Building Damage Classification

The emergence of social media sites specially twitter has been considered a useful platform for humanitarian difficulties. During natural disasters it provides the most recent information ranging from number of injuries to deaths, infrastructure damage and urgent rescue to people. All this informati

Project Title

Building Damage Classification

Project Area of Specialization

Artificial Intelligence

Project Summary

The emergence of social media sites specially twitter has been considered a useful platform for humanitarian difficulties. During natural disasters it provides the most recent information ranging from number of injuries to deaths, infrastructure damage and urgent rescue to people. All this information is in the form of text and images. Since to rescue the injured people by responding immediately social media sites plays a significant role to provide information in seconds or minutes. In order to identify useful tweets supervised learning approach can be used to classify data into two classes i.e. relevant and irrelevant but it will be difficult to obtain large amount of labelled data thus semi supervised model will be useful to deal with large amount of unlabeled data.

A graph based deep learning framework has been selected for learning an inductive semi supervised model to classify tweets. CNN is combined with graph based network in order to define the similarity between labelled and unlabeled training objects. The evaluation is conducted by using two real-world twitter datasets.

The graph has been constructed by using k-nearest neighbor approach for finding nearest neighbors of objects. In order to find the nearest objects efficiently k-d tree data structure has been used.

In the architecture of the neural network model the first layer maps each of the words into a distributed representation. Its output is passed through a number of convolutional and pooling layers to learn high-level feature representations. A convolutional operation applies different filters to obtain different feature maps. Max pooling is performed before the final activations for classification in which Softmax is used.

The crisis dataset contains different types of tweets therefore in preprocessing all the irrelevant text i.e. URLs, username etc. are removed. All the crisis dataset has been collected from Nepal Earthquake 2015 and the 2013 Queensland floods. For training dataset has been splitted into 60% as training, 30% as test and 10% as development. For training purpose word2vec model is used.

Thus a graph based semi-supervised deep learning framework has been presented based on CNN which combines the loss for predicting the class labels with a loss for predicting the context defined by a similarity graph.

Project Objectives

As most of the people use social media sites in order to post images and news about current situations and mostly people consider these sources as a useful tool for retrieving information. Therefore, to estimate, damages that occur during natural disasters the model will collect the raw data from social media sites. The collectted raw data will be filttered ans then the filttered data will be classified to predict the damage caused to the buildings.

Project Implementation Method

First of all, we will create a social media scrapper to scrape data from social media related to any hash tag. Since the data is scraped from social media, it will contain irrelevant information more than relevant information. The first thing we need to do is process the data and filter out irrelevant images and text. Afterwards data will be prepared for deep hybrid model for damage estimation.

Benefits of the Project

The model uses semi-supervised deep learning approach to classify the incoming data into three classes i.e. high, low and irrelevant. In previous approaches mostly two classes have been implemented i.e. relevant and irrelevant but in this case the relevant class has been divided into two classes i.e. high damage and low damage. As the obtained labelled data always have a less amount as compared to unlabeled data that’s why semi-supervised method has been implemented to provide efficient and effective results.

Technical Details of Final Deliverable

The Hybrid model will be delivered predicting the results with efficient accuracy on both text and images using Convolutional Nueral Network (CNN). It will be deployed on Amazon Cloud and a web portal will be devloped.

Final Deliverable of the Project

Software System

Type of Industry

IT

Technologies

Artificial Intelligence(AI)

Sustainable Development Goals

Decent Work and Economic Growth, Industry, Innovation and Infrastructure

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com