency cues. Performance depends on image resolution, and very coarse ground sampling distance reduces the reliability of the FFT anchor and shallow anchors. Computationally, 23.9M parameters and 90.3 GFLOPs are described as lightweight-moderate rather than minimal.

Future work is framed around temporal fusion, SAR–optical synergy, domain adaptation, and broader adverse scenarios such as rain, snow, sandstorm, and general multi-class all-weather segmentation. The code and dataset are publicly available at the project repository given in the paper.