Joshi et al. (2026) AANSER: A Novel Attention-based Residual U-Net architecture for Satellite data segmentation

Identification

• Journal: Remote Sensing Applications Society and Environment
• Year: 2026
• Date: 2026-01-01
• Authors: Prakhar Joshi, Manisha Kaushal, Akashdeep Sharma
• DOI: 10.1016/j.rsase.2026.101906

Research Groups

• UIET, Panjab University, Chandigarh, India
• CSED TIET DB Campus, Punjab, India

Short Summary

This paper introduces AANSER, a novel attention-based residual U-Net architecture for semantic segmentation of agricultural satellite imagery, achieving a mean Intersection over Union (IoU) of 94% across five land-cover classes.

Objective

• To develop a reliable and accurate deep learning model for semantic segmentation of complex satellite imagery, specifically for agricultural land-cover classification, addressing limitations of existing U-Net variants in multi-scale feature representation and semantic gap bridging.

Study Configuration

• Spatial Scale: Pixel-level semantic segmentation of satellite imagery, specifically Sentinel-2 data.
• Temporal Scale: Snapshot analysis of satellite imagery.

Methodology and Data

• Models used: AANSER (Attention-based Residual U-Net architecture), which integrates attention-guided feature selection, nested skip connections, squeeze-and-excitation modules, and Atrous Spatial Pyramid Pooling (ASPP).
• Data sources: Sentinel-2 satellite imagery.

Main Results

• The proposed AANSER architecture achieved a mean Intersection over Union (IoU) score of 94% for land-cover segmentation.
• Class-wise IoU scores were 98.00%, 97.82%, 91.33%, 95.13%, and 92.46% for five different land-cover classes, demonstrating robust performance across diverse categories.
• The integrated architectural design provides reliable and accurate segmentation performance for complex satellite imagery.

Contributions

• Introduction of AANSER, a novel deep learning architecture that integrates attention mechanisms, nested skip connections, squeeze-and-excitation modules, and Atrous Spatial Pyramid Pooling (ASPP) into a residual U-Net framework.
• This design effectively addresses limitations of existing U-Net variants by improving multi-scale feature representation, reducing semantic gaps between encoder and decoder, and enhancing the combination of low-level spatial details with high-level semantic information for robust and accurate satellite image segmentation.

Funding

• Not specified in the provided text.

Citation

@article{Joshi2026AANSER,
  author = {Joshi, Prakhar and Kaushal, Manisha and Sharma, Akashdeep},
  title = {AANSER: A Novel Attention-based Residual U-Net architecture for Satellite data segmentation},
  journal = {Remote Sensing Applications Society and Environment},
  year = {2026},
  doi = {10.1016/j.rsase.2026.101906},
  url = {https://doi.org/10.1016/j.rsase.2026.101906}
}