On Transfer Learning for Building Damage Assessment from Satellite Imagery in Emergency Contexts

Abstract

When a natural disaster occurs, humanitarian organizations need to be prompt, effective, and efficient to support people whose security is threatened. Satellite imagery offers rich and reliable information to support expert decision-making, yet its annotation remains labour-intensive and tedious. In this work, we evaluate the applicability of convolutional neural networks (CNN) in supporting building damage assessment in an emergency context. Despite data scarcity, we develop a deep learning workflow to support humanitarians in time-constrained emergency situations. To expedite decision-making and take advantage of the inevitable delay to receive post-disaster satellite images, we decouple building localization and damage classification tasks into two isolated models. Our contribution is to show the complexity of the damage classification task and use established transfer learning techniques to fine-tune the model learning and estimate the minimal number of annotated samples required for the model to be functional in operational situations.

Evaluation of Deep Learning Models for Building Damage Mapping in Emergency Response Settings

“…In this case, IND refers to a proportion of samples that are intentionally split from all samples for testing purposes. Previous studies such as [35], [16], [6], [8], [36], [37] and [5] have attempted to evaluate the generalization ability to serve emergency response efforts. [6] evaluated the generalization ability of several well-known CNN-based networks namely VGG-16, InceptionV3, ResNet50, and DenseNet121.…”

“…In their study, the generalization of models was assessed in terms of geographic location (testing on satellite images of an unseen area) and data generalization (transferring from satellite to aerial photo), mainly in binary classification (damage or no-damage). [37] proposed a two-step model design comprised of BuildingNet for building localizer followed by DamageNet for damage classifier. Each model was trained before disasters to make them ready for inference, thus minimizing the postincident execution time.…”

“…For example, [36] evaluate the geographical transferability in three different datasets: satellite, airborne, and UAV in two building damage classes, namely damage and intact. Similarly, [38], [37] and [33] work in binary mapping. [37] combined nodamage and minor-damage into one class and major-and destroyed class into another.…”

“…Similarly, [38], [37] and [33] work in binary mapping. [37] combined nodamage and minor-damage into one class and major-and destroyed class into another. They argue that minor-damage levels do not require immediate emergency attention from humanitarian organizations.…”

Unsupervised learning framework for region based damage assessment on xBD, a large satellite imagery

“…can be quantified using remote sensing images. The basic methodology in doing this task is to detect changes in building structures before and after the disaster and then classify the building damage according to predefined scale (Bouchard et al (2022)). Several recent works have shown remarkable precision in determining the degree of damage to buildings (Li et al (2022); Gilani et al (2018)).…”