BDANet: Multiscale Convolutional Neural Network with Cross-directional Attention for Building Damage Assessment from Satellite Images

Abstract: Fast and effective responses are required when a natural disaster (e.g., earthquake, hurricane, etc.) strikes. Building damage assessment from satellite imagery is critical before relief effort is deployed. With a pair of pre- and post-disaster satellite images, building damage assessment aims at predicting the extent of damage to buildings. With the powerful ability of feature representation, deep neural networks have been successfully applied to building damage assessment. Most existing works simply concatenate pre- and post-disaster images as input of a deep neural network without considering their correlations. In this paper, we propose a novel two-stage convolutional neural network for Building Damage Assessment, called BDANet. In the first stage, a U-Net is used to extract the locations of buildings. Then the network weights from the first stage are shared in the second stage for building damage assessment. In the second stage, a two-branch multi-scale U-Net is employed as backbone, where pre- and post-disaster images are fed into the network separately. A cross-directional attention module is proposed to explore the correlations between pre- and post-disaster images. Moreover, CutMix data augmentation is exploited to tackle the challenge of difficult classes. The proposed method achieves state-of-the-art performance on a large-scale dataset -- xBD. The code is available at https://github.com/ShaneShen/BDANet-Building-Damage-Assessment.

• The paper introduces BDANet, a novel two-stage CNN that uses cross-directional attention between pre- and post-disaster images for improved building segmentation and damage classification.
• It leverages a multi-branch U-Net design and CutMix data augmentation to mitigate class imbalance and boost performance on the extensive xBD dataset.
• Results demonstrate state-of-the-art accuracy and reduced training time, making BDANet a promising tool for real-time disaster management and response.

"BDANet: Multiscale Convolutional Neural Network with Cross-directional Attention for Building Damage Assessment from Satellite Images" (2105.07364)

Introduction

The paper introduces BDANet, a multiscale convolutional neural network specifically designed for assessing building damage using satellite images. This framework addresses the need for accurate and efficient damage assessment following natural disasters, leveraging high-resolution pre- and post-disaster imagery. Unlike traditional methods, BDANet integrates a cross-directional attention module to effectively model correlations between image pairs, improving damage prediction accuracy.

BDANet Architecture

BDANet follows a two-stage approach:

1. Stage 1: Building Segmentation
   • Utilizes a U-Net architecture to segment building locations from pre-disaster images, providing clear delineation of buildings before disaster impact.

     Figure 1: Visual examples of building segmentation in Stage 1. First row to third row are: pre-disaster images, ground truth of building segmentation, and segmentation results of Stage 1.

2. Stage 2: Damage Assessment
   • Implements a two-branch multi-scale U-Net where pre- and post-disaster images are processed separately. The network weights from Stage 1 are used to initialize this stage, encouraging efficient learning.
   • Introduces a cross-directional attention module which enhances feature representation by exploring spatial and channel correlations between the two image sets.

     Figure 2: Overview of the proposed framework (BDANet). The U-Net structure is used in both stages. It consists of two stages: (a) Stage 1: building segmentation, (b) Stage 2: damage assessment.

Cross-Directional Attention Module

The CDA module aids in identifying informative sections of images by recalibrating feature weights. This consists of spatial and channel squeezes and excitations, performed in a cross-modal fashion between pre- and post-disaster features, ensuring robust damage level classification.

Figure 3: Framework of the proposed cross-directional attention (CDA) module.

Data Augmentation and CutMix

To overcome class imbalance and improve model generalization, BDANet employs CutMix augmentation for difficult-to-classify damage levels. This technique combines image patches to increase training set diversity, focusing the learning process on challenging scenarios.

Figure 4: Data augmentation with CutMix for difficult classes.

Results and Evaluation

BDANet was evaluated on the xBD dataset, the largest available dataset for building damage assessment, demonstrating state-of-the-art performance in terms of overall classification accuracy. Critical metrics such as F1 score showed substantial improvements, particularly in minor damage levels, due to the focused augmentation strategies.

Figure 5: Damage assessment results. From left to right: (a) pre-disaster image, (b) post-disaster image, (c) the ground-truth, (d) FCN, (e) SegNet, (f) DeepLab and (g) our proposed network.

Computational Efficiency

The integration of weights from Stage 1 for Stage 2 initialization significantly reduced training time, while maintaining high validation accuracy from early epochs. Computational costs were analyzed showing the efficiency of BDANet compared with other networks, with only marginal increases in resource use due to CDA and MFF modules.

Figure 6: Evaluation of the training efficiency in Stage 2 by applying network weights from Stage 1 as initialization. (a) Validation accuracy (\%). (b) Cross-entropy loss.

Conclusion

BDANet provides an advanced framework for building damage assessment, offering significant improvements in damage detection accuracy through innovative network design components like the CDA module and strategic data augmentation. Future work may explore adaptive learning strategies for real-time assessment scenarios, optimizing deployment in disaster management operations. The methodologies and results present promising directions for further research in automated remote sensing and disaster response technologies.