Unsupervised domain adaptation with self-attention for post-disaster building damage detection

Li, Yundong; Chen, Lin; Li, Hongguang; Hu, Wei; Han, Dong; Liu, Yi

doi:10.1016/j.neucom.2020.07.005

Cited by 24 publications

(13 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Xu et al (2019) report a drop in performance when training on two earthquakes and testing on a third one from a different region. Li et al (2020) study the performance of domain adaptation methods in the context of transferring damage models trained on airborne imagery of one hurricane to another. Gupta & Shah (2020) report mediocre OOD generalization of their xView 2 challenge model.…”

Section: Measuring Generalizationmentioning

confidence: 99%

Assessing out-of-domain generalization for robust building damage detection

Benson,

Ecker

2020

Preprint

View full text Add to dashboard Cite

An important step for limiting the negative impact of natural disasters is rapid damage assessment after a disaster occurred. For instance, building damage detection can be automated by applying computer vision techniques to satellite imagery. Such models operate in a multi-domain setting: every disaster is inherently different (new geolocation, unique circumstances), and models must be robust to a shift in distribution between disaster imagery available for training and the images of the new event. Accordingly, estimating real-world performance requires an out-of-domain (OOD) test set. However, building damage detection models have so far been evaluated mostly in the simpler yet unrealistic in-distribution (IID) test setting. Here we argue that future work should focus on the OOD regime instead. We assess OOD performance of two competitive damage detection models and find that existing state-of-the-art models show a substantial generalization gap: their performance drops when evaluated OOD on new disasters not used during training. Moreover, IID performance is not predictive of OOD performance, rendering current benchmarks uninformative about real-world performance. Code and model weights are available at https://github.com/ecker-lab/robust-bdd.

show abstract

Section: Measuring Generalizationmentioning

confidence: 99%

Assessing out-of-domain generalization for robust building damage detection

Benson,

Ecker

2020

Preprint

View full text Add to dashboard Cite

show abstract

“…Li et al [13], taking into consideration the amount of time that labeled data needs to be produced, have employed an unsupervised domain adaptation model based on unlabeled post-hurricane imagery. The model, despite its complexity, as it consists of several Generative Adversarial Networks (GANs), a classifier and a self-attention module, was evaluated by the authors as successful with regards to the transfer learning tasks that were assigned to it.…”

Section: Deep Learning For Urban Damage Assessmentmentioning

confidence: 99%

“…The studies that utilize small-scale datasets for damage classification categorize the buildings into two classes: damaged/undamaged [5,7], collapsed/non-collapsed [6,12], debris/mild damage [10]. On the other hand, studies based on larger datasets further split the wreckage level into more categories [11,13]. A finer division is beneficial for prioritizing the emergency response in the affected areas.…”

Section: Study Data Source Dataset Size ML Approach Classesmentioning

confidence: 99%

“…The regional seismicity can also be monitored using ML if the automatically captured ambient noise data are subjected to Convolutional Neural Network (CNN)-based classifiers in order to detect earthquake events [4]. In the post-disaster phase, ML approaches have been employed to locate and measure post-earthquake urban damage, confirming the relevance of this field for such kinds of applications [5][6][7][8][9][10][11][12][13][14][15]. Pre-event seismic vulnerability assessment, earthquake magnitude evaluation and post-event damage detection are complementary aspects of disaster mitigation and, if combined, can provide valuable insights to local stakeholders for traffic network disruption modeling, restoration planning and cost estimation [16].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Few-Shot Learning for Post-Earthquake Urban Damage Detection

2021

View full text Add to dashboard Cite

Among natural disasters, earthquakes are recorded to have the highest rates of human loss in the past 20 years. Their unexpected nature has severe consequences on both human lives and material infrastructure, demanding urgent action to be taken. For effective emergency relief, it is necessary to gain awareness about the level of damage in the affected areas. The use of remotely sensed imagery is popular in damage assessment applications; however, it requires a considerable amount of labeled data, which are not always easy to obtain. Taking into consideration the recent developments in the fields of Machine Learning and Computer Vision, this study investigates and employs several Few-Shot Learning (FSL) strategies in order to address data insufficiency and imbalance in post-earthquake urban damage classification. While small datasets have been tested against binary classification problems, which usually divide the urban structures into collapsed and non-collapsed, the potential of limited training data in multi-class classification has not been fully explored. To tackle this gap, four models were created, following different data balancing methods, namely cost-sensitive learning, oversampling, undersampling and Prototypical Networks. After a quantitative comparison among them, the best performing model was found to be the one based on Prototypical Networks, and it was used for the creation of damage assessment maps. The contribution of this work is twofold: we show that oversampling is the most suitable data balancing method for training Deep Convolutional Neural Networks (CNN) when compared to cost-sensitive learning and undersampling, and we demonstrate the appropriateness of Prototypical Networks in the damage classification context.

show abstract

“…is kind of work efficiency cannot meet the growing demand for translation. However, the total amount and speed of translation that an English machine translation system can complete are thousands of times that of human translation [11,12]. In actual work, English machine translation can shorten delivery time and greatly increase work efficiency.…”

Section: Introductionmentioning

confidence: 99%

English Machine Translation Model Based on an Improved Self-Attention Technology

Pan

2021

Scientific Programming

View full text Add to dashboard Cite

English machine translation is a natural language processing research direction that has important scientific research value and practical value in the current artificial intelligence boom. The variability of language, the limited ability to express semantic information, and the lack of parallel corpus resources all limit the usefulness and popularity of English machine translation in practical applications. The self-attention mechanism has received a lot of attention in English machine translation tasks because of its highly parallelizable computing ability, which reduces the model’s training time and allows it to capture the semantic relevance of all words in the context. The efficiency of the self-attention mechanism, however, differs from that of recurrent neural networks because it ignores the position and structure information between context words. The English machine translation model based on the self-attention mechanism uses sine and cosine position coding to represent the absolute position information of words in order to enable the model to use position information between words. This method, on the other hand, can reflect relative distance but does not provide directionality. As a result, a new model of English machine translation is proposed, which is based on the logarithmic position representation method and the self-attention mechanism. This model retains the distance and directional information between words, as well as the efficiency of the self-attention mechanism. Experiments show that the nonstrict phrase extraction method can effectively extract phrase translation pairs from the n-best word alignment results and that the extraction constraint strategy can improve translation quality even further. Nonstrict phrase extraction methods and n-best alignment results can significantly improve the quality of translation translations when compared to traditional phrase extraction methods based on single alignment.

show abstract

Unsupervised domain adaptation with self-attention for post-disaster building damage detection

Cited by 24 publications

References 28 publications

Assessing out-of-domain generalization for robust building damage detection

Assessing out-of-domain generalization for robust building damage detection

Few-Shot Learning for Post-Earthquake Urban Damage Detection

English Machine Translation Model Based on an Improved Self-Attention Technology

Contact Info

Product

Resources

About