ATT-CR: Adaptive Triangular Transformer for Cloud Removal

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Engineering & Applied Sciences, Environmental Science & Earth Systems · Depth: Expert, extended

Summary

The Adaptive Triangular Transformer for Cloud Removal (ATT-CR) is a novel model designed to accurately reconstruct ground objects obscured by clouds in remote sensing images. It addresses the high computational complexity of traditional Transformer self-attention and the interference caused by treating cloudy and clean pixels uniformly. ATT-CR features two core components: Triangular Attention (TAN) and the Feature Selected Gating Module (FSGM). TAN employs lower and upper triangular matrices to approximate Softmax attention with ℴ(N) computational complexity, effectively reducing costs and alleviating the low-rank limitation of linear attention. FSGM adaptively distinguishes between cloudy and clean features at both channel and spatial levels, minimizing invalid information propagation. Extensive experiments on RICE1, RICE2, T-CLOUD, and SEN12MS-CR datasets demonstrate ATT-CR's superior performance, achieving PSNR gains of up to 0.48dB on RICE2 and significantly reducing parameters and FLOPs compared to existing state-of-the-art methods like CR-former and ACA-CRNet.

Key takeaway

For Machine Learning Engineers developing remote sensing image restoration models, ATT-CR provides a robust solution for cloud removal. Its efficient Triangular Attention and adaptive Feature Selected Gating Module significantly improve accuracy while drastically reducing computational costs compared to prior Transformer-based methods. You should consider adapting this architecture for other image degradation tasks like haze or shadow removal, leveraging its ability to handle large-scale occlusions and ambiguous textures with high fidelity.

Key insights

Adaptive Triangular Transformer (ATT-CR) efficiently removes clouds from remote sensing images by combining linear triangular attention and feature-selective gating.

Principles

• Linear attention can achieve full-rank with triangular matrices.
• Adaptive gating improves robustness to occluded features.
• Multi-scale tokens enhance feature diversity.

Method

ATT-CR employs a multi-stage encoder-decoder. Triangular Attention (TAN) uses split heads and triangular matrices for ℴ(N) complexity and full-rank attention. Feature Selected Gating Module (FSGM) adaptively modulates TAN outputs to filter cloudy features.

In practice

• Apply triangular matrices to linear attention.
• Integrate gating modules for feature selection.
• Use multi-scale convolutions for diverse tokens.

Topics

• Cloud Removal
• Remote Sensing Imagery
• Transformer Architectures
• Linear Attention
• Feature Gating
• Image Restoration

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• ATT-CR: Adaptive Triangular Transformer for Cloud Removal — ATT-CR uses adaptive triangular attention and feature gating to efficiently remove clouds, reducing complexity and interference.
• ATT-CR: Adaptive Triangular Transformer for Cloud Removal — ATT-CR uses triangular attention and feature gating to efficiently remove clouds from remote sensing images, improving accuracy.
• LALE: Lightweight-Transformer Architecture for Land-Cover Estimation — LALE achieves efficient remote sensing image segmentation by combining resolution-specific encoders and an all-MLP decoder.
• RPC-GS: Gaussian Splatting with native RPC Rendering for Satellite Imagery — Native RPC integration in Gaussian Splatting significantly improves satellite imagery 3D reconstruction accuracy.
• EfficientPENet: Real-Time Depth Completion from Sparse LiDAR via Lightweight Multi-Modal Fusion — EfficientPENet achieves real-time depth completion on edge devices by combining a lightweight ConvNeXt backbone with sparsity-invariant convolutions and CSPN refinement.
• Adaptive Slicing-Assisted Hyper Inference for Enhanced Small Object Detection in High-Resolution Imagery — ASAHI adaptively slices high-resolution images to enhance small object detection, reducing computation and improving performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.