ATT-CR Beats 8 Rivals Removing Clouds from Satellite Images With O(N) Attention
ATT-CR cuts transformer attention complexity from O(N²) to O(N) and filters cloudy pixels from attention, beating prior cloud removal methods on four benchmarks
Takeaways
- ATT-CR is a new transformer for satellite cloud removal that reduces attention complexity from O(N²) to O(N) using triangular matrix decomposition.
- A Feature Selected Gating Module (FSGM) adaptively suppresses cloudy pixel features before they corrupt subsequent attention layers.
- Evaluated on four benchmarks: RICE-I, RICE-II, T-Cloud, and SEN12MS-CR (~110,000 samples across 169 global regions).
- Outperforms prior methods including pix2pix, SPA-GAN, Restormer, CMNet, ACA-CRNet, and CR-former on all datasets.
- Code and pre-trained checkpoints are publicly available on GitHub.
- The O(N) attention design is general and applicable beyond cloud removal to any restoration task with structured input corruption.
Satellite imagery is the backbone of environmental monitoring, precision agriculture, disaster response, and land-use mapping. But there is a persistent, frustrating obstacle: on average, 66% of the Earth's sky is covered by clouds annually. That means the majority of optical satellite captures are partially or fully obscured. Cloud removal -- computationally reconstructing what lies beneath -- has become a critical preprocessing step, and transformer-based models have emerged as the leading approach. A new paper published in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing proposes ATT-CR, a model that tackles two long-standing weaknesses in transformer-based cloud removal at once.
Why transformers struggle with clouds
Transformer-based models, leveraging self-attention, demonstrate remarkable capability in modeling long-range dependencies and integrating multi-scale features. However, their quadratic computational complexity poses challenges for high-resolution remote sensing data. For a 256x256 image patch, that means the attention matrix has over 4 billion entries -- and satellite images are often much larger. Scaling up is essentially prohibitive without approximation.
But there is a second, subtler problem that most efficiency-focused work ignores. Treating both cloudy and clean pixels as valid within the attention computation brings disturbances in subsequent layers, leading to suboptimal performance. In other words, when a transformer attends to a corrupted, cloud-covered pixel as if it were valid signal, that noise propagates forward through the network and degrades the final reconstruction. Existing methods do not explicitly separate good pixels from bad ones during attention.