Seeing through fog and night: new neural network learns to extract building outlines from satellite images in bad weather

On April 16, 2026, Chinese researchers presented a practical solution to this problem. They created the first specialized benchmark HaLoBuilding (Hazy + Low-light) — a large set of satellite images of the same area taken at different times of day and in different weather. Thanks to this, the building outline labels turned out perfectly aligned even on heavily degraded images.

Based on this dataset, the authors developed the HaLoBuild-Net neural network. Its architecture features three key modules. The first focuses on stable low-frequency features and suppresses noise from fog and haze. The second helps maintain the overall structure of buildings at different scales. The third cleans up noise and refines blurred boundaries.

In tests, the new network significantly outperformed all existing methods, including cascaded schemes of "first improve the photo — then extract buildings." Additionally, HaLoBuild-Net performs well on other well-known datasets: WHU (images of Chinese cities), INRIA (aerial photos of France and Austria), and LoveDA (diverse landscapes).

The practical benefits are obvious and quite broad. Such technology enables faster and more accurate map creation in poor visibility conditions, monitoring new construction, assessing damage after floods, fires, or earthquakes, when ordinary optical images become almost useless. This is especially important for regions with frequent fogs, monsoon rains, or polar night. The authors immediately released the code and dataset into open access.

Of course, this is not a magic wand that will see absolutely everything in pitch darkness. But the step turned out to be noticeable and practical: the network immediately learns to work with real complex shooting conditions, rather than trying to artificially "clarify" the image first.

Source: Feifei Sang, Wei Lu, Hongruixuan Chen, Sibao Chen, Bin Luo. Building Extraction from Remote Sensing Imagery under Hazy and Low-light Conditions: Benchmark and Baseline. arXiv:2604.15088 (2026). [https://arxiv.org/abs/2604.15088](https://arxiv.org/abs/2604.15088).