Correction to "Scene-Driven Multitask Parallel Attention Network for Building Extraction in High-Resolution Remote Sensing Images"

Haonan Guo, Qian Shi, Bo Du, Liangpei Zhang, Dongzhi Wang, Huaxiang DingIEEE TGRS 2021 · DOI

This document provides a formal correction to the research paper "Scene-Driven Multitask Parallel Attention Network for Building Extraction in High-Resolution Remote Sensing Images," originally published in the IEEE Transactions on Geoscience and Remote Sensing (Vol. 59, No. 5, May 2021). The correction addresses a critical visual error in the original publication where ground truth labels were mistakenly displayed instead of model predictions in a comparative figure.

The Challenge of Building Extraction

Building extraction—automatically identifying and outlining building footprints from high-resolution satellite or aerial imagery—is a fundamental task in urban planning, disaster response, and cartography. The complexity of remote sensing data presents significant challenges: varying architectural styles, shadows, tree occlusions, and diverse geographic contexts (e.g., dense urban vs. industrial zones). Standard deep learning models often struggle to maintain sharp boundaries or distinguish buildings from similar-looking man-made structures like roads or parking lots.

To address these issues, the original paper introduced the Scene-Driven Multitask Parallel Attention Network (MTPA-Net). This architecture leverages scene-level context and multitask learning to improve the precision and geometric accuracy of building extractions.

Methodology: MTPA-Net

The MTPA-Net architecture is built upon three core innovations:

  • Multitask Learning: Beyond simple binary segmentation, the network handles auxiliary tasks to refine feature representation.
  • Parallel Attention Mechanism: This allows the model to focus on relevant spatial features and channel-wise information simultaneously. By doing so, it effectively filters out "noise" such as shadows or vegetation that might mimic building textures.
  • Scene-Driven Approach: The model incorporates global scene information, allowing it to adapt its extraction strategy based on the broader geographic context of the image.

Corrected Visual Comparison

The primary purpose of this correction is to rectify Figure 12 of the original paper (reproduced below as Figure 1). In the initial publication, the columns for SegNet and PSPNet accidentally displayed the "Ground Truth" (the perfect answer key) rather than the actual model predictions. This error inadvertently made these baseline models appear significantly more accurate than they were in practice.

Qualitative comparison of building extraction results across different models, showing the corrected predictions for SegNet (d) and PSPNet (e).

Key Insights from the Corrected Results

With the corrected visual data, the qualitative superiority of MTPA-Net becomes evident:

  1. Visual Integrity: The corrected columns for SegNet (d) and PSPNet (e) now accurately reflect their performance, showing more fragmented footprints and noise compared to the ground truth (b).
  2. Superiority of MTPA-Net: Older models like MBI (c) or SegNet (d) exhibit significant noise or merged footprints in dense areas. In contrast, the proposed MTPA-Net (column k) produces clean, geometrically accurate boundaries that closely align with the ground truth.
  3. Ablation Clarity: The correction also clarifies the performance of the Single-MTPA-Net (column j). This demonstrates that while the simplified version is effective, the full multitask parallel attention framework provides a visible incremental improvement in boundary definition.

Conclusion

The necessity of this correction highlights the complexity of evaluating high-performance deep learning models in remote sensing. When models achieve high levels of accuracy, the visual difference between a prediction and the ground truth can be subtle, requiring rigorous verification to ensure scientific accuracy. The corrected evidence confirms that MTPA-Net effectively handles complex building shapes and dense clusters where earlier state-of-the-art models often fail.

Summary updated

BibTeX
@article{guo2021correction,
  author = {Haonan Guo and Qian Shi and Bo Du and Liangpei Zhang and Dongzhi Wang and Huaxiang Ding},
  title = {{Correction to "Scene-Driven Multitask Parallel Attention Network for Building Extraction in High-Resolution Remote Sensing Images"}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2021},
  doi = {10.1109/tgrs.2020.3021422},
}
Figure