INTERNSHIP @ GIC

DEVANSHI VERMA

AGENDA

Mapping Disasters for South- East Asia from Twitter using ​Python

  • Why Twitter and how do we get the data?
  • Problem and Solution
  • Framework
  • LIVE DEMO
  • Output
  • Testing and Results
  • Questions?
PROJECT 1

AGENDA

Classification of Satellite Images using Machine Learning in Python

  • Image Classification Techniques in Remote Sensing
  • Framework
  • Output
  • Questions?
PROJECT 2

MAPPING DISASTERS  FOR SOUTHEAST ASIA USING TWITTER

TWITTER DATA

WHY AND HOW?
  • Twitter is the top source of breaking news averaged 336 million users in the first quarter of 2018 with an average of around 6,000 tweets per second.
  • It provides fast, real-time information about a large-scale disaster and can produce a map within around a minute of messages being posted.

TWITTER DATA

WHY AND HOW?

TWITTER API

APPLICATION PROGRAMMING INTERFACE
  • API is a way to request and deliver information. This is done using "End points" which is an address (generally unique) that corresponds with a specific type of information.
  • 2 types of API's 
    1. Search API
    2. Stream API

PROBLEM?

AND THE SOLUTION
  • With a tweet we have 4 types of Data dictionaries : Tweet object, User object,Twitter entities and extended entities.
  • In Tweet object dictionary we have the co-ordinates in geoJSON format
  • PROBLEM: A user has to enable the precise location to add this information. This feature is OFF by default!

NAMED ENTITY RECOGNITION

Text

EXAMPLE

NAMED ENTITY RECOGNITION

EXAMPLE

NAMED ENTITY RECOGNITION

SOLUTION
  • NER is used to classify named entities in text into pre-defined categories such as the names of persons, organizations, locations, expressions of times, quantities etc.
  • For Training NER : 
    1. Collect a set of representative documents
    2. Label each token for its entity class. 
    3. Design Feature extractor appropriate. Features can be words, the position of words, POS Tags, character subsequences etc.
    4. Train a sequence classifier to predict the labels

FRAMEWORK

TESTING

RESULTS - FLOOD
No Disaster Location ​​Status ​Source
1 Flood Japan - Kamo River, Hiroshima, Kyoto, Fukuoka, Okayama, Moyotama, Mabi town, Kurashiki, Nagasaki,​Kyusyu island Detected Floodlist
2 Flood Pakistan - Lahore  Detected Floodlist
3 Flood Nepal Not Detected Floodlist
4 Flood India - Jammu, and Kashmir, Karimganj, Srinagar Detected Floodlist

TESTING

RESULTS - EARTHQUAKE
No Disaster Location ​​Status ​Source
1 Earthquake Japan - Chiba, Tokyo,  ​Fukushima Detected USGS
2 Earthquake Indonesia -  ​Sumatra  Detected USGS
3 Earthquake Japan - Osaka False Detection ------------
4 Earthquake India -  Rajasthan Detected USGS
5 Earthquake Australia - Adelaide SA Detected USGS
6 Earthquake  Taiwan- Taitung County Detected USGS

TESTING

RESULTS - LANDSLIDE
No Disaster Location ​​Status ​Source
1 Landslide India -  Tamenglong district, Manipur, Jammu, ​ Baltal Route, Jammu and Kashmir Detected Google
2 Landslide Japan - Hiroshima,  Kurashiki,  ​Kyushu Detected Google
3 Landslide China - ​Beichuan Qiang Autonomous County Detected Google

DEMO

QUESTIONS?

CLASSIFICATION OF SATELLITE IMAGES USING MACHINE LEARNING

TYPES OF IMAGE CLASSIFICATION IN REMOTE SENSING

  1. UNSUPERVISED CLASSIFICATION

  • In unsupervised clustering, group pixels into "clusters" on the basis of clustering algorithms like K-Means etc.
  • Identify the number of groups you want
  • Manually identify each cluster with land cover classes

SUPERVISED CLASSIFICATION

  • Another Pixel Based Classification Technique, select representative samples for each land cover class
  • Uses Spectral Signature (variation of reflectance or emittance of a material with respect to wavelengths)
  • It determines each class on what it resembles most in the training set.

OBIA CLASSIFICATION

  • Unsupervised and Supervised are pixel based i.e. it creates square pixels and each pixel has a class
  • In OBIA, we group them into representative groups and sizes i.e. Multi- Resolution Segmentation i.e. producing homogenous image objects by grouping pixel.
  • Features : Shape,Context and Geometry. We can take Multiple bands or even a shape file to classify object

PROPERTIES OF SATELLITE IMAGE

WHAT MAKES IT DIFFERENT

GREYSCALE IMAGES

  • Where the value of each pixel is a single sample representing only an amount of light
  • An 8-bit gray image implies 2^8 combinations and therefore, 0 - Black  and 255 - White 

RGB IMAGES

  • To define any color, you must specify the Red (R), Green (G), and Blue (B) values.
  • 0 is minimum for any RGB color and 255 is max for any RGB color

WHY?

  • Combine any band to get a different depiction of the Earth's Surface. Eg: RGB image forms a Natural color image
  • Eg: Green, NIR, SWIR2 combined give a better contrast between trees and shrubs and sawgrass marsh. These are false color images which highlight or enhance different surface features.
  • Eg: Green, Red and ​NIR, it turns out that healthy vegetation reflects a lot of NIR light

NDVI INDEX

DEMO

QUESTIONS?

APPLICATION OF MACHINE LEARNING IN REMOTE SENSING

  • These guys have trained a machine-learning algorithm to create ground-level images simply by looking at satellite pictures from above.
  • Technique used in the given article is GAN which has two parts in it: the generator that generates images and the discriminator that classifies real and fake images.
  • Team trained the discriminator with 16,000 pairs of overhead and ground-level images.

Text

APPLICATION OF MACHINE LEARNING IN REMOTE SENSING

  • Generator was fed a set of 4,000 satellite images
  • The team tested the system with 4,000 overhead images and compared them with the ground truth images.
  • Result :“The generated ground-level images looked natural although, as expected, they lacked the details of real images”.
  • Use: Classify as Rural/Urban Region!
  • Limitation : Existing DataBase is sparse

Text

THANK YOU

thisisdevanshi

thisisdevanshi

thisisdevanshi

Made with Slides.com